An unmanned aerial vehicle failure mode analysis method, device, equipment and storage medium

Abstract

The application relates to an unmanned aerial vehicle failure mode analysis method, device, equipment and storage medium, which comprises the following steps: constructing an unmanned aerial vehicle failure domain knowledge graph by fusing structured flight log data and unstructured maintenance text data; based on the failure domain knowledge graph, performing probability graph reasoning on abnormal events detected in real-time flight data flow to determine a fault propagation path and a fault cause; and generating a human-computer interactive failure analysis report by using a domain large model according to user input failure phenomenon description, the fault propagation path and the fault cause. The application effectively integrates multi-source heterogeneous data, improves the completeness and systematicness of knowledge, can clearly reveal the cause-and-effect chain and confidence of fault propagation, realizes intelligent closed loop from data analysis to decision support, and changes the traditional failure analysis mode which depends on manual work, is low in efficiency and insufficient in adaptability.

Description

Technical Field

[0001] This invention relates to the field of fault analysis technology, and in particular to a method, apparatus, device, and storage medium for analyzing the failure modes of unmanned aerial vehicles (UAVs). Background Technology

[0002] Unmanned Aerial Vehicles (UAVs) have been widely used in many key fields such as logistics and transportation, agricultural plant protection, and disaster relief. As application scenarios continue to deepen, their reliability requirements are increasing. UAV systems have complex structures, involving multiple subsystems such as flight control, power, and communication. The potential failure of any component can trigger a chain reaction, leading to mission failure or even safety accidents. Therefore, achieving rapid and accurate analysis and localization of UAV failure modes is of great significance for ensuring flight safety and improving operational efficiency.

[0003] Currently, the analysis of failure modes of unmanned aerial vehicles (UAVs) mainly relies on the following traditional methods: First, Failure Mode and Effects Analysis (FMEA) based on expert experience. This method manually enumerates potential failure modes and assesses their impact, but it heavily relies on expert knowledge, is time-consuming and labor-intensive, and is difficult to update dynamically. Second, data-driven diagnostic methods, such as using machine learning models to detect anomalies in flight logs. Although these methods can detect statistical anomalies, they often lack interpretability and are difficult to clearly reveal the causal transmission mechanism between failure modes. Third, rule-based expert systems. Their diagnostic logic relies on pre-defined "if-then" rules. When faced with complex and novel compound failures, the rule base has limited coverage and insufficient adaptability.

[0004] The aforementioned existing technologies generally suffer from a core deficiency: they fail to achieve deep fusion of multi-source heterogeneous data and structured representation of knowledge, resulting in insufficient capabilities of failure mode analysis in terms of knowledge completeness, interpretability of reasoning, and intelligent human-computer interaction. Specifically, the structured time-series data in flight logs and the unstructured textual knowledge in maintenance reports are disconnected, unable to be effectively correlated and constructed into a unified knowledge model that allows for deep machine reasoning. This limits existing analytical methods to either superficial statistical correlations or rigid rule frameworks, making it impossible to automatically uncover the causal propagation paths hidden behind failure events, or to generate comprehensive analytical reports that are easy for maintenance personnel to understand and use. Summary of the Invention

[0005] In view of this, it is necessary to provide a method, apparatus, device and storage medium for analyzing the failure modes of unmanned aerial vehicles (UAVs) to solve the problem that existing technologies have failed to achieve deep fusion of multi-source heterogeneous data and structured expression of knowledge, resulting in insufficient capabilities of failure mode analysis in terms of knowledge integrity, reasoning interpretability and human-computer interaction intelligence.

[0006] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for analyzing the failure modes of unmanned aerial vehicles (UAVs), comprising: By fusing structured flight log data with unstructured maintenance text data, a failure domain knowledge graph for UAVs is constructed. Based on the failure domain knowledge graph, probabilistic graphical reasoning is performed on abnormal events detected in real-time flight data streams to determine the failure propagation path and failure cause; Based on the user-input description of the failure phenomenon, the failure propagation path, and the cause of the failure, an interactive failure analysis report is generated using a domain-wide model.

[0007] In one possible implementation, a failure domain knowledge graph for the UAV is constructed by fusing structured flight log data with unstructured maintenance text data, including: The multidimensional time-series data in the flight log is symbolically discretized, and a single flight mission is constructed as a directed graph with discretized state events as nodes and the temporal relationships between events as edges; A frequent subgraph mining algorithm is used to extract frequently connected subgraphs with support exceeding a preset threshold from directed graphs of multiple flight missions as fault propagation patterns. Perform named entity recognition and relation extraction on maintenance text data to construct semantic relation triples between entities; The initial knowledge graph is obtained by fusing frequent connected subgraphs with semantic relation triples. Assign conditional probability weights to the relation edges in the initial knowledge graph based on historical data statistical calculations to obtain the knowledge graph of the failed domain.

[0008] In one possible implementation, a frequent subgraph mining algorithm is used to extract frequently connected subgraphs with support exceeding a preset threshold from the directed graphs of multiple flight missions as fault propagation patterns, including: Search for all frequently connected subgraphs that satisfy the preset minimum number of nodes from the set of directed graphs from multiple flight missions, and use them as candidate graph patterns; Count the number of times each candidate graph pattern appears in the entire set of directed graphs and calculate its support. Candidate graph patterns whose support exceeds a preset threshold are identified and output as frequently connected subgraphs representing fault propagation patterns.

[0009] In one possible implementation, conditional probability weights calculated based on historical data are assigned to the relation edges in the initial knowledge graph to obtain a failure domain knowledge graph, including: Based on the historical data covered by the initial knowledge graph, count the number of times any two entities with a causal relationship co-occur in the historical records; For each relation edge pointing from the head entity to the tail entity, its weight value is calculated as the conditional probability of the tail entity event occurring given that the head entity event has already occurred. Using the calculated conditional probabilities as the weights of the relation edges, we obtain the knowledge graph of the failed domain.

[0010] In one possible implementation, based on a failure domain knowledge graph, probabilistic graphical reasoning is performed on anomaly events detected in the real-time flight data stream to determine the failure propagation path and cause of the failure, including: Using abnormal events as evidence nodes, all connected failure paths are traversed and retrieved in the failure domain knowledge graph; For each retrieved path, the conditional probability weights of each edge on the path are multiplied together to calculate the propagation probability of that path. Based on Bayes' theorem, the posterior probability of each potential root cause of failure is calculated under given anomalous evidence. The retrieved fault propagation paths and root causes are sorted according to the propagation probability and posterior probability to obtain a fault hypothesis list sorted by risk level, thus determining the fault propagation path and fault cause.

[0011] In one possible implementation, based on the user-input description of the failure phenomenon, the fault propagation path, and the cause of the failure, a large domain model is used to generate an interactive failure analysis report, including: By fusing information such as fault propagation path, fault cause and its corresponding probability with the description of failure phenomenon, structured prompt words are constructed. By inputting structured prompts into the domain-wide model, a comprehensive natural language report is generated, which includes at least the interpretation of fault phenomena, causal chain analysis, risk level assessment, and maintenance recommendations. If a follow-up question is received from a user, the question is combined with relevant contextual information in the failure domain knowledge graph to form a new prompt word, which is then input into the domain big model to obtain a failure analysis report.

[0012] In one possible implementation, the domain-wide model is obtained through domain adaptation training via the following steps: Acquire professional textual data in the field of drone failure, including at least maintenance manuals, failure case libraries, and technical specifications; Using the knowledge graph of the failed domain as a structured knowledge source, question-answer pairs containing graph entities, relationships and reasoning paths are constructed as training samples. Using professional text corpora and training samples, a basic large language model is trained to obtain a domain-specific large model with the ability to understand and reason about drone failures.

[0013] Secondly, the present invention also provides an analysis apparatus for unmanned aerial vehicle (UAV) failure modes, comprising: The graph construction module is configured to construct a failure domain knowledge graph for UAVs by fusing structured flight log data with unstructured maintenance text data. The anomaly reasoning module is configured to perform probabilistic graphical reasoning on anomaly events detected in real-time flight data streams based on a failure domain knowledge graph to determine the fault propagation path and fault cause. The fault analysis module is configured to generate an interactive fault analysis report based on the user-input description of the failure phenomenon, the fault propagation path, and the fault cause, using a large domain model.

[0014] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein, Memory, used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the UAV failure mode analysis method in any of the above implementations.

[0015] Fourthly, the present invention also provides a computer-readable storage medium for storing a computer-readable program or instruction, which, when executed by a processor, can implement the steps in the UAV failure mode analysis method in any of the above implementations.

[0016] This invention provides a method for analyzing the failure modes of unmanned aerial vehicles (UAVs). By fusing structured flight logs and unstructured maintenance texts, a unified and quantifiable failure domain knowledge graph is constructed. This achieves deep fusion and systematic modeling of multi-source heterogeneous data, integrating previously scattered and fragmented numerical signals and expert experience into a structured relational network, fundamentally solving the problems of fragmented and unreusable failure mode knowledge. Based on the failure domain knowledge graph, probabilistic graphical reasoning is performed on abnormal events in real-time flight data streams. This automatically mines and identifies potential fault propagation paths and their root causes, revealing not only the causal transmission mechanism between faults but also quantifying the magnitude of risk through probability weights. This overcomes the shortcomings of traditional black-box models, such as poor interpretability and limited coverage of rule systems. Finally, combined with a large domain model, an interactive failure analysis report is generated, transforming complex graph reasoning results into easily understandable and usable professional-grade natural language descriptions. This lowers the technical threshold and analysis time for maintenance personnel, achieving an end-to-end intelligent analysis closed loop from data perception and intelligent reasoning to decision support. Attached Figure Description

[0017] Figure 1 A flowchart illustrating an embodiment of the method for analyzing the failure modes of unmanned aerial vehicles provided by the present invention; Figure 2 Provided by the present invention Figure 1 A schematic flowchart of an embodiment of step S101; Figure 3 A schematic diagram of the structure of an embodiment of the unmanned aerial vehicle (UAV) failure mode analysis device provided by the present invention; Figure 4 This is a schematic diagram of the structure of the drone failure mode analysis device provided in an embodiment of the present invention. Detailed Implementation

[0018] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0019] In the description of this application, "multiple" means two or more, unless otherwise expressly and specifically defined.

[0020] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0021] This invention provides a method, apparatus, device, and storage medium for analyzing the failure modes of unmanned aerial vehicles (UAVs), which will be described below.

[0022] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the drone failure mode analysis method provided by the present invention. A specific embodiment of the present invention discloses a drone failure mode analysis method, comprising: S101. By fusing structured flight log data and unstructured maintenance text data, a failure domain knowledge graph for UAVs is constructed. S102. Based on the failure domain knowledge graph, perform probabilistic graphical reasoning on abnormal events detected in the real-time flight data stream to determine the fault propagation path and fault cause; S103. Based on the user-input description of the failure phenomenon, the failure propagation path, and the cause of the failure, generate an interactive failure analysis report using a large domain model.

[0023] In the above embodiments, the processing of structured flight logs involves: extracting multi-dimensional time-series data such as voltage, current, temperature, and rotational speed from the flight logs; cleaning and aligning the data; and using a symbolic aggregation approximation method to discretize continuous numerical values ​​into symbolic sequences to represent different state events. Based on the discretization results, a single flight mission is constructed as a directed graph, where nodes represent state events and directed edges represent the temporal relationships between events. The processing of unstructured maintenance text involves: using natural language processing techniques to extract component, fault, and symptom entities from the text; and using a relation extraction model to extract semantic relationships between entities from sentences, forming triples of (head entity, relation, tail entity). The frequent event sequences mined from the flight logs are fused with the semantic triples extracted from the maintenance text to construct a unified domain knowledge graph in the graph database. Conditional probability weights based on historical data statistical calculations are assigned to the relation edges in the graph, thereby enabling the graph to possess quantitative reasoning capabilities.

[0024] Continuous monitoring of real-time flight data streams allows for the direct identification of anomalous events using existing anomaly detection algorithms. Starting with detected anomalous events, a traversal search is performed within the knowledge graph to retrieve all connected potential failure propagation paths. Treating the knowledge graph as a probabilistic graphical model, for each retrieved failure propagation path, the conditional probability weights of all edges along the path are multiplied to calculate the overall propagation probability of that path. Simultaneously, based on Bayes' theorem, the posterior probability of each potential root cause of the failure is calculated given evidence of the anomalous event. The calculated path probabilities and posterior probabilities are then ranked, and the failure propagation path with the highest probability and its root cause are identified as the final analysis results and output.

[0025] The user-input natural language description and the structured results obtained from the S102 step reasoning are integrated to construct a structured prompt word. The obtained prompt word is input into a domain-specific large model that has been fine-tuned with UAV-related text and knowledge. Through command control, the large model is guided to generate a comprehensive analysis report. The report includes, but is not limited to, interpretation of fault phenomena, root cause reasoning, propagation path analysis, risk level assessment, and specific maintenance recommendations.

[0026] In addition, the present invention can also support users to ask follow-up questions about the report content based on the interactive interface, receive the user's follow-up questions, and generate new prompt words by combining the user's follow-up questions with the knowledge graph context, thereby driving the large model to supplement or correct the report and realize true human-computer interactive analysis.

[0027] Compared with existing technologies, this embodiment provides a method for analyzing UAV failure modes. By fusing structured flight logs and unstructured maintenance texts, it constructs a unified, quantifiable, and reasonable failure domain knowledge graph. This achieves deep fusion and systematic modeling of multi-source heterogeneous data, integrating previously scattered and fragmented numerical signals and expert experience into a structured relational network, fundamentally solving the problems of fragmented and unreusable failure mode knowledge. Based on the failure domain knowledge graph, probabilistic graphical reasoning is performed on abnormal events in real-time flight data streams. This automatically mines and identifies potential fault propagation paths and their root causes, revealing not only the causal transmission mechanism between faults but also quantifying the risk magnitude through probability weights. This overcomes the shortcomings of traditional black-box models, such as poor interpretability and limited coverage of rule systems. Finally, combined with a large domain model, an interactive failure analysis report is generated, transforming complex graph reasoning results into easily understandable and usable professional-grade natural language descriptions. This lowers the technical threshold and analysis time for maintenance personnel, achieving an end-to-end intelligent analysis closed loop from data perception and intelligent reasoning to decision support.

[0028] Please see Figure 2 , Figure 2 Provided by the present invention Figure 1 A flowchart illustrating an embodiment of step S101. In some embodiments of the present invention, a failure domain knowledge graph of the UAV is constructed by fusing structured flight log data and unstructured maintenance text data, including: S201. The multidimensional time-series data in the flight log is symbolically discretized, and a single flight mission is constructed as a directed graph with discretized state events as nodes and time-series relationships between events as edges. S202. The frequent subgraph mining algorithm is used to extract frequently connected subgraphs with support exceeding a preset threshold from the directed graphs of multiple flight missions as fault propagation modes. S203. Perform named entity recognition and relation extraction on the maintenance text data, and construct semantic relation triples between entities; S204. The frequent connected subgraphs and semantic relation triples are fused to obtain the initial knowledge graph. S205. Assign conditional probability weights to the relation edges in the initial knowledge graph based on historical data statistical calculations to obtain the failure domain knowledge graph.

[0029] In the above embodiments, the multidimensional time-series data in the flight logs, including but not limited to voltage, current, temperature, and rotational speed, are cleaned, aligned, and normalized to eliminate noise and dimensional effects. A symbolic aggregation approximation method is used to segment each dimension of the processed continuous time-series data into segments of equal or unequal length, and a symbol is used to represent the aggregation characteristics of each segment, thereby transforming the continuous numerical sequence into a discrete symbolic sequence. The system state represented by each symbol is defined as an event node. If event A occurs and event B occurs within a set time window Δt, a directed edge is created from node A to node B, representing a temporal sequence relationship, thus constructing a directed graph model for a single flight mission.

[0030] The directed graphs from the multiple flight missions are compiled into a graph dataset. Frequent subgraph mining algorithms, such as gSpan, can be used to search for all connected subgraphs in the dataset that meet a preset minimum support threshold. The mined frequent connected subgraphs, especially those chain-like structures containing nodes from different sensor domains (such as the power system and flight control system) and whose edges have clear temporal directions, are identified as potential, physically meaningful cross-system fault propagation patterns.

[0031] Using a named entity recognition model based on a pre-trained language model (such as BERT), key entities representing components, failure modes, phenomena, and measures are automatically identified and extracted from unstructured historical maintenance reports, fault records, and other texts. Relation extraction techniques based on dependency parsing or end-to-end models are employed to extract semantic relationships between entities from sentences containing these entities, forming triples in the format (head entity, relation, tail entity).

[0032] Based on entity names, thesaurus, or embedded representation similarity, nodes representing state events in frequent subgraphs are aligned and merged with entities extracted from the text to eliminate synonyms and homonyms. Edges representing temporal relationships in frequent subgraphs are fused and unified with semantic relationships extracted from the text to construct a unified pattern library that combines numerical statistical patterns with expert semantic descriptions. The fused triples and frequent patterns are organized according to predefined UAV domain ontology (such as categories and relationships of components, faults, phenomena, consequences, etc.) to construct an initial knowledge graph pattern in a graph database.

[0033] Based on all historical data covered by the initial knowledge graph, the number of times any two entities with a direct causal or relational relationship co-occur in the historical records is counted. For each relation edge in the knowledge graph pointing from a head entity to a tail entity, its weight value is calculated as the conditional probability of the tail entity event occurring given that the head entity event has already occurred. The calculated conditional probability value is used as the weight of the relation edge, assigned and stored in the knowledge graph, thereby upgrading the graph from a static relational network to a dynamic probabilistic graphical model with quantitative reasoning capabilities.

[0034] To enrich and refine the patterns mined from purely numerical data, the solution simultaneously mines unstructured historical maintenance case texts. Utilizing natural language processing techniques, it first extracts key entities from the text using Named Entity Recognition (NER) models (such as those based on BERT-BiLSTM-CRF architecture), such as components ("motor," "battery"), faults ("seized," "overheating"), and symptoms ("abnormal current," "jitter"). Next, it employs Relation Extraction models (such as those based on dependency parsing or the pre-trained REBEL model) to extract semantic relationships between entities from sentences, forming (head entity, relation, tail entity) triples (e.g., extracting (motor, fault status, seized) from "inspection found motor seized"). These triples extracted from the text are then fused with frequent patterns mined from flight logs (which can be considered a type of triple) to construct a unified, structured domain knowledge graph. The edges of this graph not only represent associations, but are also assigned weights. The weight values ​​are calculated based on the conditional probability P (effect | cause) of historical data statistics, that is, Weight(A->B) = P(B|A) = Count(A and B co-occur) / Count(A), which gives the graph the ability to quantify reasoning.

[0035] In some embodiments of the present invention, a frequent subgraph mining algorithm is used to extract frequently connected subgraphs with support exceeding a preset threshold from directed graphs of multiple flight missions as fault propagation patterns, including: Search for all frequently connected subgraphs that satisfy the preset minimum number of nodes from the set of directed graphs from multiple flight missions, and use them as candidate graph patterns; Count the number of times each candidate graph pattern appears in the entire set of directed graphs and calculate its support. Candidate graph patterns whose support exceeds a preset threshold are identified and output as frequently connected subgraphs representing fault propagation patterns.

[0036] In the above embodiments, the search begins with the simplest single event node, and more complex graph patterns are generated by gradually adding connected edges and nodes. During this process, a minimum number of nodes is set to filter out overly simple patterns that lack analytical value. All connected subgraphs that meet this node count requirement and appear at least once in the entire graph set are listed as candidate graph patterns.

[0037] Support refers to the frequency with which a candidate map pattern appears across all flight mission map sets. To assess the prevalence and importance of each candidate map pattern, its support needs to be calculated. Specifically, this can be calculated by counting how many different flight mission maps the pattern appears on and then dividing by the total number of flight mission maps. Support reflects the prevalence of the failure propagation pattern's recurrence in historical flights.

[0038] The support calculated for each candidate graph pattern is compared with a pre-set support threshold. All candidate graph patterns whose support reaches or exceeds this threshold are ultimately determined to be statistically significant fault propagation patterns and are output as key knowledge. The frequently connected subgraphs selected from this process intuitively reveal the transmission chain of recurring fault events in the UAV swarm. For example, "abnormal motor current leads to increased controller temperature, which in turn causes flight attitude jitter" is a typical result.

[0039] In some embodiments of the present invention, conditional probability weights calculated based on historical data are assigned to the relation edges in the initial knowledge graph to obtain a failure domain knowledge graph, including: Based on the historical data covered by the initial knowledge graph, count the number of times any two entities with a causal relationship co-occur in the historical records; For each relation edge pointing from the head entity to the tail entity, its weight value is calculated as the conditional probability of the tail entity event occurring given that the head entity event has already occurred. Using the calculated conditional probabilities as the weights of the relation edges, we obtain the knowledge graph of the failed domain.

[0040] In the above embodiments, based on all historical data covered by the initial knowledge graph, co-occurrence statistics are performed on any two entities in the graph that have a causal relationship. Specifically, the historical records are traversed, and the number of times that a subsequent failure or event represented by the tail entity also occurs simultaneously or within a certain time window when the failure or event represented by the head entity occurs is counted, thereby obtaining the co-occurrence frequency of these two entities in history. It can be understood that such a head entity and a tail entity are two entities with a causal relationship.

[0041] For each directed edge in the knowledge graph that points from a head entity to a tail entity, its weight is calculated as a conditional probability, representing the likelihood of the tail entity event occurring given that the head entity event has already occurred. The specific calculation method involves dividing the number of times the two entities co-occur in history by the total number of times the head entity appears alone, resulting in a probability value between 0 and 1, used to quantify the strength of the causal relationship.

[0042] The calculated conditional probability values ​​are used as weights for the corresponding edges and assigned to the initial knowledge graph. Through this weight assignment process, the initial knowledge graph is enhanced into a failure domain knowledge graph with quantitative reasoning capabilities. Each edge not only represents the causal relationship between failures but also reflects the credibility and strength of the relationship through probability weights.

[0043] In some embodiments of the present invention, based on a failure domain knowledge graph, probabilistic graphical reasoning is performed on abnormal events detected in real-time flight data streams to determine the fault propagation path and fault cause, including: Using abnormal events as evidence nodes, all connected failure paths are traversed and retrieved in the failure domain knowledge graph; For each retrieved path, the conditional probability weights of each edge on the path are multiplied together to calculate the propagation probability of that path. Based on Bayes' theorem, the posterior probability of each potential root cause of failure is calculated under given anomalous evidence. The retrieved fault propagation paths and root causes are sorted according to the propagation probability and posterior probability to obtain a fault hypothesis list sorted by risk level, thus determining the fault propagation path and fault cause.

[0044] In the above embodiments, the process of traversing and searching in the failure domain knowledge graph starts with the abnormal event and searches layer by layer along the direction of all directed relation edges in the knowledge graph to find all possible upstream failure causes and downstream impact consequences, thereby finding multiple complete potential failure propagation paths.

[0045] For each retrieved fault propagation path, the likelihood of the path occurring is assessed by calculating its overall propagation probability. Specifically, the conditional probability weights of each relation edge along the path from the initial fault to the final abnormal event are multiplied together, and the resulting product is the overall propagation probability of the fault propagation path, which intuitively reflects the probability that the fault chain will occur completely from beginning to end.

[0046] Simultaneously, to determine the root cause of the anomalous event, probabilistic reasoning is performed based on Bayes' theorem. The posterior probability of each potential root cause of the failure is calculated given the currently observed evidence of the anomalous event. This calculation requires considering both the prior probability of each cause occurring in historical data and the conditional probability that that cause leads to the current evidence of the anomalous event.

[0047] All retrieved fault propagation paths and their corresponding root causes are sorted from high to low according to their propagation probability and posterior probability, generating a fault hypothesis list arranged by risk level. This helps operations and maintenance personnel quickly locate the most likely fault propagation path and root cause.

[0048] This phase leverages a failure mode knowledge graph to analyze and mine failure modes. The UAV flight logs contain rich, multi-dimensional time-series data, such as voltage, current, temperature, rotational speed, and GPS coordinates. To discover high-frequency correlation patterns, such as "battery overheating leading to voltage drop, which in turn causes communication interruption," a frequent subgraph mining method is employed. The time-series data is preprocessed, including cleaning and alignment, and continuous numerical values ​​are discretized into symbol sequences using symbol aggregation approximation and threshold-based segmentation methods.

[0049] Subsequently, a single flight mission is constructed as a directed graph g(V, E), where vertex V represents a discretized state event, and edge E represents the temporal relationship between events. Specifically, if event B occurs within a time interval Δt after event A, a directed edge is created pointing from A to B. Based on this, the gSpan frequent subgraph mining algorithm is used to search for all frequently connected subgraphs with support exceeding a preset threshold in the entire flight graph dataset D ={G1, G2, ..., Gn}. This is the number of unique nodes in graph g mapped to by nodes based on pattern p. This method avoids the computation of MIS (Maximum Independent Set).

[0050] ; In the formula, σ ^ (p,g) represents the pattern support defined using the Minimum Number of Node Mappings (MNI), where V(p) is the set of nodes for pattern p. This represents a first-order subgraph isomorphism of pattern p in graph g. For pattern nodes v The graph node corresponding to the i-th mapping. This support is defined by taking the minimum number of distinct graph nodes mapped to by each node across all occurrences of the pattern, thus avoiding the computational complexity of the maximum independent set problem. The discovered high-frequency subgraphs represent potential fault propagation patterns with temporal order.

[0051] Once the system is operational in real-time, its inference module continuously monitors the real-time flight data stream. When an anomaly event E_t, such as "abnormal motor current," is detected based on the anomaly detection algorithm, the inference engine is immediately triggered. The engine uses E_t as the starting point for its query and traverses the pre-built knowledge graph, retrieving all possible fault paths connected to it. Subsequently, the system treats the knowledge graph as a probabilistic graphical model and employs a confidence propagation algorithm for probabilistic inference. For each possible path retrieved, such as "motor jamming → insufficient thrust → decreased flight altitude → attitude instability," the system calculates its propagation probability. The probability of this path is obtained by multiplying the conditional probabilities of all edges k on the path, i.e.: P(path) = P(N1 | E_t) * P(N2 | N1) * ... * P(Nk | N_{k-1}).

[0052] Simultaneously, to assess the likelihood of a fundamental failure, the system calculates the posterior probability of a certain failure cause F_i given the evidence E_t, based on Bayes' theorem: P(F_i | E_t) = [P(E_t | F_i) * P(F_i)] / [∑_k P(E_t | F_k) * P(F_k)], Where P(F_i) is the prior probability obtained statistically from historical data. Based on the failure mode map generated by the system, the currently detected modes and their possible propagation paths are highlighted. Experts can confirm or revise the system's recommendations, and the feedback results are written into the knowledge base to achieve iterative optimization.

[0053] In some embodiments of the present invention, based on the user-input description of the failure phenomenon, the fault propagation path, and the fault cause, a human-computer interactive failure analysis report is generated using a large domain model, including: By fusing information such as fault propagation path, fault cause and its corresponding probability with the description of failure phenomenon, structured prompt words are constructed. By inputting structured prompts into the domain-wide model, a comprehensive natural language report is generated, which includes at least the interpretation of fault phenomena, causal chain analysis, risk level assessment, and maintenance recommendations. If a follow-up question is received from a user, the question is combined with relevant contextual information in the failure domain knowledge graph to form a new prompt word, which is then input into the domain big model to obtain a failure analysis report.

[0054] In the above embodiments, the fault propagation path obtained through reasoning, the identified root cause of the fault and its corresponding probability of occurrence are integrated with the user-input description of the failure phenomenon. This structured and unstructured information is organized into structured prompts that include task background, key evidence and generation requirements, providing standardized input content for subsequent report generation.

[0055] The constructed structured prompts are input into a specially trained domain-specific large-scale model. Based on its powerful language understanding and domain knowledge, the large-scale model can generate a professional and comprehensive natural language analysis report. This report includes at least a detailed interpretation of the fault phenomenon, causal chain analysis based on graph reasoning, a risk level assessment incorporating probability, and specific and feasible maintenance operation suggestions.

[0056] Furthermore, this invention also supports multi-turn interactive intelligent question answering. If a user raises new follow-up questions after reading the report, the system will retrieve highly relevant entities, attributes, and historical cases from the failure domain knowledge graph. This contextual information will be combined with the user's follow-up questions to form prompts containing new knowledge, which will then be input again into the domain-wide model. In this way, a supplementary analysis report that accurately responds to follow-up questions can ultimately be generated.

[0057] In some embodiments of the present invention, the domain-large model is obtained through domain adaptation training via the following steps: Acquire professional textual data in the field of drone failure, including at least maintenance manuals, failure case libraries, and technical specifications; Using the knowledge graph of the failed domain as a structured knowledge source, question-answer pairs containing graph entities, relationships and reasoning paths are constructed as training samples. Using professional text corpora and training samples, a basic large language model is trained to obtain a domain-specific large model with the ability to understand and reason about drone failures.

[0058] In the above embodiments, in order to construct a high-quality training dataset, the professional text corpus in the field of UAV failure should at least include professional documents such as official maintenance manuals, historical failure case libraries, and equipment technical specifications to ensure coverage of possible failure modes, diagnostic methods, and solutions for each subsystem of the UAV.

[0059] Using the existing failure domain knowledge graph as a structured knowledge source, high-quality question-answer pairs are constructed as training samples through a specific transformation mechanism. Specifically, key entity nodes, relation edges, and complete reasoning paths are extracted from the knowledge graph and transformed into training data in the form of "question-answer". For example, based on the relationship "motor-cause-overheating", a question-answer pair "What causes the motor to overheat?" is generated, thereby enabling the model to learn structured domain logic and reasoning patterns.

[0060] By combining collected professional text corpora with constructed graph-based question-answering pairs as training samples, a general-purpose basic large language model is subjected to supervised domain adaptation training. Through multiple rounds of iterative optimization, the model gradually masters the professional terminology, fault reasoning logic, and analytical framework of the UAV failure domain, ultimately obtaining a specialized domain-specific large model with deep domain knowledge understanding and accurate reasoning capabilities. It is understood that the specific training process can be implemented using existing technologies, and this invention will not elaborate on this further.

[0061] To better implement the UAV failure mode analysis method in this embodiment of the invention, based on the UAV failure mode analysis method, please refer to the corresponding documentation. Figure 3 , Figure 3 This is a schematic diagram of an embodiment of the UAV failure mode analysis device provided by the present invention. The embodiment of the present invention provides a UAV failure mode analysis device 300, comprising: The graph construction module 310 is configured to construct a failure domain knowledge graph for UAVs by fusing structured flight log data with unstructured maintenance text data. The anomaly reasoning module 320 is configured to perform probabilistic graphical reasoning on anomaly events detected in real-time flight data streams based on a failure domain knowledge graph to determine the fault propagation path and fault cause. The fault analysis module 330 is configured to generate an interactive fault analysis report based on the user-input description of the failure phenomenon, the fault propagation path, and the fault cause, using a large domain model.

[0062] It should be noted that the device 300 provided in the above embodiments can implement the technical solutions described in the above method embodiments. The specific implementation principles of the above modules or units can be found in the corresponding content in the above method embodiments, and will not be repeated here.

[0063] Please see Figure 4 , Figure 4This is a schematic diagram of the structure of a drone failure mode analysis device provided in an embodiment of the present invention. Based on the above-described drone failure mode analysis method, the present invention also provides a drone failure mode analysis device, which can be a computing device such as a mobile terminal, desktop computer, laptop, handheld computer, or server. The drone failure mode analysis device 400 includes a processor 410, a memory 420, and a display 430. Figure 4 Only some components of the drone failure mode analysis device are shown; however, it should be understood that implementation of all shown components is not required, and more or fewer components may be implemented instead.

[0064] In some embodiments, memory 420 may be an internal storage unit of the drone failure mode analysis device 400, such as a hard disk or memory of the drone failure mode analysis device 400. In other embodiments, memory 420 may be an external storage device of the drone failure mode analysis device 400, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the drone failure mode analysis device 400. Furthermore, memory 420 may include both internal and external storage units of the drone failure mode analysis device 400. Memory 420 is used to store application software and various types of data installed on the drone failure mode analysis device 400, such as the program code of the drone failure mode analysis device 400. Memory 420 may also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 420 stores a drone failure mode analysis program 440, which can be executed by the processor 410 to implement the drone failure mode analysis method of the various embodiments of this application.

[0065] In some embodiments, processor 410 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 420 or process data, such as executing a method for analyzing the failure modes of a drone.

[0066] In some embodiments, display 430 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 430 is used to display information from the UAV failure mode analysis device 400 and to display a user interface for visualization. Components 410-430 of the UAV failure mode analysis device 400 communicate with each other via a system bus.

[0067] In one embodiment, when the processor 410 executes the drone failure mode analysis program 440 in the memory 420, the steps in the drone failure mode analysis method described above are implemented.

[0068] This embodiment also provides a computer-readable storage medium storing a drone failure mode analysis program, which, when executed by a processor, performs the following steps: By fusing structured flight log data with unstructured maintenance text data, a failure domain knowledge graph for UAVs is constructed. Based on the failure domain knowledge graph, probabilistic graphical reasoning is performed on abnormal events detected in real-time flight data streams to determine the failure propagation path and failure cause; Based on the user-input description of the failure phenomenon, the failure propagation path, and the cause of the failure, an interactive failure analysis report is generated using a domain-wide model.

[0069] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0070] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0071] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0072] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0073] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0074] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0075] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0076] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0077] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of embodiments of this disclosure upon considering the specification and practicing the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

[0078] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0079] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for analyzing the failure modes of unmanned aerial vehicles (UAVs), characterized in that, include: By fusing structured flight log data with unstructured maintenance text data, a failure domain knowledge graph for UAVs is constructed. Based on the aforementioned failure domain knowledge graph, probabilistic graphical reasoning is performed on abnormal events detected in the real-time flight data stream to determine the fault propagation path and fault cause. Based on the user-input description of the failure phenomenon, the failure propagation path, and the failure cause, an interactive failure analysis report is generated using a large domain model.

2. The method for analyzing the failure modes of unmanned aerial vehicles according to claim 1, characterized in that, The method involves fusing structured flight log data with unstructured maintenance text data to construct a failure domain knowledge graph for the UAV, including: The multidimensional time-series data in the flight log is symbolically discretized, and a single flight mission is constructed as a directed graph with discretized state events as nodes and time-series relationships between events as edges; A frequent subgraph mining algorithm is used to extract frequently connected subgraphs with support exceeding a preset threshold from directed graphs of multiple flight missions as fault propagation patterns. Named entity recognition and relation extraction are performed on the maintenance text data to construct semantic relation triples between entities; The initial knowledge graph is obtained by fusing the frequently connected subgraph with the semantic relation triples. Assign conditional probability weights based on historical data statistical calculations to the relation edges in the initial knowledge graph to obtain the failure domain knowledge graph.

3. The method for analyzing the failure modes of unmanned aerial vehicles according to claim 2, characterized in that, The method employs a frequent subgraph mining algorithm to extract frequently connected subgraphs with support exceeding a preset threshold from directed graphs across multiple flight missions, serving as fault propagation patterns. This includes: Search the directed graph set of the multiple flight missions for all frequently connected subgraphs that satisfy the preset minimum number of nodes, and use them as candidate graph patterns; Count the number of times each candidate graph pattern appears in the entire set of directed graphs, and calculate its support. Candidate graph patterns whose support exceeds the preset threshold are identified and output as frequently connected subgraphs representing fault propagation patterns.

4. The method for analyzing the failure modes of unmanned aerial vehicles according to claim 2, characterized in that, Assigning conditional probability weights based on historical data statistical calculations to the relation edges in the initial knowledge graph to obtain the failure domain knowledge graph includes: Based on the historical data covered by the initial knowledge graph, the number of times any two entities with a causal relationship co-occur in the historical records is counted. For each relation edge pointing from the head entity to the tail entity, its weight value is calculated as the conditional probability of the tail entity event occurring given that the head entity event has already occurred. Using the calculated conditional probabilities as the weights of the relation edges, we obtain the knowledge graph of the failed domain.

5. The method for analyzing the failure modes of unmanned aerial vehicles according to claim 1, characterized in that, The process of using the failure domain knowledge graph to perform probabilistic graphical reasoning on abnormal events detected in the real-time flight data stream to determine the fault propagation path and fault cause includes: Using the abnormal event as an evidence node, all connected failure paths are traversed and retrieved in the failure domain knowledge graph; For each retrieved path, the conditional probability weights of each edge on the path are multiplied together to calculate the propagation probability of that path. Based on Bayes' theorem, the posterior probability of each potential root cause of failure is calculated under given anomalous evidence. The retrieved fault propagation paths and root causes are sorted according to the propagation probability and the posterior probability to obtain a fault hypothesis list sorted by risk level, thereby determining the fault propagation path and fault cause.

6. The method for analyzing the failure modes of unmanned aerial vehicles according to claim 1, characterized in that, The step of generating an interactive failure analysis report using a domain-wide model based on the user-input description of the failure phenomenon, the failure propagation path, and the failure cause includes: The fault propagation path, the fault cause and its corresponding probability are fused with the description of the failure phenomenon to construct structured prompt words; The structured prompts are input into the domain-wide model to generate a comprehensive natural language report that includes at least fault phenomenon interpretation, causal chain analysis, risk level assessment, and maintenance recommendations. If a follow-up question is received from a user, the follow-up question is concatenated with relevant contextual information in the failure domain knowledge graph to form a new prompt word, which is then input into the domain big model to obtain a failure analysis report.

7. The method for analyzing the failure modes of unmanned aerial vehicles according to claim 1, characterized in that, The domain-specific large model is obtained through the following steps: Domain-adaptive training. Acquire professional textual data in the field of drone failure, including at least maintenance manuals, failure case libraries, and technical specifications; Using the aforementioned failure domain knowledge graph as a structured knowledge source, question-answer pairs containing graph entities, relationships, and reasoning paths are constructed as training samples. Using the aforementioned professional text corpus and training samples, the basic large language model is trained to obtain a domain-specific large model with the ability to understand and reason about drone failures.

8. A device for analyzing the failure modes of unmanned aerial vehicles (UAVs), characterized in that, include: The graph construction module is configured to construct a failure domain knowledge graph for UAVs by fusing structured flight log data with unstructured maintenance text data. An anomaly reasoning module is configured to perform probabilistic graphical reasoning on anomaly events detected in the real-time flight data stream based on the failure domain knowledge graph to determine the fault propagation path and fault cause. The fault analysis module is configured to generate an interactive fault analysis report using a large domain model based on the user-input description of the failure phenomenon, the fault propagation path, and the fault cause.

9. An electronic device, characterized in that, Including memory and processor, among which, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the method for analyzing the failure modes of the unmanned aerial vehicle according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer-readable programs or instructions, which, when executed by a processor, are capable of implementing the steps in the method for analyzing the failure modes of a UAV as described in any one of claims 1 to 7.

Images