A UAV fault diagnosis method, device, equipment and storage medium

Abstract

The present invention discloses a method, device, equipment and storage medium for diagnosing a drone fault, and relates to the technical field of fault diagnosis, and includes the following steps: constructing a fault knowledge graph; obtaining a drone fault problem and converting it into a to-be-queried graph; matching the to-be-queried graph with the fault knowledge graph to obtain a plurality of fault subgraphs; calculating the structural similarity and semantic similarity of the plurality of fault subgraphs with the to-be-queried graph, linearly combining the structural similarity and the semantic similarity to obtain a comprehensive score, and arranging the comprehensive scores from high to low to obtain an optimal fault subgraph; querying the optimal fault subgraph, and if the query result is a single fault cause, the drone fault diagnosis is completed; if the query result includes multiple fault causes, the optimal fault subgraph is mapped to a fuzzy Bayesian network, and the fuzzy Bayesian network is reasoned and analyzed through fuzzy reasoning to obtain the optimal fault cause. The present invention performs reasoning analysis on the optimal fault subgraph including multiple fault causes through a fuzzy Bayesian network, clarifies the possibility of occurrence of these fault causes, obtains the optimal fault cause, and enables auxiliary maintenance personnel to quickly and accurately locate the faulty unit.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a method, device, equipment and storage medium for diagnosing faults of unmanned aerial vehicles. Background Art

[0002] The purpose of UAV fault diagnosis is to find out the cause of the fault and locate the corresponding fault unit. The traditional UAV fault diagnosis process is that after the fault occurs, professional and technical backbones use their own business capabilities, experience and auxiliary data to find out all possible causes of the fault based on the observed fault phenomenon or signs, and then investigate and locate them one by one. This fault diagnosis method relies too much on expert experience and has low troubleshooting efficiency, which is difficult to meet the military aircraft combat support tasks under the new situation.

[0003] Knowledge graphs have powerful semantic expression, knowledge storage and reasoning capabilities. They are the cornerstone of the transition from perceptual intelligence to cognitive intelligence, and can make up for the lack of interpretability of traditional manual fault diagnosis results. At the same time, knowledge graphs can effectively process and utilize unstructured texts such as historical fault cases, operating procedures, and troubleshooting manuals, which effectively solves the problem of large amounts of unstructured data waste and "loose" knowledge in the field of fault diagnosis. In this regard, some scholars have realized the advantages of knowledge graph technology and introduced it into corresponding industry fields to carry out research on fault diagnosis technology. Deng Xianbo et al. analyzed and processed cable fault data in the past 10 years, constructed a fault diagnosis system based on cable fault knowledge graph, and initially realized the rapid analysis and diagnosis of cable line faults; Qiao Ji et al. used knowledge graph technology to condense the unstructured fault data in the form of a large amount of text in the power grid dispatching operation link into a structured knowledge network that can be represented, operated, and reasoned, in order to realize the analysis and judgment of fault information and intelligent auxiliary decision-making; Nie Tongpan et al. proposed a knowledge graph construction method for aircraft power system fault diagnosis, and based on the constructed knowledge graph, realized intelligent search, recommendation and intelligent question and answer of aircraft power system related faults; Chen Xi proposed an intelligent fault diagnosis method based on domain knowledge graph based on the characteristics of diesel engine fault data and diagnosis business needs, and realized fault cause reasoning to a certain extent.

[0004] In summary, the application of knowledge graph technology to the field of fault diagnosis has shortened the process time of fault diagnosis to a certain extent and preliminarily realized the diagnosis of the cause of the fault. However, for a certain fault phenomenon, the possible causes of the fault can be simply inferred through the knowledge graph, but the possibility of these causes of faults cannot be clearly determined, resulting in the inability of auxiliary maintenance personnel to quickly and accurately locate the faulty unit. Therefore, in the face of this situation, further reasoning and analysis based on the knowledge graph is needed to determine the possibility of the cause of the fault, more efficiently assist maintenance personnel in fault diagnosis, and improve maintenance efficiency. Summary of the invention

[0005] The present invention provides a UAV fault diagnosis method, device, equipment and storage medium, which solves the problem that the existing knowledge graph can only simply infer the possible causes of the fault, but cannot clarify the possibility of these fault causes, resulting in the inability of auxiliary maintenance personnel to quickly and accurately locate the faulty unit.

[0006] The present invention provides a method for diagnosing a fault of an unmanned aerial vehicle, comprising the following steps:

[0007] Build a fault knowledge graph based on the drone fault database;

[0008] Obtain drone failure problems and convert them into query graphs;

[0009] Match the query graph with the fault knowledge graph to obtain multiple fault subgraphs;

[0010] The structural similarity and semantic similarity of multiple fault subgraphs and the query graph are calculated, and the structural similarity and semantic similarity are linearly combined to obtain a comprehensive score. The comprehensive scores are arranged from high to low to obtain the optimal fault subgraph;

[0011] The optimal fault subgraph is queried. If the query result is a single fault cause, the UAV fault diagnosis is completed; if the query result includes multiple fault causes, the optimal fault subgraph is mapped to a fuzzy Bayesian network, and the fuzzy Bayesian network is reasoned and analyzed through fuzzy reasoning to obtain the optimal fault cause.

[0012] Preferably, the query graph is matched with the fault knowledge graph by a minimum spanning tree method to obtain multiple fault subgraphs.

[0013] Preferably, the structural similarity calculation is performed on multiple fault subgraphs and the query graph through a similarity matrix composed of nodes and edges, and the similarity matrix of the nodes is as follows:

[0014] SSim(X)={x 1 (k),x 2 (k)…}

[0015] The similarity matrix of the edges is as follows:

[0016] SSim(Y)={y 1 (k),y 2 (k)…}

[0017] In the formula, SSim(X) represents the similarity score matrix between the fault subgraph and the nodes in the query graph, x 1 (k),x 2(k)…represents the similarity value of each point in the two graphs after k iterations, SSim(Y) represents the similarity score matrix of the edges in the two graphs, y 1 (k),y 2 (k)…represents the similarity value of each edge in the two graphs after k iterations.

[0018] Preferably, semantic similarity calculation is performed on multiple fault subgraphs and the query graph by likelihood probability, and the likelihood probability calculation formula is as follows:

[0019]

[0020] In the formula, p(G 1 |G 2 ) represents the likelihood probability, G 1 represents the query graph, G 2 represents multiple fault subgraphs, e i represents the i-th word in the fault question, n represents the number of words in the fault question, and f j represents the jth triple in the fault knowledge graph, and m represents the number of triples in the fault knowledge graph.

[0021] Preferably, the structural similarity and the semantic similarity are linearly combined and calculated using the following formula:

[0022] Score=αScore SSim +(1-α)p(G 1 |G 2 )

[0023] In the formula, Score represents the comprehensive score, α represents a variable parameter with a value range of [0,1], and Score SSim Represents the weighted average of the score matrices SSim(X) and SSim(Y).

[0024] Preferably, mapping the optimal fault subgraph into a fuzzy Bayesian network comprises the following steps:

[0025] The entities in the optimal fault subgraph are converted into nodes of the Bayesian network, and the relationships between the entities are converted into directed edges of the Bayesian network. A topological structure is constructed based on the nodes and directed edges. If the entity is also associated with a fault cause, it is mapped as a root node, and if not, it is mapped as an intermediate node.

[0026] Determine the prior probability of the Bayesian network by the ratio of the fault in all fault cases;

[0027] Determine the membership function of fuzzy set theory to represent the fuzziness of the strength of association between nodes;

[0028] The fuzziness of the association strength between nodes is evaluated by the expert group decision method, and the conditional probability fuzzy number of the node under the occurrence of its parent node is obtained;

[0029] Solve the conditional probability fuzzy number to obtain the fuzzy probability of the association strength between the node and the parent node;

[0030] A conditional probability table is established based on the prior probability and the fuzzy probability of the association strength between the node and the parent node;

[0031] Combining the topological structure and the conditional probability table, a fuzzy Bayesian network is obtained. Preferably,

[0032] Preferably, the Bayesian network is analyzed by fuzzy reasoning to obtain the optimal fault cause, including the following steps:

[0033] The forward reasoning of Bayesian network is carried out through fuzzy causal reasoning. Based on the known prior probability of the cause layer nodes, the fuzzy Bayesian network reasoning is carried out to find the probability of the fault occurring in this case.

[0034] The Bayesian network is reversely inferred through fuzzy diagnostic reasoning. When a certain fault state is confirmed to occur, the posterior probability of the fault cause in the fuzzy Bayesian network is calculated according to the fuzzy Bayesian network reasoning algorithm.

[0035] A UAV fault diagnosis device, comprising:

[0036] A construction module is used to build a fault knowledge graph based on the drone fault database;

[0037] The acquisition module is used to obtain the drone fault problem and convert it into a query graph;

[0038] A matching module is used to match the query graph with the fault knowledge graph to obtain multiple fault subgraphs;

[0039] A calculation module is used to calculate the structural similarity and semantic similarity of multiple fault subgraphs and the query graph, perform linear joint calculation of the structural similarity and the semantic similarity to obtain a comprehensive score, and arrange the comprehensive scores from high to low to obtain the optimal fault subgraph;

[0040] The diagnosis module is used to query the optimal fault subgraph. If the query result is a single fault cause, the UAV fault diagnosis is completed; if the query result includes multiple fault causes, the optimal fault subgraph is mapped to a fuzzy Bayesian network, and the fuzzy Bayesian network is reasoned and analyzed through fuzzy reasoning to obtain the optimal fault cause.

[0041] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned UAV fault diagnosis method when executing the program.

[0042] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned drone fault diagnosis method is implemented.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The present invention first constructs a fault knowledge graph based on a fault database, then obtains actual fault problems and converts them into a query graph, matches the query graph with the fault knowledge graph, obtains multiple related fault subgraphs, and calculates scores for multiple fault subgraphs to obtain the optimal fault subgraph. This greatly reduces the time required for fault location. Then, for the optimal fault subgraph including multiple fault causes, a fuzzy Bayesian network is used to perform reasoning analysis on it, clarify the possibility of these fault causes, and obtain the optimal fault cause, so that auxiliary maintenance personnel can quickly and accurately locate the faulty unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0046] Figure 1 A flowchart of a method for diagnosing a fault of a drone according to the present invention;

[0047] Figure 2 It is a schematic diagram of the sub-graph matching process of the present invention;

[0048] Figure 3 Constructing a flow chart for the fuzzy Bayesian network of the present invention;

[0049] Figure 4 An example of a mapping construction for the present invention;

[0050] Figure 5 is a membership function of a fuzzy trigonometric function of the present invention;

[0051] Figure 6 A fault sub-map corresponding to a lubrication system fault according to an embodiment of the present invention;

[0052] Figure 7 A fuzzy Bayesian network topological structure diagram of a lubrication system fault according to the present invention;

[0053] Figure 8 It is a fuzzy Bayesian network diagram of lubrication system failure on GeNIe software of the present invention;

[0054] Fig. 9 It is a schematic diagram of the importance analysis results of the lubrication system failure causes of the present invention. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] A method for diagnosing UAV faults: when a fault occurs, maintenance personnel can input relevant fault problems into the knowledge graph based on the observed fault phenomenon, obtain the fault subgraph with the highest similarity to the fault phenomenon through subgraph matching, map it into a Bayesian network, and infer and analyze the most likely fault cause. The method includes the following steps:

[0057] Step 1: Build a fault knowledge graph based on the drone fault database.

[0058] Reference Figure 2 Based on a certain type of UAV fault corpus, through technical means such as knowledge extraction, knowledge fusion, and knowledge completion, we have realized the identification of fault entities, extraction of relations, and entity disambiguation from unstructured data texts, and constructed a fault knowledge graph G = (E, R, E). The UAV fault knowledge graph data stores a large number of entities E and relations R.

[0059] Step 2: Obtain the fault problem and convert it into a query graph.

[0060] Maintenance personnel can obtain fault problems based on observed fault phenomena, and transform the fault problems into query graphs Q = (E Q , R Q , E Q ).

[0061] Step 3: Match the query graph with the fault knowledge graph through the minimum spanning tree method to obtain multiple fault subgraphs.

[0062] Subgraph matching essentially means converting the fault problem or fault phenomenon input by the maintenance personnel into a fault graph (query graph Q), and then traversing the entire UAV fault knowledge graph (database graph G) to find a graph with the same structure and similar content as graph Q. A simple matching diagram is shown as follows: Figure 1 As shown. f(Q) is defined as a subgraph in G that satisfies the mapping function f, which maps the point X in the query graph Q to f(X) in G, and maps the edge r(X, Y) in the query graph Q to the edge f(r) = (f(X), f(Y)) in G.

[0063] The drone fault knowledge graph constructed by the present invention has a clear hierarchy and not particularly complex entity types, and can be matched by using the minimum spanning tree method. Its main process includes three parts: query graph decomposition, vertex selection, and pruning judgment.

[0064] The core of the decomposition of the query graph is to assign corresponding weights to each edge of the query graph Q. The larger the weight, the more times the edge appears in the database graph G. The core subgraph based on the minimum spanning tree uses the more characteristic edges in the query graph as the core edges for query matching. The weighting method has stronger screening capabilities and can better reflect the user's query intention. The purpose of using the minimum spanning tree method is to find the most accurate and ideal query graph at the cost of the minimum number of queries.

[0065] The purpose of vertex selection is to select a query vertex and effectively filter out data vertices that do not meet the criteria, so as to better map the query graph. The main factors that have a greater impact on the screening ability of vertex selection are the degree of the query vertex in the query graph Q (the number of edges associated with it) and the number of times the vertex with the same label as the query graph appears in the database graph G. The formula for calculating the vertex selection screening ability is as follows:

[0066]

[0067] In the formula, in Q (X) represents the in-degree of the query vertex, that is, the number of edges ending with the X entity relationship in the knowledge graph triple relationship structure; out Q (X) represents the out-degree of the query vertex, that is, the number of edges starting with X as the entity relationship; count G (m Q (X)) represents the number of times the vertex with the same label as vertex X appears in the database graph G; rank(X) represents the ranking of screening ability. The larger the value of rank(X), the stronger the screening ability.

[0068] Pruning decisions are made comprehensively from multiple dimensions, including vertex degrees, vertex labels, and edge labels. They mainly include pruning decisions based on graph structure features and pruning decisions based on label features. By comprehensively considering the labels and structures of the query graph, more fine-grained pruning can be achieved.

[0069] Step 4: Calculate the structural similarity and semantic similarity of multiple fault subgraphs and the query graph, perform linear joint calculation of the structural similarity and semantic similarity to obtain a comprehensive score, and arrange the comprehensive scores from high to low to obtain the optimal fault subgraph.

[0070] By calculating the structural similarity and content relevance scores of the fault subgraph to be queried and the graph to be queried, the best result is obtained by sorting the comprehensive scores. The knowledge graph has the characteristics of points and edges in structure, so when considering the similarity of the graph structure, the similarity of nodes and edges is also mainly considered. At the same time, the knowledge graph can convert sentences into vector representations through sentence encoding, and adding the semantic similarity of the sentence vector can calculate more accurate query results.

[0071] According to the graph structure similarity calculation theory, the following assumptions are made: if two graphs G 1 and G 2 , if G 1 Midpoint s and graph G 2 If the nodes around node t are similar, then node s can be considered similar to node t. 1 The edge l and graph G in 2 If the first and last nodes of edge h in are similar, then edge h and m can be considered similar. The present invention mainly evaluates the structural similarity between the query graph and the fault subgraph through the similarity matrix composed of nodes and edges. In short, the similarity value of any node or edge is always between 0 and 1. The closer to 1, the higher the structural similarity, and the higher the degree of subgraph matching.

[0072] If G 1 and G 2 There are n and m nodes in the matrix, respectively. The size of the node similarity matrix is ​​n×m. st ,y hl They represent the similarity between nodes s and t, and the similarity between edges h and l respectively. Then the similarity score formula for nodes and edges is:

[0073] SSim(X)={x 1 (k),x 2 (k)…} (2)

[0074] SSim(Y)={y 1 (k),y 2 (k)…}| (3)

[0075] Among them, SSim(X) represents the similarity score matrix of the nodes in the two graphs, x 1 (k),x 2 (k)…represents the similarity value of each point in the two graphs after k iterations; SSim(Y) represents the similarity score matrix of the edges in the two graphs, y 1 (k),y 2 (k)…represents the similarity value of each edge in the two graphs after k iterations.

[0076] By converting the natural language questions input by maintenance personnel into vector representation, the structure composition G of the query graph is obtained. 1 ={e 1 ,e 2 ,e 3 ,…,e n}, where e i is the word contained in the maintenance personnel’s question. At the same time, let the fault subgraph G 2 ={f 1 ,f 2 ,f 3 ,…,f n}, where f i It is composed of triples in the fault knowledge graph database. The similarity calculation based on the statistical language model is mainly based on the likelihood probability p(G 1 |G 2 ) indicates that the larger the probability value, the higher the similarity. The likelihood probability calculation formula is:

[0077]

[0078] In the formula, p(e i |G 2 ) is Figure G 1 The statistical language model generates word e i The probability of e i The probability p(e) generated by multiple triple models i |f j ) average value. At the same time, the present invention comprehensively considers the graph structure similarity and semantic similarity, and uses the graph structure similarity Score SSim and semantic similarity p(G 1 |G 2 ) for linear joint calculation, the calculation formula is:

[0079] Score=αScore SSim +(1-α)p(G 1 |G 2 ) (5)

[0080] In the formula, α is a variable parameter with a value range of [0,1], which is used to adjust the proportion of graph structure similarity and semantic similarity in the result score. By comprehensively considering the two similarities, the comprehensive score Score is calculated to finally obtain the comprehensive score of each fault subgraph, and the optimal query result is obtained by ranking the scores.

[0081] Through the above method, the matching of a single fault description or representation query graph to the fault subgraph in the fault knowledge graph is realized, and the matching results are fed back to the user in the form of a visual graph, which narrows the scope of fault troubleshooting and assists maintenance personnel to achieve UAV system fault isolation based on subgraph matching.

[0082] Through the sub-graph matching method, one or more fault causes related to the fault phenomenon can be quickly found and the scope of investigation can be locked.

[0083] Step 5: Query the optimal fault subgraph. If the query result is a single fault cause, the fault diagnosis is completed. If the query result includes multiple fault causes, the optimal fault subgraph is mapped to a Bayesian network, and the Bayesian network is analyzed through fuzzy reasoning to obtain the optimal fault cause.

[0084] For maintenance personnel, when the query result is a single fault cause, fault handling can be carried out immediately. However, if the query result is multiple fault causes, although the scope of fault investigation is narrowed, each fault cause has the same probability, and targeted maintenance measures cannot be formulated. Therefore, the present invention introduces the fuzzy Bayesian reasoning algorithm and utilizes its advantages in causal reasoning logic to achieve optimization of UAV fault diagnosis strategy.

[0085] A Bayesian network is a directed acyclic network consisting of a directed acyclic graph (DAG) and several conditional probability tables (CPT). It can be represented by the symbol B(G,P), where G represents the structure of random variable nodes, which is the topological structure of the Bayesian network, and P represents the conditional probability of the directed edges.

[0086] Establishing a Bayesian network model includes topological structure construction and quantitative parameter setting. Common Bayesian network construction mainly adopts methods based on expert knowledge and data learning modeling. The construction method based on expert knowledge (such as the fault tree mapping construction method) has low construction efficiency. The method based on data learning modeling can overcome the limitations of relying on expert experience and has high construction efficiency, but the accuracy and interpretability of the constructed Bayesian network are insufficient. To this end, the present invention proposes a "knowledge graph-Bayesian network" mapping construction method, and its specific process is as follows Figure 3 As shown, it is divided into the following three steps:

[0087] (1) Node mapping. Using the better visualization graph structure of the knowledge graph, the entities in the fault subgraph are converted into nodes of the Bayesian network, where fault cause I, fault cause II, and fault mode entities are mapped to the root node, intermediate node, and leaf node, respectively. The difference between fault cause I and fault cause II is that there is no fault cause association behind them. If there is no fault cause association, it is mapped to the root node, otherwise it is mapped to the intermediate node.

[0088] (2) Relationship mapping: Convert the “failure cause” and “cause” associations in the fault subgraph into directed edges in the network. Figure 4 is a simple example of building a map.

[0089] (3) Parameter setting. This mainly includes the prior probability and conditional probability of the Bayesian network. The prior probability of fault can reflect the historical fault conditions of the aircraft engine and is an important basis for diagnosing the current fault type. The prior probability of fault can be determined by the ratio of the fault in all fault cases. Figure 4 Medium V 5 Of the 200 failure cases that occurred, 12 were caused by V 1 The prior probability P(V 1 )=12 / 200=0.06.

[0090] Conditional probability cannot be expressed with precise mathematical formulas or numerical values ​​due to the existence of uncertain factors such as environment, human factors, and quality in actual maintenance, and simulation experiments cannot fully consider all influencing factors in actual situations. Therefore, the present invention introduces fuzzy set theory in the process of determining conditional probability to deal with the uncertainty problem in engineering practice. The specific method is as follows:

[0091] (1) Determine the fuzzy function and comment level

[0092] There are many forms of membership functions in fuzzy set theory, such as triangular fuzzy numbers, trapezoidal fuzzy numbers and normal fuzzy numbers. Among them, triangular fuzzy numbers are easy to calculate and practical, and are widely used in engineering theory research. Therefore, triangular fuzzy numbers are used to represent the fuzziness of the strength of association between nodes. The membership function expression is:

[0093]

[0094] In the process of setting parameters of the uncertainty Bayesian network, the correlation strength between some nodes usually cannot obtain accurate probability values. In this case, it can be effectively evaluated by the expert decision-making method. To this end, the present invention introduces seven linguistic variables of "extremely high", "high", "relatively high", "average", "low", "low", and "extremely low" in combination with the military aviation engine maintenance example. The corresponding triangular fuzzy numbers are shown in Table 1.

[0095] Table 1 Semantic values ​​of node association strength and corresponding fuzzy numbers

[0096]

[0097] (2) Obtaining fuzzy probability from linguistic variables

[0098] The expert group decision-making method can better describe the strength of the association between nodes. If the number of experts is s, the kth expert pair at a given node V i Under the condition of the parent node, V i The evaluation language value of the occurrence probability is converted into a triangular fuzzy number according to Table 1:

[0099]

[0100] At the same time, the evaluation ability level of experts is taken into consideration, and the evaluation weight of each expert is determined by using the literature method according to the expert's position, title, academic qualifications, etc. The calculation formula is:

[0101]

[0102] Where S k is the weighted total score of the kth expert; S ek , S sk , S pk , S rk Respectively represent the expert's scores in terms of education, years of work experience, professional title, and position. k represents the weight of the evaluation result of the kth expert. Based on this, the evaluation results of multiple experts are combined to obtain V i The conditional probability fuzzy number of a node under the occurrence of its parent node:

[0103]

[0104] (3) Defuzzification of fuzzy probability

[0105] The conditional probability after fuzzification is a fuzzy possibility interval, and it is often necessary to adopt a defuzzification algorithm to calculate a more accurate value to represent the eigenvalue of the fuzzy set. Compared with the centroid method, the bisector method and the maximum mean method, the α-weighted valuation method can ensure the integrity of information in the defuzzification process. Therefore, the present invention adopts the α-weighted valuation method to complete the defuzzification process. Figure 5 It is a membership function of a fuzzy trigonometric function.

[0106] Formula (10) is a generalized formula for defuzzification of fuzzy probability, where F α = {x|F(x)>α} is the α-cut of F(x), Average(F α) is the average value of the α-cut set, f(α) is the weighted valuation function, and Average(F) can be calculated by formula (11): α ).

[0107]

[0108]

[0109] l α =(m′ i -a′ i )×α+a′ i (12)

[0110] h α = b′ i -(b′ i -m′ i )×α (13)

[0111] Among them l α ,h α They represent the lower and upper limits of the α-cut set, respectively. Usually, f(α)=1 is assumed, so that the probability value after defuzzification can be obtained by equation (14).

[0112]

[0113] (4) Establish FBN conditional probability table

[0114] The above method can be used to obtain the fuzzy probability of the association strength between a child node and a parent node, but how to obtain the CPT in FBN is widely considered to be a more complex problem, especially when the FBN is large and there are many parent nodes, the number of parameters that require expert evaluation will increase exponentially. Therefore, the present invention adopts the Noisy-OR model to achieve the establishment of CPT.

[0115] Assume an x 1 , x 2 , …, x n are n fault causes that cause fault y, and their causal relationships with fault y are independent of each other, then cause x i The probability of y fault occurring under the condition:

[0116]

[0117] In the formula, Indicates that the node event did not occur, then there are multiple fault causes x 1 , x 2 , …, x n The coupling effect on fault y is calculated as:

[0118]

[0119] After adopting the Noisy-OR model, only the association strength between a single faulty node and its parent node needs to be fuzzily set. The probability P(y|x) of a child node failing under different cause state combinations is obtained by coupling the effect of a single cause on the failure, without the need to set each association strength separately. For example, for a node with n parent nodes, only 2n association strengths need to be evaluated, instead of the original 2. n The correlation strength is calculated, which greatly simplifies the complexity of CPT establishment in the Bayesian network parameter setting.

[0120] The fuzzy Bayesian network constructed in the above way can be used for various types of evaluation and analysis. The most important use of FBN is to correct the probability according to the actual observation of the event, so as to calculate the probability distribution of the cause of the fault, find the most likely potential cause when the fault occurs, and then optimize the troubleshooting diagnosis strategy.

[0121] Fuzzy Bayesian reasoning is mainly divided into fuzzy causal reasoning and fuzzy diagnostic reasoning. Fuzzy causal reasoning is forward reasoning, that is, based on the known prior probability of the cause layer nodes, fuzzy Bayesian network reasoning is performed to find the probability of failure in this case. Fuzzy causal reasoning can be used to calculate the operating failure probability of a certain system of an aircraft engine based on historical failure data, thereby obtaining the operating reliability of the system. The fuzzy causal reasoning formula is:

[0122]

[0123] Where π(V i ) indicates V i The parent node set of Figure 4 Medium V 5 The inference calculation process of probability is:

[0124] P(V 5 )=P(V 3 )P(V 5 |V 3 )+P(V 5 |V 4 )×(P(V 1 )P(V 4 |V 1 )+P(V 2 )P(V 4 |V 2 )) (18)

[0125] Fuzzy diagnostic reasoning is reverse reasoning, that is, when a certain fault state is determined to occur, the posterior probability of the fault cause in the fuzzy Bayesian network is calculated according to the fuzzy Bayesian network reasoning algorithm. The calculation results can be used to optimize the fault diagnosis strategy and formulate targeted measures to improve task reliability. The fuzzy diagnostic reasoning formula is:

[0126]

[0127] by Figure 4 The node in the middle is explained, assuming that the fault layer node V 5 Occurs, find the cause layer node V 1 The posterior probability of occurrence is:

[0128]

[0129] When searching for the most critical cause of a fault, relying solely on a priori or a posteriori probability may lead to inaccurate results, thereby increasing unnecessary maintenance costs. The present invention uses the ratio of variation (RoV) as an indicator of importance to identify the most critical cause of a fault, which is an important part of quantitative analysis of system reliability and fault diagnosis. i The RoV importance calculation formula is:

[0130]

[0131] In the formula ξ(x i ) and ζ(x i ) represent the posterior probability and prior probability of the i-th fault root cause respectively.

[0132] Example

[0133] Through sub-graph matching, the fault sub-graph corresponding to the lubrication system fault can be obtained, such as Figure 6 Then, according to the “fault knowledge graph-Bayesian network” mapping construction method, the Bayesian network topology of the lubrication system fault is established, as shown in Figure 7 The fuzzy parameter setting results and fault diagnosis of lubrication system faults are as follows:

[0134] The historical data of lubrication system failures of this type of aircraft engine in the past five years is used as the data sample for verification of this method. The prior probability statistics are shown in Table 2.

[0135] Five domain experts were invited to evaluate the association relationship of nodes in the fuzzy Bayesian network of lubrication system faults. The relevant information, score and evaluation weight of each expert are shown in Table 3. Then, the association strength of the individual influence of child nodes and parent nodes in the Bayesian network of lubrication system faults was obtained through expert evaluation, as shown in Table 4. Finally, the conditional probability table in FBN was established. In order to clearly describe its calculation process, the node y 3 "High metal content in lubricating oil" and its three parent nodes x 5 , x 6 , x 7 Take 2 as an example for explanation, see Table 5 for details. At the same time, in actual engineering, even if all components are not faulty, the system may still be abnormal. Therefore, this embodiment introduces the Leak node when calculating the conditional probability, and assumes that its effect on the occurrence of the fault always exists, that is, when the fault causes listed in the model have not occurred, the fault may still occur. The influence intensity of the Leak node on the fault is uniformly set to 0.05, such as P Ly3 =0.05.

[0136] Table 2 Fault prior probability

[0137]

[0138]

[0139] Table 3 Expert ability evaluation results

[0140]

[0141] Table 4 Summary of expert evaluation

[0142]

[0143] Table 5 Conditional probability calculation process of node y3

[0144]

[0145]

[0146] After all CPTs are obtained, FBN reasoning can be applied for quantitative analysis. This embodiment uses GeNIe software to perform reasoning analysis on the fuzzy Bayesian network model of lubrication system failure. The model construction and parameter settings are as follows: Figure 8 , “State 1” and “State 0” respectively indicate that the node event “occurs” and “does not occur”.

[0147] Based on the above Bayesian network diagram of lubrication system failure, it can be seen that based on historical failure statistics, the probability of lubrication system failure T occurring is 19% through Bayesian forward reasoning, that is, the system operation reliability is 81%. Then, the Bayesian network can be used to diagnose and reason about lubrication system failures and find the key causes. That is, assuming that a lubrication system failure has occurred, the State1 state value of leaf node T is set to 100%, and the posterior probability of each root cause of the failure is obtained according to Bayesian reverse reasoning, as shown in Table 6. Then, the importance analysis of each root cause node is carried out, as shown in Table 6. Fig. 9 shown.

[0148] From the above results, we can see that x 8 , x 5 , x 10 , x 4 The root causes of failures such as these have a greater impact on the failure of the lubrication system and are the weak link of the system. To this end, on the one hand, targeted measures such as shortening the maintenance interval, reliability optimization, and redundant design can be taken to improve system reliability; on the other hand, during fault diagnosis, such root causes of failures should be prioritized.

[0149] Table 6 Fault posterior probability

[0150]

[0151] Based on the same concept, the present invention also provides a UAV fault diagnosis device, which is characterized by comprising a construction module, an acquisition module, a matching module, a calculation module and a diagnosis module.

[0152] The building module is used to build a fault knowledge graph based on the drone fault database.

[0153] The acquisition module is used to obtain drone fault problems and convert them into query graphs.

[0154] The matching module is used to match the query graph with the fault knowledge graph to obtain multiple fault subgraphs.

[0155] The calculation module is used to calculate the structural similarity and semantic similarity of multiple fault subgraphs and the query graph, linearly combine the structural similarity and semantic similarity to obtain a comprehensive score, and arrange the comprehensive scores from high to low to obtain the optimal fault subgraph.

[0156] The diagnosis module is used to query the optimal fault subgraph. If the query result is a single fault cause, the UAV fault diagnosis is completed; if the query result includes multiple fault causes, the optimal fault subgraph is mapped to a fuzzy Bayesian network, and the fuzzy Bayesian network is reasoned and analyzed through fuzzy reasoning to obtain the optimal fault cause.

[0157] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the above-mentioned UAV fault diagnosis method is implemented when the processor executes the program.

[0158] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned drone fault diagnosis method is implemented.

[0159] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0160] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for diagnosing a drone fault, characterized in that: The following steps are involved: Build a fault knowledge graph based on the drone fault database; Obtain drone failure problems and convert them into query graphs; Match the query graph with the fault knowledge graph to obtain multiple fault subgraphs; The structural similarity and semantic similarity of multiple fault subgraphs and the query graph are calculated, and the structural similarity and semantic similarity are linearly combined to obtain a comprehensive score. The comprehensive scores are arranged from high to low to obtain the optimal fault subgraph; The optimal fault subgraph is queried. If the query result is a single fault cause, the UAV fault diagnosis is completed; if the query result includes multiple fault causes, the optimal fault subgraph is mapped to a fuzzy Bayesian network, and the fuzzy Bayesian network is reasoned and analyzed through fuzzy reasoning to obtain the optimal fault cause; Mapping the optimal fault subgraph into a fuzzy Bayesian network includes the following steps: The entities in the optimal fault subgraph are converted into nodes of the Bayesian network, and the relationships between the entities are converted into directed edges of the Bayesian network. A topological structure is constructed based on the nodes and directed edges. If the entity is also associated with a fault cause, it is mapped as a root node, and if not, it is mapped as an intermediate node. Determine the prior probability of the Bayesian network by the ratio of the fault in all fault cases; Determine the membership function of fuzzy set theory to represent the fuzziness of the strength of association between nodes; The fuzziness of the association strength between nodes is evaluated by the expert group decision method, and the conditional probability fuzzy number of the node under the occurrence of its parent node is obtained; Solve the conditional probability fuzzy number to obtain the fuzzy probability of the association strength between the node and the parent node; A conditional probability table is established based on the prior probability and the fuzzy probability of the association strength between the node and the parent node; Combining the topological structure and the conditional probability table, a fuzzy Bayesian network is obtained; The optimal fault cause is obtained by reasoning and analyzing the Bayesian network through fuzzy reasoning, including the following steps: The forward reasoning of Bayesian network is carried out through fuzzy causal reasoning. Based on the known prior probability of the cause layer nodes, the fuzzy Bayesian network reasoning is carried out to find the probability of the fault occurring in this case. The Bayesian network is reversely inferred through fuzzy diagnostic reasoning. When a certain fault state is confirmed to occur, the posterior probability of the fault cause in the fuzzy Bayesian network is calculated according to the fuzzy Bayesian network reasoning algorithm.

2. A method for diagnosing a drone fault as claimed in claim 1, characterized in that: The query graph is matched with the fault knowledge graph through the minimum spanning tree method to obtain multiple fault subgraphs.

3. A method for diagnosing a drone fault as claimed in claim 1, characterized in that: The structural similarity of multiple fault subgraphs and the query graph is calculated through a similarity matrix composed of nodes and edges. The similarity matrix of the nodes is as follows: SSim(X)={x1(k),x2(k)…} The similarity matrix of the edges is as follows: SSim(Y)={y1(k),y2(k)…} Where SSim(X) represents the similarity score matrix between the fault subgraph and the nodes in the query graph, x1(k), x2(k)… represent the similarity values ​​of the points in the two graphs after k iterations, SSim(Y) represents the similarity score matrix of the edges in the two graphs, and y1(k), y2(k)… represent the similarity values ​​of the edges in the two graphs after k iterations.

4. A method for diagnosing a drone fault as claimed in claim 3, characterized in that: The semantic similarity between multiple fault subgraphs and the query graph is calculated by likelihood probability. The likelihood probability calculation formula is as follows: In the formula, p(G1|G2) represents the likelihood probability, G1 represents the query graph, G2 represents multiple fault subgraphs, and e i represents the i-th word in the fault question, n represents the number of words in the fault question, and f j represents the jth triple in the fault knowledge graph, and m represents the number of triples in the fault knowledge graph.

5. A method for diagnosing a drone fault as claimed in claim 4, characterized in that: The structural similarity and semantic similarity are linearly combined and calculated using the following formula: Score=αScore SSim +(1-α)p(G1|G2) In the formula, Score represents the comprehensive score, α represents a variable parameter with a value range of [0,1], and Score SSim Represents the weighted average of the score matrices SSim(X) and SSim(Y).

6. A UAV fault diagnosis device, characterized in that: include: A construction module is used to build a fault knowledge graph based on the drone fault database; The acquisition module is used to obtain the drone fault problem and convert it into a query graph; A matching module is used to match the query graph with the fault knowledge graph to obtain multiple fault subgraphs; A calculation module is used to calculate the structural similarity and semantic similarity of multiple fault subgraphs and the query graph, perform linear joint calculation of the structural similarity and the semantic similarity to obtain a comprehensive score, and arrange the comprehensive scores from high to low to obtain the optimal fault subgraph; The diagnosis module is used to query the optimal fault subgraph. If the query result is a single fault cause, the UAV fault diagnosis is completed; if the query result includes multiple fault causes, the optimal fault subgraph is mapped to a fuzzy Bayesian network, and the fuzzy Bayesian network is reasoned and analyzed through fuzzy reasoning to obtain the optimal fault cause; Mapping the optimal fault subgraph into a fuzzy Bayesian network includes the following steps: The entities in the optimal fault subgraph are converted into nodes of the Bayesian network, and the relationships between the entities are converted into directed edges of the Bayesian network. A topological structure is constructed based on the nodes and directed edges. If the entity is also associated with a fault cause, it is mapped as a root node, and if not, it is mapped as an intermediate node. Determine the prior probability of the Bayesian network by the ratio of the fault in all fault cases; Determine the membership function of fuzzy set theory to represent the fuzziness of the strength of association between nodes; The fuzziness of the association strength between nodes is evaluated by the expert group decision method, and the conditional probability fuzzy number of the node under the occurrence of its parent node is obtained; Solve the conditional probability fuzzy number to obtain the fuzzy probability of the association strength between the node and the parent node; A conditional probability table is established based on the prior probability and the fuzzy probability of the association strength between the node and the parent node; Combining the topological structure and the conditional probability table, a fuzzy Bayesian network is obtained; The optimal fault cause is obtained by reasoning and analyzing the Bayesian network through fuzzy reasoning, including the following steps: The forward reasoning of Bayesian network is carried out through fuzzy causal reasoning. Based on the known prior probability of the cause layer nodes, the fuzzy Bayesian network reasoning is carried out to find the probability of the fault occurring in this case. The Bayesian network is reversely inferred through fuzzy diagnostic reasoning. When a certain fault state is confirmed to occur, the posterior probability of the fault cause in the fuzzy Bayesian network is calculated according to the fuzzy Bayesian network reasoning algorithm.

7. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for diagnosing a fault of a UAV as described in any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the drone fault diagnosis method described in any one of claims 1 to 5 is implemented.

Images