Vertical take-off and landing aircraft diagnosis generalization enhancement method based on causal weight

Abstract

The invention relates to the technical field of aircraft fault diagnosis, in particular to a vertical take-off and landing aircraft diagnosis generalization enhancement method based on causal weight. Firstly, a directed acyclic graph between flight features and fault types is established through an information flow causal discovery method, the causal relationship between the flight features and the fault types can be clearly presented, and a basis is provided for subsequent analysis. Then, mutual information between nodes is calculated and converted into causal strength for causal enhancement, and expression of the causal relationship in the data is enhanced; in the convolutional neural network, causal weights are used for dynamic weighting to extract local features, causal features of a directed acyclic graph are combined, and feature information of different types is fully utilized. According to the method, features are fused through a graph convolutional network to obtain high-order features, adversarial training is completed through a full-connection network, the constructed fault prediction model integrates various technical advantages, the generalization ability of the model is effectively improved, the model can better adapt to different working conditions and data changes, and aircraft faults are accurately diagnosed.

Description

Technical Field

[0001] The invention relates to the technical field of aircraft fault diagnosis, and in particular to a method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weights. Background Art

[0002] With the rapid development of artificial intelligence technology, vertical take-off and landing aircraft have been widely used in many fields such as drones, military and commercial aviation. These aircraft play a key role in performing missions, ensuring the efficiency and safety of flight missions. However, as the use time increases, vertical take-off and landing aircraft often face problems such as overload, fatigue, wear, sensor failure, etc. when flying under complex working conditions, especially under different wind speeds and wind directions. These problems may cause aircraft failures or cause serious safety accidents.

[0003] At present, intelligent diagnosis methods based on machine learning and deep learning have emerged in the field of aircraft fault diagnosis. These methods are of great significance in improving the accuracy of fault diagnosis and reducing the reliance on manual feature extraction. However, most of the existing intelligent fault diagnosis frameworks focus on the correlation between data for classification and identification, but ignore the causal relationship between data and labels. This makes these methods insufficient in the diagnosis of complex multi-factor systems, resulting in poor generalization ability of the fault prediction models obtained, which needs to be solved urgently. Summary of the invention

[0004] In order to avoid and overcome the technical problems existing in the prior art, the present invention provides a method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weights. The present invention can effectively improve the generalization ability of the fault prediction model.

[0005] To achieve the above object, the present invention provides the following technical solutions: A method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weights includes the following enhancement steps: S1. Obtain fault flight data of vertical take-off and landing aircraft, and establish a directed acyclic graph between flight characteristics and fault types through information flow causal discovery method; S2. Calculate the mutual information between nodes in the directed acyclic graph and convert the mutual information into the causal strength of each node to causally strengthen each node; S3, inputting the causally strengthened nodes into the convolutional neural network for convolution operation, and using the causal weights for dynamic weighting during the convolution operation to extract local features; S4, concatenating the local features with the causal features in the directed acyclic graph to obtain concatenated features; and processing the concatenated features using a multi-layer perceptron to obtain an adjacency matrix of the nodes; S5. Use graph convolutional network to fuse the concatenated features and adjacency matrix to obtain high-order features; S6. Input the high-order features into the fully connected network to complete adversarial training with the convolutional neural network and the graph convolutional network as feature extractors and the fully connected network as the fault classifier; the feature extractor and the fault classifier constitute a fault prediction model.

[0006] As a further solution of the present invention: the specific steps of step S2 are as follows: S21. Based on the time series data of each node in the directed acyclic graph, the information entropy of each node is calculated; the calculation formula of the information entropy is as follows: ; In the formula, Represents the directed acyclic graph Nodes Information entropy of Representation Node exist The value of the moment; express The probability distribution of represents the logarithmic function with base 10; S22. Use information entropy to calculate the mutual information between nodes. The calculation formula of mutual information is as follows: ; In the formula, Represents the directed acyclic graph Nodes and Nodes The mutual information between Represents the directed acyclic graph Nodes Information entropy of Representation Node and nodes The joint information entropy between S23, arranging the calculated mutual information in order to form a causal relationship matrix in the form of a square matrix , where each element represents the mutual information between two corresponding nodes; S24. Causality Matrix Perform normalization to obtain the causal strength matrix ; S25. Combined Causal Strength Matrix Each causal strength in the causal strength strengthens each node causally; S26. Calculate the value of each node after causal reinforcement at each moment through the calculation formula of causal reinforcement.

[0007] As a further solution of the present invention: the calculation formula of causal reinforcement is expressed as follows: ; In the formula, Represents the node after causal reinforcement exist The value of the moment; Representation Node exist The value of the moment.

[0008] As a further solution of the present invention: Causal Strength Matrix The calculation formulas for each element in are as follows: ; In the formula, Represents the directed acyclic graph Nodes and Nodes the causal strength between them; Representation Node The weight factor of Representation Node The standard deviation of the time series data; Representation Node The mean of the time series data; Represents the directed acyclic graph Nodes and Nodes The mutual information between ; represents the scaling factor; Representation and Node The minimum value of all relevant mutual information; Representation and Node The maximum value of all relevant mutual information.

[0009] As a further solution of the present invention: the specific steps of step S3 are as follows: S31, arranging the values ​​of each node at each time after causal reinforcement into a data set at the corresponding time; S32, inputting the data sets at each moment into the convolutional neural network in sequence for convolution operation, so as to extract the local features between each node through the convolutional neural network.

[0010] As a further solution of the present invention: the convolution operation is expressed as follows: ; In the formula, Representation Node exist Local characteristics of the moment; Indicates Nodes exist The value of the moment; express and The similarity between express and The similarity between express and Between softmax Function computes value; Indicates the total number of nodes; Represents the node after causal reinforcement exist The value of the moment; Represents the convolution kernel of the convolutional neural network; Represents the bias term of the convolutional neural network; Represents the convolution operation; Indicated by natural constant e An exponential function with base .

[0011] As a further solution of the present invention: the specific steps of step S4 are as follows: S41. The local features and the causal features in the directed acyclic graph are spliced ​​together to obtain spliced ​​features. The spliced ​​features are expressed as follows: ; In the formula, express The splicing features formed after splicing; express Nodes in the directed acyclic graph at any moment causal characteristics; Represents a splicing operation; S42, then perform nonlinear transformation on the concatenated features to obtain nonlinear features; the nonlinear features are expressed as follows: ; In the formula, express The nonlinear characteristics of represents the nonlinear bias term; represents nonlinear weight; represents the activation function; S43. Use a multi-layer perceptron to transform the concatenated features into high-dimensional features. The high-dimensional features are expressed as follows: ; In the formula, Represents the feature matrix formed by the combination of various nonlinear features; represents a multi-layer perceptron; express Corresponding high-dimensional features; S44. Multiply the feature matrix by its transposed matrix to obtain an adjacency matrix; the adjacency matrix is ​​expressed as follows: ; In the formula, represents the adjacency matrix; Represents a matrix multiplication operation; express The transposed matrix of S45. Select the front of each node nearest neighbors, forming a sparse adjacency matrix ; ; In the formula, Indicates the front nearest neighbor operation.

[0012] As a further solution of the present invention: the specific content of step S5 is as follows: S51, input the feature matrix and the sparse adjacency matrix into the graph convolutional network; S52, using two layers of receptive field convolution to perform convolution operation to obtain corresponding high-order features; ; In the formula, Represents the features learned by the first layer of receptive field convolution; Represents the weight matrix of the first layer of receptive field convolution; represents the receptive field convolution; Represents the weight matrix of the second layer receptive field convolution; Represents the features learned by the second-layer receptive field convolution, i.e., high-order features.

[0013] As a further solution of the present invention: the objective function in the adversarial training process is as follows: ; In the formula, represents the objective function; represents the classification loss; represents the domain alignment loss; represents a hyperparameter.

[0014] As a further solution of the present invention: the calculation formula of each loss is specifically expressed as follows: ; In the formula, represents source domain samples; represents the target domain sample; Indicates the fault type. Represents a fault classifier model, inputs a source domain sample, and outputs the fault type to which the source domain belongs The probability of Represents a fault classifier model, inputs a target domain sample, and outputs the fault type to which the target domain sample belongs probability; represents the source domain classifier output; represents the target domain classifier output; ; In the formula, represents the domain discriminator; represents the probability that the domain discriminator judges the generated source domain sample as fake; represents the loss of target domain samples.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. First, a directed acyclic graph between flight characteristics and fault types is established through the information flow causal discovery method, which can clearly present the causal relationship between the two and provide a basis for subsequent analysis. Then, the mutual information between nodes is calculated and converted into causal strength for causal reinforcement, which enhances the expression of causal relationships in the data. In the convolutional neural network, causal weights are used to dynamically weight local features, combined with the causal characteristics of the directed acyclic graph, and full use is made of different types of feature information. The high-order features are obtained by fusing features through the graph convolutional network, and adversarial training is completed through the fully connected network. The fault prediction model constructed combines multiple technical advantages and effectively improves the generalization ability of the model, so that it can better adapt to different working conditions and data changes, and accurately diagnose aircraft faults.

[0016] 2. Start by calculating the node information entropy, and calculate the mutual information between nodes based on the information entropy. This method based on information theory can accurately measure the degree of association between nodes. The mutual information is arranged into a causal relationship matrix and normalized to obtain a causal strength matrix, which provides a quantitative basis for causal reinforcement. According to specific steps, the nodes are causally strengthened in combination with the causal strength matrix, so that the nodes can more prominently reflect the causal relationship between them and other nodes, and further mine the causal information in the data, thereby providing more valuable data features for subsequent model training, which helps to improve the accuracy and generalization of the fault diagnosis model.

[0017] 3. The causal reinforcement calculation formula clarifies how to adjust the values ​​of nodes at various times according to the causal strength matrix. Through this formula, the value of each node at a certain moment can comprehensively consider the values ​​of other nodes at the same moment and the causal strength between them, realizing the dynamic adjustment and fusion of node information. This makes the information transmission of nodes in the time series more reasonable, enhances the coherence of causal relationships in the data, and helps the model more accurately capture the relationship between aircraft failures and flight characteristics over time, thereby improving the reliability and generalization of fault prediction.

[0018] 4. The calculation formula of the causal strength matrix elements takes into account multiple factors. Including node weight factors, mean and standard deviation of time series data, maximum value of mutual information, etc. Through this complex calculation method, the strength of the causal relationship between nodes can be more carefully characterized. The weight factors of different nodes can be set according to their importance, and the statistical characteristics of time series data and the maximum value of mutual information further calibrate the causal strength, so that the causal strength matrix can more accurately reflect the true causal relationship between nodes, provide more accurate parameters for subsequent causal reinforcement and model training, and improve the performance of the fault diagnosis model.

[0019] 5. The node values ​​after causal reinforcement are organized into a data set, providing the convolutional neural network with orderly and causal information-rich data input. Inputting the data set in sequence for convolution operation enables the convolutional neural network to effectively extract local features between each node. This data processing and feature extraction method not only retains the data advantages after causal reinforcement, but also utilizes the advantages of convolutional neural networks in local feature extraction, laying a good foundation for subsequent model fusion and fault diagnosis, and helping to improve the model's ability to identify aircraft fault features.

[0020] 6. The convolution operation formula details how to use causal weights for dynamic weighting in convolutional neural networks. The formula includes elements such as node similarity calculation, softmax function, and causal strength matrix. Through this complex calculation method, the convolution operation can more specifically weight the information of different nodes. When extracting local features, the node information with a close causal relationship with the current node is highlighted, and irrelevant or weakly related information is suppressed, so that the extracted local features can better reflect the causal relationship between aircraft failures and flight characteristics, improve the quality of feature extraction, and thus enhance the fault diagnosis ability and generalization of the model.

[0021] 7. By concatenating local features and causal features to obtain concatenated features, feature information of different levels and types is integrated to enrich the expression of data. Nonlinear transformation, multi-layer perceptron processing and other operations are performed on the concatenated features to gradually transform low-dimensional features into high-dimensional features, which can better capture complex patterns in the data. Finally, the adjacency matrix and sparse adjacency matrix are obtained through matrix operations, which provide a suitable input format for the graph convolutional network, help the graph convolutional network to more effectively mine the graph structure information in the data, and improve the accuracy and generalization of the model for aircraft fault diagnosis.

[0022] 8. Input the feature matrix and sparse adjacency matrix into the graph convolutional network, and use two layers of receptive field convolution to perform convolution operations to obtain high-order features. This operation mode can fully utilize the advantages of graph convolutional networks in processing graph structure data, and learn different levels of graph structure features through convolutions of different layers. The features learned by the first layer of receptive field convolution provide the basis for the second layer, and the second layer further mines more complex features. The obtained high-order features can more comprehensively reflect the complex relationship between aircraft flight characteristics and fault types, which helps to improve the performance and generalization ability of fault diagnosis models.

[0023] 9. The adversarial training objective function comprehensively considers classification loss and domain alignment loss. Classification loss is used to measure the accuracy of the model in classifying different samples to ensure that the model can correctly identify the type of aircraft failure. Domain alignment loss is dedicated to reducing the difference between the source domain and the target domain, so that the model can maintain good performance under different working conditions or data distributions. By adjusting the hyperparameters λ Balancing the two losses can improve the classification accuracy of the model while enhancing its generalization ability, making it better suited to aircraft fault diagnosis tasks in different environments.

[0024] 10. The classification loss formula can effectively guide the model to learn common fault features and improve the classification ability of different samples by comparing the classification of source domain and target domain samples. The domain alignment loss formula uses domain discriminators to measure the probability of misclassification of source domain samples and the loss of target domain samples, respectively, prompting the model to align the data distribution of the source domain and target domain during training, reducing the impact of inter-domain differences on model performance, and further enhancing the generalization of the model, so that the model can diagnose aircraft faults more stably and accurately when facing different working conditions and data changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of the enhancement steps of the present invention. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] See also Figure 1 In an embodiment of the present invention, a method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weights includes the following contents: 1. Construct a directed acyclic graph: The faulty flight data of vertical take-off and landing aircraft are obtained, and a directed acyclic graph between flight characteristics and fault types is established through the information flow causal discovery method.

[0028] 1. Collect fault flight data under different operating conditions from the sensor system of the vertical take-off and landing aircraft.

[0029] During the process of fault flight data collection, the vertical take-off and landing aircraft is placed under different operating conditions for operation. By setting different flight modes (such as vertical take-off and landing, hovering, horizontal flight, etc.) and simulating different wind speed and wind direction conditions, the sensor data of the aircraft under these conditions is collected. The time interval for data collection is usually uniform to ensure that each data point is from the same time period for subsequent processing. In this process, the fault flight data collected includes: Three-axis vibration (X, Y, and Z axis vibration): These vibration data reflect the dynamic response of the aircraft in different directions, recording the vibration amplitude and frequency in each direction.

[0030] Torque: records the torque generated by the aircraft engine or propulsion system, reflecting the load and power changes of the power system.

[0031] Rotational speed: records the rotational speed of key components such as rotors and motors, reflecting the operating status of the power system.

[0032] Each sensor continuously monitors the aircraft's operating status throughout the experiment, generating periodic values. For example, the three-axis vibration sensor records the following: ; ; ; in, Represents the vector composed of vibration data in the X-axis direction, Represents the vibration data in the X-axis direction at the first moment, Represents the vibration data in the X-axis direction at the second moment, Represents the vibration data in the X-axis direction at the third moment, Indicates Vibration data in the X-axis direction at each moment. Represents a vector consisting of vibration data in the Y-axis direction, Represents the vibration data in the Y-axis direction at the first moment, Represents the vibration data in the Y-axis direction at the second moment, Represents the vibration data in the Y-axis direction at the third moment, Indicates Vibration data in the Y-axis direction at each moment. Represents the vector composed of vibration data in the Z-axis direction, Represents the vibration data in the Z-axis direction at the first moment, Represents the vibration data in the Z-axis direction at the second moment, Represents the vibration data in the Z-axis direction at the third moment, Indicates Vibration data in the Z-axis direction at each moment.

[0033] Torque data: ; Speed ​​data: .in, Represents a vector composed of vibration data, Represents the torque data at the first moment, Represents the torque data at the second moment, Represents the torque data at the third moment, Indicates Torque data at each moment. Represents a vector composed of speed data, Indicates the speed data at the first moment, Indicates the speed data at the second moment, Indicates the speed data at the third moment, Indicates Speed ​​data at a moment.

[0034] In order to ensure data diversity, the experiment not only collects normal aircraft operation data, but also simulates a variety of common aircraft failure scenarios. For example: Fault condition 1: The battery overheats, causing unstable power output.

[0035] Fault condition 2: Engine failure, manifested as abnormal fluctuations in torque and speed.

[0036] Fault state 3: Sensor failure, resulting in distortion or complete loss of sensor data.

[0037] Through the above data collection process, the sensor data of the vertical take-off and landing aircraft in different working states and different fault modes can be obtained. Using this data, causal analysis, feature weight allocation and fault diagnosis model training can be carried out in the future, so as to achieve efficient diagnosis and early warning of aircraft faults.

[0038] 2. Use information flow causal discovery methods to perform causal discovery on time series data.

[0039] 1) Data preprocessing: Perform noise filtering (Butterworth low-pass filtering, median filtering), missing value interpolation (linear interpolation) and standardization (Z-score) on sensor time series data (such as vibration, torque, and speed) to ensure data quality.

[0040] 2) Discovery of causal structure: Fast causal inference: Construct partial ancestry graphs (PAGs) through conditional independence tests (such as kernel conditional independence tests, KCIT) to identify adjacency relationships between variables , forming a Markov equivalence class.

[0041] Information flow algorithm: Based on the asymmetry of information transmission, it calculates the information flow intensity between variables, directs causal edges, and generates a complete directed acyclic graph (DAG).

[0042] Output and application: Finally, a DAG reflecting the causal relationship of the fault is obtained (such as "battery temperature → abnormal torque → increased vibration"), which provides structural priors for subsequent feature weighting and graph convolutional networks, and guides the model to focus on key causal features.

[0043] 2. Causal reinforcement: The mutual information between nodes in the directed acyclic graph is calculated and converted into the causal strength of each node to causally strengthen each node.

[0044] In faulty network diagnosis, it is a challenge to ensure that the model assigns weights according to the importance of faults in a multi-factor system. To optimize model learning, the input data is processed by computing information dependencies, fusing causal relationships, and applying a weighting matrix.

[0045] The information entropy of each node is calculated to measure the randomness or uncertainty in the time series data of the node.

[0046] Assume Node exist The value of the moment is discrete , and has a corresponding probability distribution , then the node Information entropy Defined as: ; In the formula, Represents the directed acyclic graph Nodes Information entropy of Representation Node exist The value of the moment; express The probability distribution of Represents the base 10 logarithmic function.

[0047] node and nodes The mutual information between quantifies the degree of information sharing between them. The calculation formula of mutual information is as follows: ; In the formula, Represents the directed acyclic graph Nodes and Nodes The mutual information between Represents the directed acyclic graph Nodes Information entropy of Representation Node and nodes The joint information entropy between .

[0048] By calculating the mutual information between all node pairs, the mutual information can be arranged in sequence to form a causal relationship matrix in the form of a square matrix. , and each element represents the mutual information between the corresponding two nodes.

[0049] After receiving the causal relationship matrix, it is first normalized to scale the feature values ​​to the range [0,1]. This normalization ensures that each feature has a clear relative importance in the subsequent analysis or modeling task, eliminating scale differences. Standard normalization techniques such as weighted min-max normalization and Z-score normalization are applied to achieve this. After the normalization operation, the causal strength matrix is ​​obtained It is expressed as follows: ; In the formula, Represents the directed acyclic graph Nodes and Nodes the causal strength between them; Representation Node The weight factor of Representation Node The standard deviation of the time series data; Representation Node The mean of the time series data; Represents the directed acyclic graph Nodes and Nodes The mutual information between ; Represents the scaling factor, which is used to control the impact of weighted minimum and maximum normalization on the final result; Representation and Node The minimum value of all relevant mutual information; Representation and Node The maximum value of all relevant mutual information.

[0050] This process produces a normalized causal strength matrix , which is then used to adjust the input data of the convolutional neural network.

[0051] In the first convolutional layer of the convolutional neural network, the nodes The weighted input It is obtained by summing the weighted contributions of all other nodes, that is, combining the causal strength matrix The causal strength in each node is used to causally strengthen each node. The calculation formula of causal strengthening is as follows: ; In the formula, Represents the node after causal reinforcement exist The value of the moment; Representation Node in The value of the moment.

[0052] This input is adjusted based on the causal relationships between nodes to produce a weighted input sequence for the convolutional neural network to use.

[0053] 3. Extract local features: In order to consider the impact of the cause of the failure, an adaptive attention adjustment mechanism is used to enhance the flexibility of the model. This mechanism dynamically assigns weights to strengthen the contribution of nodes that are closely related to the output, while suppressing features that are not related to the failure. The values ​​of each node at each moment after causal reinforcement are organized into a data set at the corresponding moment, and then the data sets at each moment are sequentially input into the convolutional neural network for adaptively adjusted convolution operations to propose local features between each node through the convolutional neural network. The convolution operation with this adaptive adjustment is given as follows: ; In the formula, Representation Node exist Local characteristics of the moment; Indicates Nodes exist The value of the moment; express and The similarity between express and The similarity between express and Between softmax Function computes value; Indicates the total number of nodes; Represents the node after causal reinforcement exist The value of the moment; Represents the convolution kernel of the convolutional neural network; Represents the bias term of the convolutional neural network; Represents the convolution operation; Indicated by natural constant e An exponential function with base .

[0054] This dynamic adjustment improves the model's ability to understand time series data, especially in capturing potential causal relationships. Normalization, weighting, and attention mechanisms enable the model to adjust its response based on changes in the importance of failure causes, thereby improving training efficiency and prediction accuracy.

[0055] 4. Get the adjacency matrix: After processing the time series sensor data through the convolutional neural network, the fully connected layer generates new feature representations, namely local features. These local features are then input into the graph convolutional network, where two different node types need to be specially handled to mitigate the impact of different feature node types. In order to further enhance the training process by leveraging the directed acyclic graph, a node sharing mechanism is proposed. Initially, the causal features in the directed acyclic graph and the local features based on the convolutional neural network are concatenated by the following method: ; In the formula, express The splicing features formed after splicing; express Nodes in the directed acyclic graph at any moment causal characteristics; Represents a concatenation operation.

[0056] This operation merges the two feature vectors to produce a more informative and compact representation of each node. The concatenated feature vector is then processed through a nonlinear transformation, which can be expressed by the following equation: ; In the formula, express The nonlinear characteristics of represents the nonlinear bias term; Represents nonlinear weights, usually expressed in the form of a matrix, that is, a mapping matrix that transforms the concatenated feature vector into a new feature space; represents the activation function, which introduces nonlinearity in the transformation process, enabling the network to model complex relationships between features.

[0057] In order to construct the adjacency matrix in the graph convolutional network The feature matrix formed by the combination of each nonlinear feature is first processed by a multi-layer perceptron to obtain a high-dimensional feature, and then the adjacency matrix is ​​obtained by multiplying the high-dimensional feature with its transpose. Next, the first r nearest neighbors of each node are selected to form a sparse adjacency matrix The specific expressions are as follows: ; In the formula, Represents the feature matrix formed by the combination of various nonlinear features; represents a multi-layer perceptron; express Corresponding high-dimensional features; represents the adjacency matrix; Represents a matrix multiplication operation; express The transposed matrix of Indicates the front nearest neighbor operation.

[0058] 5. Obtaining high-level features: Applications have different receptive fields The receptive field of is convolved into the instance graph, and the corresponding high-order feature representation is obtained through the receptive field convolution layer: ; In the formula, Represents the features learned by the first layer of receptive field convolution; Represents the weight matrix of the first layer of receptive field convolution; represents the receptive field convolution; Represents the weight matrix of the second layer receptive field convolution; Represents the features learned by the second-layer receptive field convolution, i.e., high-order features.

[0059] 6. Build a fault prediction model: The goal of adversarial training is to make the model perform well in both the source domain (training data conditions) and the target domain (new conditions). Its core is the following two components: 1. Feature Extractor: It consists of a convolutional neural network and a graph convolutional network, which is responsible for generating high-order feature representations.

[0060] The role of graph convolutional networks: Generate high-quality features containing causal relationships through graph structure modeling and feature fusion.

[0061] 2. Domain Classifier: A separate fully connected network is used to determine whether the feature comes from the source domain or the target domain.

[0062] Adversarial Target: The feature extractor is optimized to “fool” the domain classifier into distinguishing between source and target domain features.

[0063] 3. Objective function: The objective function in the adversarial training process is as follows: ; In the formula, represents the objective function; represents the classification loss; represents the domain alignment loss; Represents a hyperparameter.

[0064] ; In the formula, represents source domain samples; represents the target domain sample; Indicates the fault type. Represents a fault classifier model, inputs a source domain sample, and outputs the fault type to which the source domain belongs The probability of Represents a fault classifier model, inputs a target domain sample, and outputs the fault type to which the target domain sample belongs probability; represents the source domain classifier output; Represents the target domain classifier output.

[0065] ; In the formula, It represents the domain discriminator; represents the probability that the domain discriminator judges the generated source domain sample as fake; represents the loss of target domain samples.

[0066] By combining these two loss functions, the ultimate goal is to achieve accurate classification on the source domain while learning features that are equally effective on the target domain.

[0067] If the target domain verification error does not decrease continuously, the training is terminated early. After the training is completed, the model used to predict faults is the optimized "feature extractor + label classifier".

[0068] The prediction process using the fault prediction model is as follows: The specific steps are as follows: 1) Enter new data: Input sensor data (such as vibration, speed, torque, etc.) of the target domain (new operating condition).

[0069] 2) Feature extraction: The feature extractor (a hybrid structure composed of convolutional neural networks and graph convolutional networks) converts raw data into high-dimensional feature representations. These features have been optimized through adversarial training and can ignore differences in operating conditions and focus on fault-related patterns.

[0070] 3) Fault classification: The label classifier (fully connected network or Softmax layer) outputs the fault type probability based on the extracted features, thus completing the fault diagnosis.

[0071] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weights, characterized in that: The enhancement steps include: S1. Obtain fault flight data of vertical take-off and landing aircraft, and establish a directed acyclic graph between flight characteristics and fault types through information flow causal discovery method; S2. Calculate the mutual information between nodes in the directed acyclic graph and convert the mutual information into the causal strength of each node to causally strengthen each node; S3, inputting the causally strengthened nodes into the convolutional neural network for convolution operation, and using the causal weights for dynamic weighting during the convolution operation to extract local features; S4, concatenating the local features with the causal features in the directed acyclic graph to obtain concatenated features; Use a multi-layer perceptron to process the concatenated features to obtain the adjacency matrix of the nodes; S5. Use graph convolutional network to fuse the concatenated features and adjacency matrix to obtain high-order features; S6. Input the high-order features into the fully connected network to complete adversarial training with the convolutional neural network and the graph convolutional network as feature extractors and the fully connected network as the fault classifier; the feature extractor and the fault classifier constitute a fault prediction model.

2. The method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weight according to claim 1, characterized in that: The specific steps of step S2 are as follows: S21. Based on the time series data of each node in the directed acyclic graph, the information entropy of each node is calculated; the calculation formula of the information entropy is as follows: ； In the formula, Represents the directed acyclic graph Nodes Information entropy of Representation Node exist The value of the moment; express The probability distribution of represents the logarithmic function with base 10; S22. Use information entropy to calculate the mutual information between nodes. The calculation formula of mutual information is as follows: ； In the formula, Represents the directed acyclic graph Nodes and Nodes The mutual information between Represents the directed acyclic graph Nodes Information entropy of Representation Node and nodes The joint information entropy between S23, arranging the calculated mutual information in order to form a causal relationship matrix in the form of a square matrix , where each element represents the mutual information between two corresponding nodes; S24. Causality Matrix Perform normalization to obtain the causal strength matrix ; S25. Combined Causal Strength Matrix Each causal strength in the causal strength strengthens each node causally; S26. Calculate the value of each node after causal reinforcement at each moment through the calculation formula of causal reinforcement.

3. The method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weight according to claim 2, characterized in that: The calculation formula of causal reinforcement is as follows: ； In the formula, Represents the node after causal reinforcement exist The value of the moment; Represents the directed acyclic graph Nodes and Nodes the causal strength between them; Representation Node exist The value of the moment.

4. The method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weight according to claim 3 is characterized in that: Causal Strength Matrix The calculation formulas for each element in are as follows: ； In the formula, Representation Node The weight factor of Representation Node The standard deviation of the time series data; Representation Node The mean of the time series data; Represents the directed acyclic graph Nodes and Nodes The mutual information between ; represents the scaling factor; Representation and Node The minimum value of all relevant mutual information; Representation and Node The maximum value of all relevant mutual information.

5. The method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weight according to claim 4, characterized in that: The specific steps of step S3 are as follows: S31, arranging the values ​​of each node at each time after causal reinforcement into a data set at the corresponding time; S32, inputting the data sets at each moment into the convolutional neural network in sequence for convolution operation, so as to extract the local features between each node through the convolutional neural network.

6. The method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weight according to claim 5, characterized in that: The convolution operation in a convolutional neural network is represented as follows: ； In the formula, Representation Node exist Local features of the moment; Indicates Nodes exist The value of the moment; express and The similarity between express and The similarity between express and Between softmax Function computes value; Indicates the total number of nodes; Represents the node after causal reinforcement exist The value of the moment; Represents the convolution kernel of the convolutional neural network; Represents the bias term of the convolutional neural network; Represents the convolution operation; Indicated by natural constant e An exponential function with base .

7. The method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weight according to claim 6, characterized in that: The specific steps of step S4 are as follows: S41. The local features and the causal features in the directed acyclic graph are spliced ​​together to obtain spliced ​​features. The spliced ​​features are expressed as follows: ； In the formula, express The splicing features formed after splicing; express Nodes in the directed acyclic graph at any moment causal characteristics; Represents a splicing operation; S42, then performing nonlinear transformation on the splicing features to obtain nonlinear features; The nonlinear characteristics are expressed as follows: ； In the formula, express The nonlinear characteristics of represents the nonlinear bias term; represents nonlinear weight; represents the activation function; S43. Use a multi-layer perceptron to transform the concatenated features into high-dimensional features. The high-dimensional features are expressed as follows: ； In the formula, Represents the feature matrix formed by the combination of various nonlinear features; represents a multi-layer perceptron; express Corresponding high-dimensional features; S44. Multiply the feature matrix by its transposed matrix to obtain an adjacency matrix; the adjacency matrix is ​​expressed as follows: ； In the formula, represents the adjacency matrix; Represents a matrix multiplication operation; express The transposed matrix of S45. Select the front of each node nearest neighbors, forming a sparse adjacency matrix ; ； In the formula, Indicates the front nearest neighbor operation.

8. The method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weight according to claim 7, characterized in that: The specific content of step S5 is as follows: S51, input the feature matrix and the sparse adjacency matrix into the graph convolutional network; S52, using two layers of receptive field convolution to perform convolution operation to obtain corresponding high-order features; ； In the formula, Represents the features learned by the first layer of receptive field convolution; Represents the weight matrix of the first layer of receptive field convolution; represents the receptive field convolution; Represents the weight matrix of the second layer receptive field convolution; Represents the features learned by the second-layer receptive field convolution, i.e., high-order features.

9. The method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weight according to claim 8, characterized in that: The objective function in the adversarial training process is as follows: ； In the formula, represents the objective function; represents the classification loss; represents the domain alignment loss; represents a hyperparameter.

10. The method for enhancing the generalization of vertical take-off and landing aircraft diagnosis based on causal weight according to claim 9, characterized in that: The calculation formulas for each loss are specifically expressed as follows: ； In the formula, represents source domain samples; represents the target domain sample; Indicates the fault type. Represents a fault classifier model, inputs a source domain sample, and outputs the fault type to which the source domain belongs The probability of Represents a fault classifier model, inputs a target domain sample, and outputs the fault type to which the target domain sample belongs The probability of represents the source domain classifier output; represents the target domain classifier output; ； In the formula, represents the domain discriminator; represents the probability that the domain discriminator judges the generated source domain sample as fake; represents the loss of target domain samples.

Images