Unmanned aerial vehicle flight control system abnormality tracing method based on deep matching network

Abstract

The present application relates to the technical field of unmanned aerial vehicle fault detection, and in particular to an unmanned aerial vehicle flight control system abnormality tracing method based on a deep matching network. An abnormal knowledge graph containing entities and semantic relationships is constructed from pre-processed multi-source heterogeneous data, and an abnormal cause is mapped to a cause label vector by a graph encoder. A flight control state abnormality representation vector is generated by a deep time series feature extraction network, and abnormalities are detected. When an abnormality is detected, the correlation score of the abnormality representation vector and the cause label vector is calculated by a deep matching network to screen the top K candidate causes. Then, a candidate causal path is constructed by searching along a preset path depth from the candidate cause node as a source node in the knowledge graph, and the candidate causal path is sorted by a node correlation scoring mechanism to output an interpretable tracing path. This method is based on deep matching network technology, improves the accuracy and interpretability of abnormality detection and tracing in complex environments, and realizes efficient and accurate abnormality cause matching and causal path reasoning.

Description

Technical Field

[0001] The present invention relates to the technical field of UAV fault detection, and in particular to a method for tracing the abnormality source of a UAV flight control system based on a deep matching network. Background Art

[0002] A drone's flight control system, the "brain" of the drone, is responsible for driving various mechanical operations, such as the motors driving the rotors, the servos adjusting direction, and controlling the aircraft's attitude. With the widespread application of drones in military reconnaissance, agricultural inspections, logistics and transportation, and emergency rescue, the reliability and safety of these flight control systems have become critical factors in ensuring the smooth execution of missions. However, the complex hardware and software integration of the flight control system makes it extremely susceptible to multiple factors during operation, such as sensor errors, actuator anomalies, bus interference, and environmental disturbances. These factors, in turn, can affect the safety and stability of the drone's flight missions. Therefore, real-time monitoring of the flight control system's status, identifying anomalies, and promptly and accurately tracing the causes of these anomalies are crucial to ensuring the safety and stability of drone flights.

[0003] During the flight of a drone, various abnormal phenomena may occur in the flight control system, such as IMU drift, electronic control failure, servo jamming, control lag, etc., and the same type of abnormal phenomenon may be caused by different inducements. In addition, in a complex flight environment, a drone may experience multiple abnormal phenomena at the same time, and multiple abnormal phenomena may originate from one abnormal inducement. There is a chain reaction causal relationship between the abnormal phenomena, such as servo jamming causing control lag, which in turn causes motor overload. Traditional flight control system anomaly detection methods mainly rely on rule setting or model driving, such as rule matching-based or physical model-based anomaly detection methods. Such methods have good accuracy and interpretability when facing known anomalies, but their ability to identify low-frequency faults is insufficient, and it is difficult to effectively handle anomaly problems in complex environments. With the rapid development of artificial intelligence technology, deep learning-based anomaly detection methods have gradually emerged. Deep learning-based anomaly detection methods can extract anomaly features from a large amount of sensor time series data in the flight control system to complete anomaly detection. Although such methods can improve the generalization ability of anomaly detection, they often lack sufficient interpretability, making it difficult to trace the cause of the anomaly and provide intuitive explanations after the anomaly occurs. They are insufficient in assisting operation and maintenance decision-making and optimization.

[0004] In recent years, the development of deep learning and knowledge graph technology provides a new path for intelligent diagnosis of flight control system anomalies. Deep learning models can extract abnormal features from a large amount of time series data, with strong abnormal detection sensitivity and generalization ability. Knowledge graph, as a structured knowledge expression carrier, has strong entity relationship modeling and causal chain representation ability, and can provide a basis for explainable reasoning of abnormal causes. For example, a Chinese patent with the application publication number CN120066004A and the patent name "Unmanned aerial vehicle flight control fault intelligent diagnosis method and system based on data and knowledge fusion" collects multi-modal flight data through on-board sensors and pre-processes them, then constructs a pre-trained large model for fault feature extraction and identification, and uses natural language processing technology to extract entities and relationships from text data to construct a flight control fault knowledge graph. Through a semantic mapping function, the large model recognition result is associated with the knowledge graph instance, and the fault root cause is reasoned based on the causal relationship in the graph to generate a diagnosis report containing fault phenomena, diagnostic analysis, maintenance solutions and historical cases, realizing collaborative anomaly diagnosis based on data-driven and knowledge-driven. However, this scheme has the problem that when facing complex abnormal scenarios, the fault cause judgment may not be accurate due to the many-to-many relationship of semantic mapping, and the traceability and explainability of complex chain abnormalities may be insufficient due to the dependence of cause and effect reasoning on semantic mapping and logical reasoning. Therefore, there is an urgent need for a method for tracing abnormal causes of unmanned aerial vehicle flight control systems that has strong abnormal cause identification and explainable cause and effect reasoning ability, can efficiently and accurately realize abnormal detection and dynamic explainability reasoning of abnormal causes, and improve the intelligent diagnosis capability of flight control systems in complex environments. SUMMARY

[0005] The technical problem solved by the present application is to provide a method for tracing abnormal causes of unmanned aerial vehicle flight control systems based on a deep matching network, which can improve the accuracy and explainability of abnormal detection and tracing in complex environments, and realize efficient and accurate abnormal cause matching and causal path reasoning.

[0006] The basic scheme provided by the present application is a method for tracing abnormal causes of unmanned aerial vehicle flight control systems based on a deep matching network, comprising the following steps:

[0007] S1, generating training samples: obtaining multi-source heterogeneous data from an unmanned aerial vehicle task execution platform and performing preprocessing operations to generate training samples with uniform format, the multi-source heterogeneous data including flight control state time series data and text records;

[0008] S2, constructing an abnormal knowledge graph: performing entity recognition and relationship extraction on the preprocessed text records to form a triple set and construct a knowledge graph , a node set representing semantic entities related to the flight control system, and an edge set representing semantic relationships between entities;

[0009] S3, generating a cause label vector: mapping the abnormal causes in the semantic entities of the knowledge graph into a unified low-dimensional cause label vector through a graph encoder :

[0010]

[0011] The vector is used to represent the semantic features of the abnormal causes in the knowledge graph structure and their relationship information with other entities;

[0012] S4, generating an abnormal representation vector and anomaly detection: constructing a deep time series feature extraction network and pre-training, using the trained deep time series feature extraction network to map the current real-time acquired flight control state time series data into a flight control state abnormal representation vector ; introducing an anomaly detection mechanism to determine the abnormal representation vector , outputting the probability value of the current flight control state belonging to an anomaly, and the anomaly detection mechanism outputs the probability value of the current flight control state belonging to an anomaly through the following formula:

[0013]

[0014] If the probability value is greater than a preset anomaly determination threshold , it is determined that the current state is an abnormal state, otherwise it is determined to be a normal state;

[0015] S5, constructing a deep matching model and outputting a candidate cause set: based on the abnormal representation vector and the cause label vector , a deep matching network model is constructed and pre-trained; the determination result of the anomaly detection mechanism in S4 is obtained, when the anomaly detection mechanism in S4 determines an abnormal state, the trained deep matching network model is used to output the relevance matching score of the current abnormal representation vector and all cause label vectors , all cause label vectors are ranked in descending order according to the matching score result, and the top K cause label vectors are selected as the candidate cause set of the current anomaly;

[0016] S6, abnormal cause and cause path reasoning: according to the cause label vectors ranked in the top K in S5, the corresponding abnormal cause nodes in the knowledge graph are obtained ​In the knowledge graph, a depth-first search is performed according to a preset path depth L, a plurality of candidate causal paths with the cause node as a source point and the target anomaly as a terminal point are constructed, a node correlation scoring mechanism is used to calculate the overall semantic correlation scores of the candidate causal paths, all candidate paths are sorted in descending order according to the score results, the top N paths are selected as the explainable root cause path set of the current anomaly, and the path with the highest score is selected as the most representative causal reasoning result.

[0017] The principle of the application is that the application realizes efficient detection and explainable root cause of the abnormality of the flight control system of the unmanned aerial vehicle by fusing deep learning and knowledge graph technology: multi-source heterogeneous data is obtained through the unmanned aerial vehicle task platform, the multi-source heterogeneous data is preprocessed to form training samples with unified format, the entity relationship of the flight control system is structured expressed through the knowledge graph, the vector representation of the abnormal state and the cause is realized through the deep time sequence feature extraction network and the graph encoder, the abnormal representation vector and the cause label vector are formed, the current state is determined through the abnormality detection mechanism, when the abnormal state is determined, the correlation degree of the abnormal representation vector and all cause label vectors is quantified through the deep matching network, so as to screen the candidate causes, finally, the depth-first search and path scoring are performed based on the causal link of the knowledge graph, the explainable root cause path set is formed, and the path with the highest score is selected as the generated causal reasoning result.

[0018] The application has the following beneficial effects:

[0019] 1. The abnormal representation vector and the cause label vector are extracted from the flight control time sequence data and the text record through the fusion architecture of deep learning and knowledge graph, the correlation degree between the abnormal representation vector and the cause label vector is quantified in the form of score through the deep matching network, the output cause set is more objective, when a plurality of causes may cause the same abnormality, the score high and low can directly reflect the possibility priority of the cause, and the fault positioning time is shortened. The deep time sequence feature extraction network can automatically extract abnormal features from the time sequence data, the knowledge graph can explicitly model the causal logic through the structured entity relationship, and the combination of the two can make the system still maintain high detection sensitivity and root cause reliability in complex environments such as multiple abnormality concurrency, chain causal relationship and new or low-frequency fault.

[0020] 2、Through the depth-first search and path scoring mechanism of the knowledge graph, the candidate causal path is ensured to meet the physical logic constraints, blind reasoning relying on data statistical correlation is avoided, the accuracy and reliability of the traceability result are improved, the traceability path set output by the entity relationship path of the knowledge graph has strong explainability, which is convenient for manual review, and the path with the highest score is taken as the explainable causal reasoning result, so that the operation and maintenance personnel can intuitively understand the logical chain of abnormal occurrence, quickly locate the fault point and develop maintenance strategies. In addition, the explainable causal reasoning result not only meets the demand of unmanned aerial vehicle operation and maintenance for traceability, but also provides data support for system optimization, and the high-frequency causal path can guide the designer to optimize the design and reduce the abnormal occurrence rate from the source.

[0021] 3、The deep time sequence feature extraction network and the deep matching network are both end-to-end models, which can process the flight control system data stream in real time, realize abnormal rapid detection and abnormal cause traceability, meet the real-time requirement of unmanned aerial vehicle flight task, and through the abnormal detection mechanism, the current abnormal feature vector is determined, when it is in an abnormal state, the subsequent abnormal cause matching and cause path traceability process is carried out, otherwise the process is stopped, so as to save the computing resources and enhance the pertinence of system response. In addition, the modular structure of the knowledge graph makes it have flexible expansion, when it is necessary to add new abnormal types, components or causal relationships, only the nodes and edges of the graph need to be updated, without the need to modify the model architecture on a large scale, which is convenient for system updating and reduces the long-term operation and maintenance cost.

[0022] The scheme combines the feature extraction capability of deep learning and the causal reasoning advantage of knowledge graph, improves the generalization of abnormal detection and the accuracy of traceability in complex environment, realizes efficient identification of causes and causal path reasoning through the deep matching network and the structured reasoning of knowledge graph, uses the semantic explainability of causal path to assist operation and maintenance decision, and has real-time processing capability and dynamic expansion characteristics of knowledge graph, which meets the intelligent and complex operation and maintenance requirements of unmanned aerial vehicle flight control system.

[0023] Further, the preprocessing operation in S1 includes timestamp synchronization, missing value processing, outlier removal and field standardization. Through the above preprocessing operations, the time consistency and integrity of the multi-source heterogeneous data are ensured, the noise data is removed, the data format is unified, the high-quality input is provided for subsequent abnormal knowledge graph construction and model training, and the detection deviation caused by data quality problems is avoided.

[0024] Further, the knowledge graph constructed in S2 is subjected to graph cleaning, semantic fusion and graph structure optimization operations. Through the above operations, redundant entities and invalid relationships in the graph are removed, multi-source semantic information is fused, graph structure connectivity is optimized, graph quality and expression ability are improved, and the reliability of causal path search is ensured.

[0025] Further, in the S4, the step of constructing a deep time sequence feature extraction network and pre-training includes:

[0026] S41, constructing a deep time sequence feature extraction network , mapping the flight control state time sequence data into a flight control state abnormality feature vector :

[0027]

[0028] wherein is the flight control state time sequence data formed in a continuous time window for each flight process obtained, denotes the state feature vector of the time step;

[0029] S42, according to the training samples obtained in S1, constructing a flight task for a typical abnormality mode relying on a flight control simulation platform, collecting simulation flight control state time sequence data and corresponding injected text information to form simulation training samples, combining the training samples obtained in S1 with the simulation training samples to form an abnormality feature training sample set with class labels, the abnormality feature training sample set including samples of normal flight states and various abnormal states;

[0030] S43, training the deep time sequence feature extraction network through the abnormality feature training sample set.

[0031] The present scheme constructs training samples of typical abnormality modes through a simulation platform, expands the diversity of data, makes the training samples cover various task environments and working condition changes, especially for low-frequency or new-type abnormal scenarios, enhances the generalization ability of the model to complex abnormality modes, and improves the robustness of feature extraction.

[0032] Further, in the pre-training process in the S4, a similarity learning mechanism based on a twin structure is introduced, and any pair of samples are respectively encoded as:

[0033]

[0034] Combining denotes whether the sample pair belongs to the same abnormality class, and the contrast loss function is defined as follows:

[0035]

[0036] wherein is a distance metric function, is a preset minimum interval threshold;

[0037] At the same time, for each sample , the abnormality probability value , the standard binary cross-entropy is used as the classification loss:

[0038]

[0039] The total training loss function is a weighted joint optimization of the two:

[0040]

[0041] wherein, .

[0042] This scheme ensures that feature vectors of the same type of anomaly are close, and feature vectors of different types of anomaly are far apart, enhancing the discriminability of the feature space, thereby achieving fine representation and differentiation of the flight control abnormal state.

[0043] Further, the step of constructing a deep matching network model based on the anomaly representation vector and the cause label vector in S5 and pre-training includes:

[0044] S51, based on the anomaly representation vector and the cause label vector construct a deep matching network model for representing the correlation between the anomaly representation vector and the cause label vector :

[0045]

[0046] wherein represents the th anomaly cause semantic entity generated cause label vector, for representing the relevance matching score between the current anomaly representation vector and the cause label vector , is a deep matching model, is a set of learnable parameters;

[0047] S52, according to the training samples obtained in S1, each training sample obtains an anomaly representation vector through a deep time sequence feature extraction network, and obtains a cause label vector through a knowledge graph. The anomaly representation vector, the corresponding correct cause label vector, and other interference label vectors of each training sample constitute a matching training sample.

[0048] S53, train the deep time sequence feature extraction network through the matching training sample set, use a supervised training method for optimization, and use cross-entropy loss as the loss function.

[0049] The scheme optimizes the model parameters by using labeled matching training samples, so that the deep matching network can accurately quantify the semantic association between abnormal representation and cause, improve the accuracy of candidate cause ranking, and reduce the interference of irrelevant causes.

[0050] Further, the node correlation scoring mechanism is adopted in S6 to calculate the overall semantic correlation score of each candidate causal path, and the formula is as follows:

[0051]

[0052] wherein represents the th candidate path, is used to represent the correlation matching score between the current abnormal representation vector and the cause label vector represents the preset path depth.

[0053] The scheme evaluates the overall correlation of the causal path according to the average matching score, avoids the influence of single node deviation on the reasoning result, ensures that the filtered causal path not only conforms to the structured relationship of the knowledge graph, but also is highly related to the current abnormal representation, and improves the credibility and explainability of the traceability result. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 Fig. 1 is a flowchart of the embodiment 1 of the unmanned aerial vehicle flight control system abnormal traceability method based on the deep matching network of the present application;

[0055] Figure 2 Fig. 2 is a schematic diagram of the flight control abnormal representation and cause label generation and matching process in the embodiment 1 of the present application;

[0056] Figure 3 Fig. 3 is a schematic diagram of the real-time state end to abnormal cause end unmanned aerial vehicle abnormal traceability mode of the embodiment 1 of the present application. DETAILED DESCRIPTION

[0057] The following will be further described in detail through specific embodiments:

[0058] The embodiment 1 is basically as shown in the accompanying drawings: Figure 1 , Figure 2 and Figure 3

[0059] The unmanned aerial vehicle flight control system abnormal traceability method based on the deep matching network comprises the following steps:

[0060] ​​S1, generating training samples: obtaining multi-source heterogeneous data from a UAV task execution platform and performing preprocessing operations to generate training samples with uniform format, the multi-source heterogeneous data including flight control state time series data and text records, the flight control state time series data including flight control log data (control instructions, attitude angles, speeds, heights, etc.), actuator sensor data (rudder deflection, steering motor current, signal feedback, etc.), abnormal record data (abnormal alarm code, timestamp, etc.); the text records include maintenance work orders and engineer text annotations (abnormal components, abnormal causes and processing methods described in natural language); performing preprocessing operations (timestamp synchronization, missing value processing, outlier removal, field standardization, etc.) on the multi-source heterogeneous data; generating training samples with uniform format through the preprocessed multi-source heterogeneous data.

[0061] S2, constructing an abnormal knowledge graph: performing entity recognition and relationship extraction on the preprocessed text records to form a triple set and construct a knowledge graph , a node set representing flight control system related semantic entities, including system components, abnormal causes, abnormal manifestations, control instructions, and maintenance operations, and an edge set representing semantic relationships between entities, including causal links, functional dependencies, time sequence, and operation results; through graph cleaning, semantic fusion and graph structure optimization operations, the graph cleaning, semantic fusion and graph structure optimization operations improve the quality and expression ability of the graph.

[0062] S3, generating a cause label vector: based on the constructed knowledge graph, selecting a set of abnormal cause semantic entities related to flight control system failures , using a graph encoder to combine graph structure information and neighbor context to encode and generate a vectorized representation of the abnormal cause:

[0063]

[0064] The vector is used to represent the semantic features of the abnormal cause in the knowledge graph structure and its relationship information with other entities.

[0065] S4, generating an abnormal representation vector and performing abnormal detection: constructing a deep time series feature extraction network and pretraining, using the trained deep time series feature extraction network to map the current real-time acquired flight control state time series data to a flight control state abnormal representation vector ; introducing an abnormal detection mechanism to determine the abnormal representation vector , outputting the probability value of the current flight control state belonging to an abnormality, the abnormal detection mechanism outputs the probability value of the current flight control state belonging to an abnormality through the following formula:

[0066]

[0067] If the probability value is greater than a preset abnormality determination threshold , it is determined that the current state is an abnormal state, otherwise it is determined to be a normal state.

[0068] In the step of constructing a deep time sequence feature extraction network and pre-training in S4, the following steps are included:

[0069] S41, constructing a deep time sequence feature extraction network , mapping the flight control state time sequence data into a flight control state abnormality representation vector

[0070]

[0071] wherein is the flight control state time sequence data formed in a continuous time window for each flight process obtained, denotes the state feature vector of the i-th time step;

[0072] S42, according to the training samples obtained in S1, constructing a flight task for a typical abnormality mode relying on a flight control simulation platform, collecting simulation flight control state time sequence data and corresponding injected text information, forming simulation training samples, combining the training samples obtained in S1 with the simulation training samples, to constitute an abnormality representation training sample set with class labels, the abnormality representation training sample set including samples of normal flight states and various abnormal states;

[0073] S43, training the deep time sequence feature extraction network through the abnormality representation training sample set.

[0074] When the deep time sequence feature extraction network is trained in S43, a similarity learning mechanism based on a twin structure is introduced, and for any pair of samples are respectively encoded as

[0075]

[0076] Combining denotes whether the sample pair belongs to the same abnormality class, and the contrast loss function is defined as follows:

[0077]

[0078] wherein is a distance metric function, is a preset minimum interval threshold;

[0079] At the same time, for each sample , its abnormality probability value is predicted, and a standard binary classification cross-entropy is used as the classification loss:​​

[0080]

[0081] The total training loss function is a weighted joint optimization of the two:

[0082]

[0083] in, .

[0084] S5. Build a deep matching model and output a candidate cause set: Based on the abnormal characterization vector and the inducement label vector Build a deep matching network model and pre-train it; obtain the judgment result of the anomaly detection mechanism in S4. When the anomaly detection mechanism in S4 determines an abnormal state, use the trained deep matching network model to output the current anomaly representation vector With all the trigger labels vector According to the correlation matching score, all inducement label vectors are sorted in descending order according to the matching score results, and the top K inducement label vectors are selected as the candidate inducement set of the current anomaly.

[0085] The step of constructing a deep matching network model based on the abnormal characterization vector and the inducement label vector and performing pre-training in S5 includes:

[0086] S51, based on the abnormal characterization vector and the inducement label vector Build a deep matching network model to represent abnormal representation vectors Vector with inducement label The degree of correlation between:

[0087]

[0088] in Indicates the The abnormal cause semantic entity The generated incentive label vector, Used to represent the current abnormal characterization vector Vector with inducement label The correlation matching score between For the deep matching model, is a set of learnable parameters;

[0089] S52. Based on the training samples obtained in S1, each training sample obtains an abnormal representation vector through a deep temporal feature extraction network, and obtains an inducement label vector through a knowledge graph. The abnormal representation vector of each training sample, the corresponding correct inducement label vector, and other interference label vectors constitute a matching training sample;

[0090] S53, train the deep time sequence feature extraction network by matching the training sample set, optimize using a supervised training method, and use cross entropy loss as the loss function.

[0091] S6, abnormal cause and effect path reasoning: according to the top K cause label vectors matched in S5, obtain the corresponding abnormal cause nodes in the knowledge graph , perform depth-first search according to the preset path depth L in the knowledge graph, construct multiple candidate cause and effect paths with the cause node as the source point and the target abnormality as the terminal point; adopt a node correlation scoring mechanism to calculate the overall semantic correlation score of each candidate cause and effect path, sort all candidate paths in descending order according to the score results, select the top N paths as the current abnormality explainable traceability path set, and select the highest score path as the most representative cause and effect reasoning result.

[0092] In S6, the formula for calculating the overall semantic correlation score of each candidate cause and effect path using the node correlation scoring mechanism is as follows:

[0093]

[0094] Wherein represents the th candidate path, is used to represent the correlation matching score between the current abnormality representation vector and the cause label vector , and represents the preset path depth.

[0095] In the actual deployment stage, the abnormality detection and abnormality cause path tracing from the real-time state end to the abnormality inducing end are realized through the following process: through time series feature extraction on the real-time acquired flight control state time series data, an abnormality characterization vector is obtained, and according to the abnormality detection mechanism, a probability value of the current flight control state belonging to an abnormality is output; when the output probability value is greater than an abnormality judgment threshold value, it is judged to be an abnormal state; at this time, the correlation matching score of the current abnormality characterization vector and all inducing label vectors is obtained through a deep matching network, all inducing label vectors are arranged in descending order according to the matching score result, and the top K inducing label vectors are selected as the candidate inducing set of the current abnormality; according to the top K inducing label vectors in the matching score, the corresponding abnormality inducing nodes in the knowledge graph are obtained, and in the knowledge graph, a depth-first search is performed according to a preset path depth L, and a plurality of candidate cause-effect paths with the inducing nodes as the source points and the target abnormality as the terminal points are constructed; a node correlation scoring mechanism is used to calculate the overall semantic correlation score of each candidate cause-effect path, all candidate paths are arranged in descending order according to the scoring result, the top N paths are selected as the explainable tracing path set of the current abnormality, and the path with the highest score is selected as the most representative cause-effect reasoning result, which can be used for decision support, maintenance suggestion or manual review.

[0096] The above is only an embodiment of the present application, and common knowledge of specific structures and characteristics in the scheme is not described in detail here. Those skilled in the art know all ordinary technical knowledge in the field of the application before the filing date or the priority date, can know all prior art in the field, and have the ability to apply conventional experimental means before that date. Those skilled in the art can improve and implement the present scheme based on their own ability under the guidance of the present application, and some typical known structures or known methods should not be an obstacle for those skilled in the art to implement the present application. It should be noted that for those skilled in the art, without departing from the structure of the present application, a number of modifications and improvements can be made, which should be considered as the protection scope of the present application, and these will not affect the effect and practicality of the present application. The protection scope claimed in the present application should be subject to the content of its claims, and the specific implementation mode and the like in the specification can be used to explain the content of the claims.

Claims

1. A method for abnormal tracing of a UAV flight control system based on a deep matching network, characterized in that: The method comprises the following steps: S1, generating training samples: obtaining multi-source heterogeneous data from a UAV task execution platform and performing preprocessing operations to generate training samples with uniform formats, wherein the multi-source heterogeneous data comprises flight control state time series data and text records; S2, construct an abnormal knowledge graph: entity recognition and relation extraction are performed on the preprocessed text records to form a triple set and construct a knowledge graph , a node set represents a semantic entity related to the flight control system, and an edge set represents the semantic relationship between entities; S3, generating a cause label vector: mapping the abnormal causes in the semantic entities of the knowledge graph into a unified low-dimensional cause label vector by a graph encoder :​ The vector is used to represent the semantic features of the abnormal inducement in the knowledge graph structure and the relationship information with other entities. S4, generating an abnormality characterization vector and abnormality detection: constructing a deep time series feature extraction network and pre-training, using the trained deep time series feature extraction network to map the current real-time acquired flight control state time series data into a flight control state abnormality characterization vector ; introducing an abnormality detection mechanism to the abnormality characterization vector determination, outputting the probability value of the current flight control state belonging to an abnormality, and the abnormality detection mechanism outputs the probability value of the current flight control state belonging to an abnormality through the following formula: If the probability value Greater than the preset abnormality judgment threshold , then the current state is determined to be abnormal, otherwise it is determined to be normal; In the S4 pre-training process, a similarity learning mechanism based on a twin structure is introduced. For any pair of samples are respectively encoded as: Combining represents whether the pair belongs to the same abnormal class or not, and the contrastive loss function is defined as follows: wherein is a distance metric function, is a preset minimum separation threshold; At the same time, for each sample predicting its anomaly probability value using standard binary cross-entropy as the classification loss: The total training loss function is a weighted joint optimization of the two: wherein ; S5. Build a deep matching model and output a candidate cause set: Based on the abnormal characterization vector and the inducement label vector Build a deep matching network model and pre-train it; obtain the judgment result of the anomaly detection mechanism in S4. When the anomaly detection mechanism in S4 determines an abnormal state, use the trained deep matching network model to output the current anomaly representation vector With all the trigger labels vector According to the correlation matching score, all inducement label vectors are sorted in descending order according to the matching score results, and the top K inducement label vectors are selected as the candidate inducement set of the current anomaly; S6, Abnormality cause and effect path reasoning: according to the top K cause label vectors matched in S5, the corresponding abnormality cause nodes in the knowledge graph are obtained In the knowledge graph, depth-first search is performed according to the preset path depth L, and a plurality of candidate cause and effect paths with the cause node as the source point and the target abnormality as the end point are constructed. A node correlation scoring mechanism is used to calculate the overall semantic correlation score of each candidate cause and effect path. According to the scoring result, all candidate paths are sorted in descending order, and the top N paths are selected as the explainable root cause path set of the current abnormality. The path with the highest score is selected as the most representative cause and effect reasoning result. 2.The method of claim 1, wherein: The preprocessing operations in S1 include timestamp synchronization, missing value processing, outlier removal, and field standardization. 3.The method of claim 2, wherein: The knowledge graph constructed in S2 is subjected to graph cleaning, semantic fusion, and graph structure optimization operations. 4.The method of claim 3, wherein: In S4, the steps of constructing a deep time series feature extraction network and pre-training include: S41, constructing a deep time sequence feature extraction network mapping the flight control state time sequence data into a flight control state abnormality representation vector : wherein for each flight procedure acquired, flight control state time series data formed over a continuous time window, denotes the state feature vector at the th time step; S42, according to the training samples obtained in S1, a flight task is constructed for typical abnormal patterns based on a flight control simulation platform, simulation flight control state time series data and corresponding injected text information are collected to form simulation training samples, and the training samples obtained in S1 and the simulation training samples are combined to form an abnormal representation training sample set with class labels, wherein the abnormal representation training sample set includes samples of normal flight states and various abnormal states; S43, the deep time series feature extraction network is trained through the abnormal representation training sample set.

5. The method of claim 4, wherein: In S5, the steps of constructing a deep matching network model based on the abnormal representation vector and the inducement label vector and pre-training include: S51, based on the anomaly representation vector and the cause label vector constructing a deep matching network model for representing the association degree between the anomaly representation vector and the cause label vector ​ wherein represents the th abnormality cause semantic entity generated cause label vector, for representing the current abnormality representation vector between the cause label vector and the relevance matching score, is a deep matching model, is a set of learnable parameters; S52, according to the training samples obtained in S1, each training sample obtains an abnormal representation vector through a deep time series feature extraction network and an inducement label vector through a knowledge graph, and the abnormal representation vector of each training sample, the corresponding correct inducement label vector, and other interference label vectors constitute a matching training sample; S53, the deep time series feature extraction network is trained through the matching training sample set, and a supervised training method is used for optimization, and the loss function adopts cross-entropy loss.

6. The method of claim 5, wherein: In S6, the node correlation scoring mechanism is used to calculate the overall semantic correlation score of each candidate causal path, and the formula is as follows: wherein represents the candidate paths, for representing the current abnormality representation vector and the cause label vector between the abnormality representation vector represents a preset path depth.

Images