Heuristic-based method for generating failure modes in the aviation domain

Abstract

This invention relates to a heuristic-based method for generating fault patterns in the aviation field, comprising: S1, extracting features from aviation fault text data using a bag-of-words model and the term frequency-inverse text frequency index method; S2, performing clustering category analysis of aviation fault text data based on k-means heuristics; S3, detecting outliers, extracting faults, and concatenating faults in the aviation fault text data to obtain fault pattern names; and S4, processing the obtained fault description text in real time to generate aviation fault patterns. This invention completes text data clustering category analysis using a bag-of-words model, the term frequency-inverse text frequency index method, and the k-means heuristic method. Furthermore, it obtains fault pattern names through outlier detection, fault extraction, and concatenation, enabling real-time detection of fault text data. By periodically re-clustering, it can discover new fault patterns by utilizing existing prior knowledge, thereby improving the accuracy and effectiveness of fault pattern detection.

Description

Technical Field

[0001] This invention belongs to the field of aviation fault analysis, and specifically relates to a heuristic-based method for generating aviation fault modes. Background Technology

[0002] Failure modes generally describe the state of a failure and its impact on equipment operation. They are a standardized and normalized form of description of failure phenomena and are widely used in FMEA, FTA, and testing in the aviation field.

[0003] During the development, testing, and use of aviation products, engineers and users record a large amount of fault text. However, due to differences in natural language expression among different recorders, the descriptions of fault phenomena vary. For example, "left brake pressure is lower than the specified value" and "left brake pressure is 9 MPa, lower than the specified 20 MPa" both express the fault mode of "low left wheel brake pressure." This massive amount of differentiated fault records poses significant challenges to fault pattern discovery, overall fault situation analysis, and fault resolution formulation. There is an urgent need to uniformly summarize and generalize this massive amount of fault text records to form standardized fault patterns, supporting full lifecycle fault analysis in the aviation industry and improving product quality.

[0004] This method takes aviation fault text as the research object and establishes a fault text data processing flow of existing fault clustering, fault pattern generation, and new fault pattern classification. This invention proposes a heuristic fault category determination method, an outlier detection method in fault text data, and a nested named entity recognition method based on a dual affine model for fault text data, so as to realize the clustering of fault text data, fault pattern generation, and classification of new data. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a heuristic-based method for generating fault modes in the aviation field. This invention utilizes a bag-of-words model, the word frequency-inverse text frequency index method, and the k-means heuristic to perform text data clustering and category analysis. Furthermore, it obtains fault mode names and enables real-time detection of fault text data through outlier detection, fault extraction, and concatenation. By periodically re-clustering, new fault modes can be discovered by leveraging existing prior knowledge, thereby improving the accuracy and effectiveness of fault mode generation.

[0006] To achieve the above objectives, the present invention discloses the following technical solution: a heuristic-based method for generating failure modes in the aviation field, comprising:

[0007] S1: Use the bag-of-words model and the term frequency-inverse text frequency index method to extract features from fault text data in the aviation field;

[0008] Obtain the set of fault description texts U = {u 1 ,u 2 ,…,u m}, then for all fault description text u i ∈U, perform word segmentation to obtain its word segmentation sequence. By combining the word segmentation results of all fault description texts, an aviation fault text data dictionary D is obtained, and an aviation text data set matrix V is constructed. Then, the constructed aviation text data set matrix V is transformed using the term frequency-inverse text frequency index method, that is, for each element... The transformation is performed as follows:

[0009]

[0010] Among them, T(V) ij ) represents the term frequency-inverse text transformation matrix element of the fault description text data; V ij The elements of the fault description text data set matrix are: m, g, f(w); m is the total number of fault description text data; g is the total number of elements in the aviation field fault text data dictionary; f(w) i ) for statistics w i Number of occurrences; w i f is the i-th word in the dictionary; ui (w i ) for w i In sequence u i The number of times it appears in the text; i is the fault description text number; j is the word segmentation sequence number;

[0011] Then, the term frequency-inverse text transformation matrix elements T(V) of all aviation text data are processed. ij The matrix is ​​regularized to obtain each row vector of the matrix, which serves as a feature vector for the fault description text data. Finally, the data feature sample matrix X = {x1, x2, ..., x...} is formed to represent the fault text data in the aviation field. m};

[0012] S2: Clustering category analysis of aviation fault text data based on k-means heuristic;

[0013] Obtain the data feature sample matrix X = {x1, x2, ..., x} of the aviation field fault text in step S1. m}, cluster them into K disjoint feature clusters, for each feature cluster c j ∈{c1,c2,…,c s The optimal feature cluster center vector μ is obtained by calculating the mean of the samples in the feature cluster and minimizing the objective value inertia. j , as shown below:

[0014]

[0015] Where inertia is the objective value to be minimized; μ j For the j-th feature cluster c j Center vector; x i represents the elements of the data feature sample matrix; s represents the total number of feature clusters;

[0016] S3: Outlier detection, fault extraction, and fault concatenation of fault text data in the aviation field to obtain fault mode names;

[0017] S31: Outlier detection in aviation fault text data using box plots; element c in feature cluster set C. l to the feature cluster center vector μ j The distance is d l =||c l -μ j ||2, based on the distance {d1,d2,…,d s} Calculate the first quartile Q1 and the third quartile Q3, and obtain the interquartile range as follows:

[0018] IQR = Q3 - Q1;

[0019] Where IQR is the interquartile range; Q1 is the first quartile; and Q3 is the third quartile.

[0020] For distance d l If it is greater than the threshold Q3 + 1.5IQR, then d is considered to be... l The corresponding c l These are outliers and should be deleted.

[0021] S32: A dual affine model is used to realize nested named entity recognition, extracting the fault subject and phenomenon; the input fault description text is processed by a transformer-based bidirectional encoder representation encoder and BiLSTM layer for feature extraction to obtain its vector representation, obtain each effective named cover J of the fault description text, calculate its predicted probability for each category, and use the cross-entropy loss function of the normalized exponential function for classification, as shown below:

[0022]

[0023] Where loss is the cross-entropy loss function; l is the entity category index; L is the total number of entity categories; and J is the effective naming coverage. For the set of all valid named covers; y Jl To efficiently name the symbolic function covering entity category l for J; p m (J l) represents the predicted probability of J for category l;

[0024] S33: Concatenate all fault topics and fault phenomena in the fault description text data of the aviation field; after obtaining all fault subjects and fault phenomenon phrases in each fault description text in each feature cluster, concatenate them as the fault mode name of this feature cluster.

[0025] S4: Real-time processing of obtained fault description text to generate aviation-related fault modes;

[0026] To obtain newly added fault description text in real time, firstly, step S1 is performed to obtain the feature matrix of fault text data in the aviation field; then, according to step S2, clustering category analysis of fault text data in the aviation field is performed; finally, the fault mode name is obtained through step S3, and the fault mode in the aviation field is generated.

[0027] Preferably, in step S1, the word segmentation results of all fault description texts are combined to obtain an aviation fault text data dictionary D, and an aviation text data set matrix V is constructed, specifically as follows:

[0028] First, obtain all word segmentation sequences u i The set of words appearing in the text yields the following aviation-related fault text data dictionary D:

[0029] D = {w1, w2, ..., w g};

[0030] Where D is the aviation field fault text data dictionary; w1 is the first word in the dictionary; w2 is the second word in the dictionary; w g The g-th word in the dictionary;

[0031] Then, determine the word segmentation sequence u j The word frequency vectors are:

[0032]

[0033] in, For word segmentation sequence u i The word frequency vector;

[0034] Finally, the aeronautical text data set matrix V is constructed as follows:

[0035]

[0036] Among them, V m×g It is a matrix of aviation text data set with m rows and g columns.

[0037] Preferably, in step S1, the word frequency-inverse text transformation matrix element T(V) of all aviation text data is... ijPerform regularization transformation, specifically:

[0038] According to T(V) ij The term frequency-inverse text transformation matrix of the constructed aviation text data is T(V), as shown below:

[0039]

[0040] Where T(V) is the term frequency-inverse text transformation matrix of the aviation text data;

[0041] Then for all elements T(V) ij Perform regularization as follows:

[0042]

[0043] Among them, N(T(V) ij The term frequency-inverse text transformation matrix element regularization result is represented by ||T(V) for the aviation text data. i ||2 is the 2-norm of the i-th row vector;

[0044] The 2-norm of the i-th row vector ||T(V) i ||2 is:

[0045]

[0046] Preferably, step S2, which clusters into K disjoint feature clusters, uses a data-based K-value estimation method, specifically:

[0047] The value of K ∈ {1,2,3,…,K} max The K value is calculated based on the clustering result inertia obtained in step S2, which is {μ1, μ2, ..., μ}. Kmax This sequence monotonically decreases as the value of K increases; when the value of K is less than or equal to the true value of the number of categories K... * At that time, sequence μ K There will be a most significant reduction, as shown below:

[0048] {μ K -μ K+1 |K <K *}>>{μ K -μ K+1 |K≥K *};

[0049] Where, μ K μ is the inertia value of the clustering results when the number of categories is K; K+1 The inertia value is the clustering result when the number of categories is K+1; K is the number of categories; K *The true value of the number of categories;

[0050] Therefore, μ K The K corresponding to the inflection point of the monotonically decreasing process is used as the reference to the true value K. * The estimate;

[0051] Specifically, first calculate the sequence The first difference is divided into:

[0052] Δ 1 μ K =μ K+1 -μ K (K = 1, 2, ..., K) max -1);

[0053] Where, Δ 1 μ K The first difference of the inertia value of the clustering result;

[0054] Based on the first-order difference, the second-order difference is then calculated as follows:

[0055] Δ 2 μ K =Δ 1 μ K+1 -Δ 1 μ K (K = 1, 2, ..., K) max -2);

[0056] Where, Δ 2 μ K The second difference of the inertia value of the clustering result;

[0057] Obtain the second-order difference Δ 2 μ K The value is used as a reference to the true value K * Best estimate As shown below:

[0058]

[0059] in, For the true value K * The best estimate; argmax is the function that takes the maximum value of the function's arguments.

[0060] Preferably, in step S31, the distance {d1,d2,…,d} is used to determine the distance. s} Calculate the first quartile Q1 and the third quartile Q3, specifically as follows:

[0061] All distances {d1, d2, ..., d sThe values ​​of} are arranged from smallest to largest and divided into four equal parts by three dividing points. The value at the first dividing point is the first quartile Q1, and the value at the third dividing point is the third quartile Q3.

[0062] Preferably, the fault description text input in step S32 is processed by a converter-based bidirectional encoder representation encoder and a BiLSTM layer, specifically as follows:

[0063] First, the beginning h of each valid named cover J is calculated using two independent feedforward neural networks. s and the ending h e Specifically:

[0064]

[0065] Among them, h s (J) is a valid name that covers the beginning of J; h e (J) is a valid name overwrite of the end of J; FFNN s For the first feedforward neural network; FFNN e This is the second feedforward neural network; This serves as the input to the first feedforward neural network; This is the input to the second feedforward neural network; s J and the starting index for effective naming coverage J; e J It is the end indicator of effective named coverage J;

[0066] Then, the score of the effective named cover J is calculated using a biaffine model, resulting in the following score tensor:

[0067]

[0068] Where, r m (J) is a valid name covering J's c r Fractional tensor of dimension U; m For the biaffine model d r ×c r ×d r tensor; W m For a 2d double affine model r ×c r The matrix; b m This represents the bias of the double affine model; d is the direct sum of the matrices; r c represents the dimension of the vector output by the LSTM layer. r The number of entity categories;

[0069] Finally, the fractional tensor r mThis includes the score for each valid named cover J belonging to each entity category, and the entity category for each valid named cover J is labeled as:

[0070] y′(J)=argmax r m (J);

[0071] Where y'(J) is the valid named coverage of the entity category marked by J.

[0072] Preferably, the effective naming coverage J in step S23 l loss parameter p m (J l Specifically:

[0073]

[0074] Where, p m (J l ) represents the predicted probability of J for category l; This is a summation index for entity category serial numbers.

[0075] Preferably, the stitching of fault description text data in step S33, which involves all fault topics and fault phenomena in the aviation field, specifically includes:

[0076] Given the set of all fault subject phrases A = {a} and the set of fault phenomenon phrases P = {p} from the extracted feature clusters, consider all concatenation combinations N = {(a,p)|a∈A,p∈P}. For any (a,p)∈N, the concatenation score is set as follows:

[0077] score((a,p))=c A (a)+c P (p)+c AP (a,p);

[0078] Where score((a,p)) is the concatenated score result; a is the fault topic phrase; p is the fault phenomenon phrase; c A (a) represents the number of times the faulty phrase 'a' appears in this cluster; c P (p) represents the number of times the fault phenomenon phrase p appears in this cluster; c AP (a,p) represents the number of times that the fault subject phrase a and the fault phenomenon phrase p appear simultaneously in this cluster, and obtains all combinations in the feature cluster;

[0079] The one with the highest score is selected as the fault mode name for that feature cluster.

[0080] Preferably, after generating aviation-related fault patterns in step S4, new fault data types will emerge in aviation-related fault text data as the acquired data accumulates; therefore, the existing clustering results need to be updated periodically to re-cluster the new data with the original data; the re-clustering only needs to reflect the new fault patterns on the newly added data.

[0081] Compared with existing technologies, the present invention has the following advantages: The present invention completes the clustering category analysis of fault text data through the bag-of-words model, the term frequency-inverse text frequency index method, and the k-means heuristic method. Furthermore, through outlier detection, fault extraction, and fault concatenation, fault pattern names are obtained, and real-time detection of fault text data is achieved. Through periodic re-clustering, new fault patterns can be discovered by utilizing existing prior knowledge while ensuring that the original data points belong to the same cluster, thereby improving the accuracy and effectiveness of fault pattern generation and identification. Attached Figure Description

[0082] Figure 1 This is a control block diagram of the heuristic-based fault mode generation method for the aviation field according to the present invention.

[0083] Figure 2 The flowchart shows the heuristic-based fault mode generation method for the aviation field according to the present invention.

[0084] Figure 3 This is a distribution diagram of the fault data of the present invention in five fault modes;

[0085] Figure 4 This is a distribution diagram of the five fault modes of the newly added fault data in this invention. Detailed Implementation

[0086] Exemplary embodiments, features, and aspects of the present invention will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0087] This invention provides a heuristic-based method for generating failure modes in the aviation field, such as... Figure 1 This is a control block diagram of the heuristic-based aviation fault mode generation method of the present invention. It extracts features from aviation fault text data using a bag-of-words model and a word frequency-inverse text frequency index method; performs k-means heuristic clustering analysis of aviation fault text data; detects outliers, extracts faults, and concatenates fault data to obtain fault mode names; processes the obtained fault description text in real time; and generates aviation fault modes, including:

[0088] Step S1: Use the bag-of-words model and the term frequency-inverse text frequency index method to extract features from aviation fault text data;

[0089] The aviation-related fault text data mentioned in this embodiment of the invention uses the fault text data of a "fuel level sensor" device as an example, and its 1147 fault text data entries serve as an exemplary embodiment of the invention. In this example, 900 randomly sampled fault text data entries are used as existing data for clustering algorithms, and the remaining 247 fault text data entries are used as examples of the new data processing portion in this invention.

[0090] Obtain the set of fault description texts U = {u 1 ,u 2 ,…,u m}, then for all fault description text u i ∈U, perform word segmentation to obtain its word segmentation sequence. By combining the word segmentation results of all fault description texts, an aviation fault text data dictionary D is obtained, and an aviation text data set matrix V is constructed, specifically as follows:

[0091] Get all word segmentation sequences u i The set of words appearing in the text yields the following aviation-related fault text data dictionary D:

[0092] D = {w1, w2, ..., w g};

[0093] Where D is the aviation field fault text data dictionary; w1 is the first word in the dictionary; w2 is the second word in the dictionary; w g It is the g-th word in the dictionary.

[0094] Determine the word segmentation sequence u j The word frequency vectors are:

[0095]

[0096] in, For word segmentation sequence u i The word frequency vector.

[0097] The aviation text data set matrix V is constructed as follows:

[0098]

[0099] Among them, V m×g It is a matrix of aviation text data set with m rows and g columns.

[0100] Then, the term frequency-inverse text frequency index transformation is performed on the constructed aviation text data set matrix V, that is, for each element... The transformation is performed as follows:

[0101]

[0102] Among them, T(V) ij ) represents the term frequency-inverse text transformation matrix element of the fault description text data; V ij The elements of the fault description text data set matrix are: m, g, f(w); m is the total number of fault description text data; g is the total number of elements in the aviation field fault text data dictionary; f(w) i ) for statistics w i Number of occurrences; w i f is the i-th word in the dictionary; ui (w i ) for w i In sequence u i The number of times it appears in the text; i is the fault description text number; j is the word segmentation sequence number.

[0103] Then, the term frequency-inverse text transformation matrix elements T(V) of all aviation text data are processed. ij Perform regularization transformation, specifically:

[0104] According to T(V) ij The term frequency-inverse text transformation matrix of the constructed aviation text data is T(V), as shown below:

[0105]

[0106] Where T(V) is the word frequency-inverse text transformation matrix of the aviation text data.

[0107] Then for all elements T(V) ij Perform regularization as follows:

[0108]

[0109] Among them, N(T(V) ij The term frequency-inverse text transformation matrix element regularization result is represented by ||T(V) for the aviation text data. i ||2 is the 2-norm of the i-th row vector.

[0110] The 2-norm of the i-th row vector ||T(V) i ||2 is:

[0111]

[0112] Each row vector of the matrix is ​​used as a feature vector of the fault description text data, forming the data feature sample matrix X = {x1, x2, ..., x} of the aviation fault text. m}

[0113] Inputting 900 pieces of aviation-related text data, after word segmentation, the dictionary formed by these 900 pieces of data contains a total of 2501 words, and finally obtaining a 900×2501-dimensional data feature sample matrix, each row of the matrix corresponds to a 2501-dimensional feature vector for each piece of data.

[0114] Step S2: Clustering category analysis of aviation fault text data based on k-means heuristic;

[0115] Obtain a 900×2501 dimensional data feature sample matrix X = {x1, x2, ..., x} from the aviation field fault text in step S1. m The clusters are then grouped into K disjoint feature clusters. The K-value estimation method based on the data is used for this clustering.

[0116] K value ∈ {1,2,3,…,K} max The K value is the inertia value obtained from the clustering result calculated in step S2. This sequence monotonically decreases as the value of K increases; when the value of K is less than or equal to the true value of the number of categories K... * At that time, sequence μ K There will be a most significant reduction, as shown below:

[0117] {μ K -μ K+1 |K <K *}>>{μ K -μ K+1 |K≥K *};

[0118] Where, μ K μ is the inertia value of the clustering results when the number of categories is K; K+1 The inertia value is the clustering result when the number of categories is K+1; K is the number of categories; K * This represents the true number of categories.

[0119] Therefore, μ K The K corresponding to the inflection point of the monotonically decreasing process is used as the reference to the true value K. * The estimate;

[0120] Specifically, first calculate the sequence The first difference is divided into:

[0121] Δ 1 μ K =μ K+1 -μ K (K = 1, 2, ..., K) max -1);

[0122] Where, Δ 1 μ K The first difference of the inertia value of the clustering result.

[0123] Based on the first-order difference, the second-order difference is then calculated as follows:

[0124] Δ 2 μ K =Δ 1 μ K+1 -Δ 1 μ K (K = 1, 2, ..., K) max -2);

[0125] Where, Δ 2 μ K The second difference is the inertia value of the clustering result.

[0126] Obtain the second-order difference Δ 2 μ K The value is used as a reference to the true value K * Best estimate As shown below:

[0127]

[0128] in, For the true value K * The best estimate; argmax is the function that takes the maximum value of the function's arguments.

[0129] Set the maximum number of clusters K MAX The value is 20. Experimental calculations show that when K = 5, the second difference Δ of the inertia value is... 2 μ K The maximum value has been reached, therefore select... As an estimate of the true number of categories.

[0130] For each feature cluster c j ∈{c1,c2,…,c s The optimal feature cluster center vector μ is obtained by calculating the mean of the samples in the feature cluster and minimizing the objective value inertia. j , as shown below:

[0131]

[0132] Where inertia is the objective value to be minimized; μ j For the j-th feature cluster c j Center vector; x i s represents the elements of the data feature sample matrix; s represents the total number of feature clusters.

[0133] At this point, the center vectors {μ1, μ2, μ3, μ4, μ5} of five feature clusters were obtained, each of which is a 2501-dimensional vector. Simultaneously, 900 aviation-related text data samples were assigned to these five feature clusters, thus assigning feature cluster information to each aviation-related text data sample. Within these five feature clusters, each cluster contains 360, 225, 155, 89, and 71 aviation-related text data samples, respectively.

[0134] Step S3: Outlier detection, fault extraction, and fault concatenation of aviation fault text data to obtain fault mode names; input the cluster information of the five feature cluster centers and each aviation text data sample.

[0135] Step S31: Use box plots to detect outliers in aviation fault text data; element c in feature cluster set C. l to the feature cluster center vector μ j The distance is d l =||c l -μ j ||2, based on the distance {d1,d2,…,d s} Calculate the first quartile Q1 and the third quartile Q3, specifically as follows:

[0136] All distances {d1, d2, ..., d s The values ​​of} are arranged from smallest to largest and divided into four equal parts by three dividing points. The value at the first dividing point is the first quartile Q1, and the value at the third dividing point is the third quartile Q3.

[0137] The interquartile range (IQR) is obtained as follows:

[0138] IQR = Q3 - Q1;

[0139] Where IQR is the interquartile range; Q1 is the first quartile; and Q3 is the third quartile.

[0140] For distance d l If it is greater than the threshold Q3 + 1.5IQR, then d is considered to be... l The corresponding c l These are outliers and should be deleted.

[0141] Outlier detection was performed on each feature cluster, resulting in 9, 5, 3, 1, and 0 outliers for the five feature clusters, totaling 18 outliers. After deleting the outliers, the text data of each feature cluster was input into step 32 to extract the fault subject and fault phenomenon from the text description. For example, from the text description "During inspection, it was found that the fuel level sensor fixing screw was broken."

[0142] Step S32: Nested named entity recognition is achieved using a dual affine model to extract the fault subject and phenomenon; the input fault description text is processed by a bidirectional encoder based on a transformer and a BiLSTM layer to extract its vector representation, specifically:

[0143] First, the beginning h of each valid named cover J is calculated using two independent feedforward neural networks. s and the ending h e Specifically:

[0144]

[0145] Among them, h s (J) is a valid name that covers the beginning of J; h e (J) is a valid name overwrite of the end of J; FFNN s For the first feedforward neural network; FFNN e This is the second feedforward neural network; This serves as the input to the first feedforward neural network; This is the input to the second feedforward neural network; s J and the starting index for effective naming coverage J; e J It is the end indicator of effective named coverage J.

[0146] Then, the score of the effective named cover J is calculated using a biaffine model, resulting in the following score tensor:

[0147]

[0148] Where, r m (J) is a valid name covering J's c r Fractional tensor of dimension U; m For the biaffine model d r ×c r ×d r tensor; W m For a 2d double affine model r ×c r The matrix; b m This represents the bias of the double affine model; d is the direct sum of the matrices; r c represents the dimension of the vector output by the LSTM layer. r This represents the number of entity categories.

[0149] Finally, the fractional tensor r m This includes the score for each valid named cover J belonging to each entity category, and the entity category for each valid named cover J is labeled as:

[0150] y'(J)=argmax rm (J);

[0151] Where y'(J) is the valid named coverage of the entity category marked by J.

[0152] For each valid named coverage J of the fault description text, calculate its predicted probability for each category, and classify it using a loss function of cross-entropy with a normalized exponential function, as shown below:

[0153]

[0154] Where loss is the cross-entropy loss function; l is the entity category index; L is the total number of entity categories; and J is the effective naming coverage. A set that covers all valid named elements; To efficiently name the symbolic function covering entity category l for J; p m (J l ) represents the predicted probability of J for category l.

[0155] Effective naming coverage J l loss parameter p m (J l Specifically:

[0156]

[0157] Where, p m (J l ) represents the predicted probability of J for category l; This is a summation index for entity category serial numbers.

[0158] In step S32, the faulty component "fixing screw" and the faulty phenomenon "breakage" are extracted, and all data samples are extracted in step S32.

[0159] Step S33: Concatenate all fault topics and fault phenomena from the fault description text data in the aviation field; after obtaining all fault subjects and fault phenomenon phrases in each fault description text of each feature cluster, concatenate them as the fault mode name of this feature cluster; for the set A = {a} of all fault subject phrases and the set P = {p} of fault phenomenon phrases in the extracted feature clusters, consider all concatenation combinations N = {(a,p)|a∈A,p∈P}, for any (a,p)∈N, set the concatenation score as:

[0160] score((a,p))=c A (a)+c P (p)+c AP (a,p);

[0161] Where score((a,p)) is the concatenated score result; a is the fault topic phrase; p is the fault phenomenon phrase; c A (a) represents the number of times the faulty phrase 'a' appears in this cluster; c P (p) represents the number of times the fault phenomenon phrase p appears in this cluster; c AP (a,p) represents the number of times the fault subject phrase a and the fault phenomenon phrase p appear simultaneously in this cluster, obtaining all combinations in the feature cluster.

[0162] The highest-scoring fault mode was selected as the fault mode name for that feature cluster. The five fault mode names obtained were: incorrect fuel gauge reading, fuel gauge oscillation, fuel measurement system malfunction, no fuel depletion signal in the engine compartment auxiliary fuel tank, and broken mounting screw. The distribution of fault data among these five fault modes is as follows: Figure 3 The figure shows the distribution of fault data in the five fault modes of the present invention; there were 351 faults in the fuel measurement system, 220 faults in the fuel level indicator, 71 faults in the fuel level indicator, 152 faults in the auxiliary fuel tank of the engine body, 88 faults in the fixing screw, and 18 outliers.

[0163] Step S4: Process the obtained fault description text in real time to generate aviation-related fault modes;

[0164] To obtain newly added fault description text in real time, firstly, step S1 is performed to obtain the feature matrix of fault text data in the aviation field; then, according to step S2, clustering category analysis of fault text data in the aviation field is performed; finally, the fault mode name is obtained through step S3, and the fault mode in the aviation field is generated.

[0165] After generating aviation failure modes, new failure data types will emerge as the data accumulates. Therefore, existing clustering results need to be updated periodically, involving re-clustering of new and existing data. Re-clustering should only reflect the new failure modes in the newly added data. After processing the existing 900 aviation text data samples, the 247 newly added data samples were processed, classifying them into these 5 failure modes, as shown in the distribution below. Figure 4 The figure shown is a distribution diagram of the five fault modes of the newly added fault data of the present invention; there were 101 faults in the fuel measurement system, 56 faults in the fuel level indicator, 26 faults in the fuel level indicator, 46 faults in the auxiliary fuel tank of the engine body, and 18 faults in the fixing screw.

[0166] like Figure 2The diagram shows a flowchart of the heuristic-based aviation fault mode generation method of the present invention; it details the flowchart of the processing and detection of fault text data in the field, and also describes the detection process of newly added data, which proves that the method achieves fault text data detection in the aviation field.

[0167] The beneficial effects of this invention are as follows: In this embodiment, the bag-of-words model, the word frequency-inverse text frequency index method, and the k-means heuristic method are used to complete the clustering category analysis of fault text data. 900 sets of aviation-related text data were processed and analyzed, resulting in 5 data feature groups. Furthermore, through outlier detection, fault extraction, and fault concatenation, fault pattern names are obtained, and real-time detection of fault text data is achieved. Periodic re-clustering can utilize existing prior knowledge or existing datasets to discover new fault patterns while ensuring that the original data points belong to the same cluster, thus improving the accuracy and effectiveness of fault pattern generation and identification. The embodiments demonstrate that this method can meet practical application requirements and has good application results.

[0168] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A heuristic based method for generating failure modes in the field of aviation, characterized in that: It includes: S1: Use the bag-of-words model and the term frequency-inverse text frequency index method to extract features from fault text data in the aviation field; Acquire a collection of fault description texts in the aviation field For all fault description text Perform word segmentation to obtain its word segmentation sequence. By combining the word segmentation results of all fault description texts, a fault text data dictionary for the aviation field is obtained. Construct an aviation text data set matrix ; To construct the matrix of the aviation text data set The term frequency-inverse document frequency index method is transformed, that is, each element is transformed into: ； in, These are the elements of the term frequency-inverse text transformation matrix for the fault description text data; The elements of the matrix containing the fault description text data set; The total number of fault description text data; The total number of dictionary elements for fault text data in the aviation field; For statistics Number of times it appears; The i-th word in the dictionary; for In sequence The number of times it appears in; Number the fault description text; Number the word segmentation sequence; Term frequency-inverse text transformation matrix elements for all aviation text data Perform regularization transformation to obtain each row vector of the matrix, which serves as a feature vector for fault description text data; and form a data feature sample matrix for aviation fault text. ; S2: Clustering category analysis of aviation fault text data based on k-means heuristic; Obtain the data feature sample matrix of aviation field fault text in step S1. Cluster them into There are three mutually exclusive feature clusters. For each feature cluster... The optimal feature cluster center vector is obtained by calculating the mean of samples in the feature cluster and minimizing the objective value. ; S3: Outlier detection, fault extraction, and fault concatenation of fault text data in the aviation field to obtain fault mode names; S31: Complete outlier detection in aviation fault text data; feature cluster set. medium elements to the feature cluster center vector The distance is According to distance Calculate the first quartile. and the third and fourth quartiles And the interquartile range is obtained as: ； wherein, is the interquartile range; is the first quartile; is the third quartile; For the distance , if it is greater than a threshold , it is considered that the corresponding is an outlier, and should be deleted. S32: A dual affine model is used to realize nested named entity recognition, extracting the fault subject and phenomenon; the input fault description text is processed by a transformer-based bidirectional encoder representation encoder and a BiLSTM layer for feature extraction to obtain its vector representation, thus obtaining each valid named overlay of the fault description text. The predicted probability for each category is calculated, and classification is performed using a loss function of cross-entropy with a normalized exponential function, as shown below: ； in, Let be the loss function for cross-entropy; For entity category serial number; Total number of entity categories; For effective naming overriding; A set that covers all valid named elements; For effective naming coverage For entity categories The sign function; To cover Category The predicted probability; S33: Concatenate all fault topics and fault phenomena in the fault description text data of the aviation field; after obtaining all fault subjects and fault phenomenon phrases in each fault description text in each feature cluster, concatenate them as the fault mode name of this feature cluster. S4: Processes the obtained fault description text in real time and generates fault modes for the aviation field.

2. The heuristic based aviation domain fault pattern generation method as claimed in claim 1, wherein: The step S1 combines the word segmentation results of all fault description texts to obtain an aviation field fault text data dictionary , and constructs an aviation text data set matrix , and specifically First, obtain all the segmented sequences The set of words appearing in the middle, get the aviation domain fault text data dictionary For: ； in, A dictionary of fault text data for the aviation industry; It is the first word in the dictionary; It is the second word in the dictionary; The g-th word in the dictionary; Then, the word frequency vector of the segmented sequence is determined is: ； wherein, is a word frequency vector of the segmented sequence . Finally, the aviation text data set matrix is constructed is: ； wherein, is a a matrix of sets of aviation text data.

3. The heuristic based aviation domain fault pattern generation method as claimed in claim 1, wherein: The step S1 of the word frequency-inverse text transformation matrix element of all the aviation text data The regularization transformation is performed, specifically: According to The word frequency-inverse document transformation matrix of the constructed aviation text data is As follows: ； wherein, is the term frequency-inverse document transformation matrix for the aviation text data; Then for all elements Perform regularization as follows: ； wherein, is the regularization result of the word frequency-inverse text transform matrix element of the aviation text data; is the 2-norm of the i-th row of the row vector; is the 2-norm of the row vector; The first 2-norm of a row vector is: 。 4. The heuristic based aviation domain fault pattern generation method as claimed in claim 2, wherein: The step S2 clusters to The data-based estimation method is used for the individual disjoint feature cluster, specifically: The ,Should Based on the clustering results calculated in step S2 Value This sequence is following The value increases and then monotonically decreases; when the K value is less than or equal to the true value of the number of categories. At that time, sequence There will be a most significant reduction, as shown below: ； wherein, is the inertia value of the clustering result when the number of classes is K; is the inertia value of the clustering result when the number of classes is K+1; is the number of classes; is the true value of the number of classes; Therefore, The K corresponding to the inflection point of the monotonically decreasing process is used as the true value. The estimate; Specifically, first, the first-order difference of the sequence is calculated as: ； wherein, is the first order difference of the inertia value of the clustering result; Based on the first-order difference, the second-order difference is then calculated as follows: ； wherein, is the second order difference of the inertia value of the clustering result; Obtaining the second-order difference The value is used as a reference to the true value Best estimate As shown below: ； wherein is the best estimate of the true value ; and is the function that takes the parameter value for which the function value is maximum.

5. The heuristic based aviation domain fault pattern generation method as claimed in claim 1, wherein: The distance in step S31 Calculate the first quartile. and the third and fourth quartiles Specifically: Arrange all the distances in ascending order and divide them into four equal parts by three split points. The value at the first split point is the first quartile , and the value at the third split point is the third quartile .

6. The heuristic based aviation domain fault pattern generation method as claimed in claim 3, wherein: The fault description text input in step S32 is processed by a converter-based bidirectional encoder representation encoder and a BiLSTM layer, specifically as follows: First, the beginning and end of each valid named entity coverage is computed by two independent feed-forward neural networks , in particular:​​ ； in, For effective naming coverage The beginning; For effective naming coverage The ending; This is the first feedforward neural network; This is the second feedforward neural network; This serves as the input to the first feedforward neural network; This serves as the input to the second feedforward neural network. and effective naming coverage The initial indicators; It is effective naming overriding The end indicator; Then, the fraction of effective named coverage is calculated using the bi-affine model resulting in the following score tensor: ； in, For effective naming coverage of Fractional tensors; For a double affine model tensor; For a double affine model Matrix; This represents the bias of the double affine model; It is the direct sum of the matrices; The dimension of the vector output by the LSTM layer; The number of entity categories; Finally, fractional tensors Includes each valid named overlay The score belonging to each entity category, for each valid named coverage. The entity category is marked as: ； wherein, to effectively name the overlay tagging entity classes.

7. The heuristic based aviation domain fault pattern generation method as claimed in claim 1, wherein: The effective naming in the step S32 covers loss parameters In particular: ； wherein, is the coverage for the prediction probability of the class ; is the sum indicator of the entity class number.

8. The heuristic based aviation domain fault pattern generation method as claimed in claim 3, wherein: The step S33, which involves splicing together all fault topics and fault phenomena from aviation-related fault text data, specifically includes: The set of all fault subject phrases in the extracted feature clusters and a collection of fault phenomenon phrases At this point, consider all possible combinations. For any The concatenated score is set as follows: ； in, To concatenate the fractional results; For fault-related phrases; These are short phrases describing the fault symptoms; The fault subject phrase The number of times it appears in this cluster; Fault phenomenon phrase The number of times it appears in this cluster; The fault subject phrase and fault symptoms phrases The number of times a feature appears in the same cluster is used to obtain all combinations in the feature cluster; The one with the highest score is selected as the fault mode name for that feature cluster.

9. The heuristic based aviation domain fault pattern generation method as claimed in claim 1, wherein: After generating aviation-related fault modes in step S4, new fault data types will emerge in the aviation-related fault text data as the acquired data accumulates. Existing clustering results need to be updated periodically, and new data should be re-clustered with the original data; the re-clustering should only reflect new fault modes on the newly added data.

Images