Civil aviation inspector competence analysis method and system based on big data

Abstract

The application discloses a kind of big data-based civil aviation inspector competency analysis method and system.Recorded inspector behavior data of multiple time points is used to distinguish the competency of inspector.Secondly, find the phrase related to competency, through the first associated neural network, the first associated convolution network and the fusion neural network, using two-dimensional and one-dimensional way respectively detect the association between multiple monitoring phrases.According to the monitoring phrase association probability value, construct a graph structure, increase the importance of the inspector behavior with an association.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a civil aviation inspector competence analysis method and system based on big data. BACKGROUND

[0002] At present, competence is the ability to apply or use knowledge, skills, abilities, behaviors and personal characteristics to successfully perform key work tasks, specific functions or operate in a given role or position. In addition, from the definition of competence, it can be clearly seen that competence contains the following characteristics: first, individual characteristics, that is, the inherent attributes of a person, which play a decisive role in the individual's thinking and behavior, including knowledge, skills, characteristics, motivation, etc.; second, behavior characteristics, that is, the specific application of individual characteristics in a specific situation; third, situational conditions, which can be understood as the external situation of the job, industry and cultural environment.

[0003] If only one time point of the inspector's behavior data is recorded, the recorded behavior will not be accurate enough due to different ideas or accidental events of the inspector. There is a correlation between multiple behavior data, and one behavior data will change with the change of another behavior data, both of which will result in a greater effect on competence determination. If the above problems are not considered, the competence determination will not be accurate enough. SUMMARY

[0004] The purpose of the present application is to provide a civil aviation inspector competence analysis method and system based on big data, to solve the above problems existing in the prior art.

[0005] In a first aspect, the embodiments of the present application provide a civil aviation inspector competence analysis method based on big data, comprising:

[0006] Obtain inspector behavior data at m time points; the inspector behavior data represents a statement describing the behavior of a civil aviation inspector;

[0007] According to the inspector behavior data at m time points, detect the word groups related to competence to obtain n inspector word groups; the inspector word group is a noun;

[0008] Obtain n*(n-1) inspector word group correlation probability values based on the n inspector word groups through a correlation word group network; the inspector word group correlation probability value represents the probability that one inspector word group appears in the inspector behavior data and another inspector word group also appears;

[0009] The correlation word group network comprises a first correlation neural network, a first correlation convolution network and a fusion neural network;

[0010] Based on the supervisor behavior data and the n*(n-1) supervision phrase association probability values, a graph structure is constructed to obtain a supervision connection matrix; the supervision connection matrix represents a graph structure with the n supervision phrases as vertices and the supervision phrase association probability values as edges;

[0011] The supervisor competence is obtained based on the supervision connection matrix through the competence judgment network.

[0012] Optionally, the n*(n-1) supervision phrase association probability values are obtained based on the n supervision phrases through the association phrase network, including:

[0013] An adjustment phrase matrix is obtained based on the n supervision phrases; the adjustment phrase matrix represents the order of the supervision phrases in the supervisor behavior data;

[0014] The adjustment phrase matrix is divided by column to obtain k order phrase vectors; one order phrase vector corresponds to one order in the supervisor behavior data;

[0015] The k order phrase vectors are input into a first association neural network to obtain a first feature vector; the first feature vector represents a feature vector of the probability of the appearance of one supervision phrase after multiple supervision phrases in the supervisor behavior data;

[0016] A second feature map is obtained based on the adjustment phrase matrix through the first association convolutional network; the second feature map represents a feature map of the probability of the appearance of one supervision phrase after multiple supervision phrases in the supervisor behavior data;

[0017] The first feature vector and the second feature map are input into a fusion neural network to obtain the n*(n-1) supervision phrase association probability values.

[0018] Optionally, all neurons corresponding to one order phrase vector in the input layer of the first association neural network are connected with one neuron of the first layer of the hidden layer;

[0019] All neurons of the first layer of the hidden layer in the first association neural network are connected with the first neuron of the second layer of the hidden layer;

[0020] The neurons of the first layer of the hidden layer of the first association neural network, except the first [1, i-1] neurons, are connected with the i th neuron of the second layer of the hidden layer; the value range of i is [2, k];

[0021] The first association neural network has z hidden layers.

[0022] Optionally, the second feature map is obtained based on the adjustment phrase matrix through the first association convolutional network, including:

[0023] obtaining a multi-scale convolution kernel; the multi-scale convolution kernel comprises a two-dimensional convolution kernel with a size of 2*2 and a size of a prime number;

[0024] using the multi-scale convolution kernel, respectively, convolving the adjusted word group matrix with a step length of 1 to detect the association relationship between the supervision word groups in different orders, and obtaining a multi-scale feature map;

[0025] performing feature fusion on the multi-scale feature map through 2*2 convolution kernel down-sampling to obtain a second feature map.

[0026] Optionally, the training method of the association word group network comprises:

[0027] obtaining a training adjusted word group matrix; the training adjusted word group matrix represents the historical adjusted word group matrix of a plurality of civil aviation supervisors;

[0028] obtaining labeled association data; the labeled association data represents the probability that one supervision word group appears in the supervisor behavior data and another supervision word group also appears in the historical plurality of civil aviation supervisors;

[0029] inputting the training adjusted word group matrix into the first association neural network to obtain a first training feature vector;

[0030] the first training feature vector is used as a first training prediction vector; the first training prediction vector represents the probability that a plurality of supervision word groups appear after one supervision word group predicted by the first association neural network;

[0031] calculating the loss between the labeled association data and the first training prediction vector to train the first association neural network;

[0032] inputting the training adjusted word group matrix into the first association convolution network to obtain a second training feature map;

[0033] inputting the second training feature map into the second full connection network to obtain a second training prediction vector; the second training prediction vector represents the probability that a plurality of supervision word groups appear after one supervision word group predicted by the first association convolution network;

[0034] calculating the loss between the labeled association data and the second training prediction vector to train the first association convolution network;

[0035] inputting the first training feature vector and the second training feature map into the fusion neural network to obtain a training supervision word group association probability vector;

[0036] calculating the loss between the labeled association data and the training supervision word group association probability vector to train the fusion neural network, the first association convolution network and the first association neural network.

[0037] Optionally, the monitoring connection matrix is input into the first discriminant network to detect the competencies corresponding to the plurality of monitoring word groups of the association relationship, and a first competency feature vector is obtained.

[0038] The monitoring word groups of the monitoring officer behavior data are taken as vertices, and the monitoring word group association probability values are taken as edges to obtain a monitoring word group graph.

[0039] The vertices are taken as rows and columns respectively to construct the monitoring connection matrix.

[0040] If the monitoring word group association probability is greater than an association threshold, the monitoring word group association probability is retained; if the monitoring word group association probability is less than or equal to the association threshold, the monitoring word group association probability is set to 0.

[0041] According to the monitoring word groups corresponding to the rows and columns of the monitoring connection matrix, the monitoring word group association probability values are filled into the monitoring connection matrix.

[0042] Optionally, the monitoring officer competency is obtained based on the monitoring connection matrix through the competency discriminant network, and the competency discriminant network comprises a first discriminant network, a first competency neural network, a 2*n two-dimensional convolutional neural network, and a 2*2 two-dimensional convolution kernel.

[0043] The competency discriminant network comprises a first discriminant network, a first competency neural network, a 2*n two-dimensional convolutional neural network, and a 2*2 two-dimensional convolution kernel.

[0044] The monitoring connection matrix is input into the first discriminant network to detect the competencies corresponding to the plurality of monitoring word groups of the association relationship, and a first competency feature vector is obtained.

[0045] In the column direction, the monitoring connection matrix is convolved with a 2*n two-dimensional convolution kernel at a step of 1 to obtain a first association feature.

[0046] In the monitoring officer behavior data, a descriptive word corresponding to the monitoring word group is found as a monitoring descriptive word.

[0047] The monitoring descriptive word is matched with the descriptive words stored in the monitoring database to obtain a monitoring degree value; the monitoring degree value represents the degree of competency discrimination of the monitoring word group.

[0048] According to the arrangement order of the vertices in the connection matrix, the monitoring degree values corresponding to the monitoring word groups are constructed into a monitoring word group degree set, and the monitoring word group degree set is input into the first competency neural network to obtain a second competency feature vector.

[0049] The first competency feature vector and the second competency feature vector are constructed into a first detection feature map.

[0050] In the wide direction, a 2*2 two-dimensional convolution kernel is used to convolve the first detection feature map to obtain the monitoring officer competency.

[0051] Optionally, the obtaining the adjusted word group matrix based on the n supervision word groups comprises:

[0052] annotating the order of appearance of the n supervision word groups according to the supervisor behavior data to obtain an order matrix; the columns of the order matrix represent the n supervision word groups, and the rows represent time points; and the values in the order matrix represent the order of appearance of a supervision word group at a time point in the supervisor behavior data;

[0053] rearranging the supervision word groups with the same order in the order matrix to obtain the adjusted word group matrix; the columns of the adjusted word group matrix represent the order of appearance of the supervision word groups, and the rows represent time points; and the values in the adjusted word group matrix represent the supervision word groups corresponding to different orders of appearance at a time point.

[0054] Optionally, the detecting the word groups related to the competency according to the supervisor behavior data at the m time points to obtain the n supervision word groups comprises:

[0055] obtaining a supervision database; the supervision database stores a plurality of word groups related to the competency;

[0056] detecting the word groups in the supervisor behavior data at the m time points in the supervision database as the supervision word groups.

[0057] In a second aspect, an embodiment of the present application provides an analysis system for the competency of civil aviation supervisors based on big data, comprising:

[0058] an obtaining module configured to obtain supervisor behavior data at m time points; the supervisor behavior data represents sentences describing the behaviors of civil aviation supervisors;

[0059] a word group module configured to detect word groups related to the competency according to the supervisor behavior data at the m time points to obtain n supervision word groups; the supervision word groups are nouns;

[0060] an association module configured to obtain n*(n-1) supervision word group association probability values based on the n supervision word groups through an association word group network; the supervision word group association probability values represent the probability that one supervision word group appears in the supervisor behavior data and another supervision word group also appears;

[0061] the association word group network comprises a first association neural network, a first association convolutional network and a fusion neural network;

[0062] a graph structure module configured to construct a graph structure based on the supervisor behavior data and the n*(n-1) supervision word group association probability values to obtain a supervision connection matrix; the supervision connection matrix represents a graph structure with the n supervision word groups as vertices and the supervision word group association probability values as edges;

[0063] The competency judgment module is configured to obtain the competency of the supervisor based on the supervision connection matrix through the competency judgment network.

[0064] Compared with the prior art, the embodiment of the present application has the following beneficial effects:

[0065] The embodiment of the present application also provides a civil aviation supervisor competency analysis method and system based on big data.

[0066] In the present application, the behavior data of the supervisor recorded at multiple time points is used to judge the competency of the supervisor, which reduces the influence of inaccurate records caused by different ideas or accidental events of the supervisor, thereby reducing the misjudgment of the competency. Secondly, the word groups related to the competency are found, and the first association neural network, the first association convolution network and the fusion neural network are used to judge the probability of the appearance of one monitoring word group after a monitoring word group in the supervisor behavior data in a two-dimensional and one-dimensional manner, detect the association relationship between the multiple monitoring word groups, and detect whether there is a relationship between the different behaviors of the supervisor for competency judgment. According to the monitoring word group association probability value, a graph structure is constructed, the importance of the behaviors of the supervisor with the association relationship is increased, and the adjectives describing the monitoring word groups are used to detect the judgment state of each behavior on the competency. The technical effect of more accurately detecting the competency of the civil aviation supervisor is achieved. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 is a civil aviation supervisor competency analysis method based on big data provided by the embodiment of the present application.

[0068] Figure 2 is a structure diagram of the first association neural network in the civil aviation supervisor competency analysis method based on big data provided by the embodiment of the present application. DETAILED DESCRIPTION

[0069] The present application will be described in detail below with reference to the accompanying drawings.

[0070] Embodiment 1

[0071] As shown in the drawings, the embodiment of the present application provides a civil aviation supervisor competency analysis method based on big data, which comprises: Figure 1

[0072] S101: Obtain the behavior data of the supervisor at m time points; the behavior data of the supervisor represents the sentences describing the behavior of the civil aviation supervisor.

[0073] Wherein, m is a positive integer greater than 2.

[0074] ​The supervisor behavior data of the m time points is randomly obtained from the supervisor behavior data of different time points.

[0075] The supervisor behavior data of the m time points can be used to determine the behavior of the supervisor by using long-term and random observation data.

[0076] S102: According to the supervisor behavior data of the m time points, the competency-related phrase is detected, and n supervision phrases are obtained. The supervision phrase is a noun.

[0077] Wherein, n is a positive integer.

[0078] Wherein, a plurality of competency-related phrases are stored in the database, and the competency-related phrases in the supervisor behavior data are matched.

[0079] Wherein, the n supervision phrases represent all the different competency-related phrases in the supervisor behavior data of the m time points.

[0080] S103: By associating the phrase network, based on the n supervision phrases, n*(n-1) supervision phrase association probability values are obtained; the supervision phrase association probability value represents the probability that one supervision phrase appears in the supervisor behavior data and another supervision phrase also appears.

[0081] The associated phrase network includes a first associated neural network, a first associated convolutional network, and a fusion neural network.

[0082] Wherein, the structure diagram of the first associated neural network is as shown in Figure 2 .

[0083] S104: Based on the supervisor behavior data and the n*(n-1) supervision phrase association probability values, a graph structure is constructed, and a supervision connection matrix is obtained; the supervision connection matrix represents a graph structure with n supervision phrases as vertices and supervision phrase association probability values as edges.

[0084] S105: By the competency judgment network, based on the supervision connection matrix, the supervisor competency is obtained.

[0085] Optionally, the n*(n-1) supervision phrase association probability values obtained by the associated phrase network based on the n supervision phrases include:

[0086] Based on the n supervision phrases, an adjustment phrase matrix is obtained; the adjustment phrase matrix represents the order of the supervision phrases appearing in the supervisor behavior data;

[0087] The adjustment phrase matrix is divided by column to obtain k order phrase vectors; one order phrase vector corresponds to one order appearing in the supervisor behavior data.

[0088] In this embodiment, k is a positive integer, and k represents the number of order labels. For example, if the order is divided into the first order, the second order, and the third order, then k is 3. The order of the order phrase vector corresponding to the first column in the phrase matrix is adjusted to the first order. The order of the order phrase vector corresponding to the second column in the phrase matrix is adjusted to the second order. The order of the order phrase vector corresponding to the third column in the phrase matrix is adjusted to the third order.

[0089] In this embodiment, the values in the order phrase vector represent the same order of the occurrence of the monitoring phrase in the monitor behavior data at the plurality of time points.

[0090] The k order phrase vectors are input into the first association neural network to obtain a first feature vector. The first feature vector represents a feature vector of the probability of the occurrence of a monitoring phrase after a monitoring phrase in the monitor behavior data.

[0091] In this embodiment, the first association neural network is a deep neural network (DNN).

[0092] In this embodiment, the number of elements in the first feature vector is a fixed value, which is 100 in this embodiment. The first n*(n-1) elements are used to represent the association relationship of the monitoring phrase. If the monitoring phrases related to competence detected in the monitor behavior data of the m time points are three monitoring phrases labeled as A, B, and C, and A is the first order, B is the second order, and C is the third order, then the first feature vector represents the probability of the occurrence of B after A, the probability of the occurrence of C after A, the probability of the occurrence of A after B, the probability of the occurrence of C after B, the probability of the occurrence of A after C, and the probability of the occurrence of A after C. Therefore, the first feature vector uses the first 3*(3-2)=6 elements to represent the association relationship between the three monitoring phrases.

[0093] The second feature map is obtained based on the adjusted phrase matrix through the first association convolutional network. The second feature map represents a feature map of the probability of the occurrence of a monitoring phrase after a monitoring phrase in the monitor behavior data.

[0094] In this embodiment, the first association convolutional network is a convolutional neural network (CNN).

[0095] The first feature vector and the second feature map are input into the fusion neural network to obtain n*(n-1) monitoring phrase association probability values.

[0096] The fusion neural network is a fully connected neural network (FCNN).

[0097] The supervision phrase association probability represents a probability that one supervision phrase appears after another supervision phrase.

[0098] Optionally, all neurons corresponding to one sequential phrase vector in the input layer of the first association neural network are connected to one neuron of the first layer of the hidden layer.

[0099] In this embodiment, the number of elements in the sequential phrase vector is m, so that m neurons in the input layer of the first association neural network correspond to one neuron of the first layer of the hidden layer.

[0100] All neurons of the first layer of the hidden layer in the first association neural network are connected to the first neuron of the second layer of the hidden layer.

[0101] The neurons of the first layer of the hidden layer in the first association neural network, except the first [1, i-1] neurons, are connected to the i-th neuron of the second layer of the hidden layer; the value range of i is [2, k].

[0102] In this embodiment, the neurons of the first layer of the hidden layer in the first association neural network, except the first neuron, are connected to the second neuron of the second layer of the hidden layer; the neurons of the first layer of the hidden layer in the first association neural network, except the first and second neurons, are connected to the third neuron of the second layer of the hidden layer; the neurons of the first layer of the hidden layer in the first association neural network, except all neurons between the first neuron and the k-1th neuron, are connected to the k-th neuron of the second layer of the hidden layer.

[0103] The first association neural network has z hidden layers.

[0104] In this embodiment, z is 6.

[0105] The hidden layer is a hidden layer connected to the input layer, and the multiple hidden layers and the hidden layer and the output layer are fully connected.

[0106] Optionally, the second feature map is obtained based on the adjustment phrase matrix through the first association convolutional network, including:

[0107] The multi-scale convolution kernel is obtained; the multi-scale convolution kernel includes a two-dimensional convolution kernel with a size of 2*2 and a size of a prime number.

[0108] Wherein, in the embodiment, the size of the two-dimensional convolution kernel is a prime number, and the two-dimensional convolution kernel is 3*3, 5*5, and 7*7.

[0109] The multi-scale convolution kernel is used to respectively perform convolution on the adjusted phrase matrix with a step of 1, to detect the association relationship between the monitoring phrases in different orders, and to obtain a multi-scale feature map.

[0110] Wherein, in the embodiment, if a convolution kernel with a size of 2*2 is used, it means that the association relationship of the monitoring phrases in the first order and the second order, the second order and the third order, the third order and the fourth order, etc. is found. And the convolution kernel with a size of 2*2 means that the features that can be obtained by other even-size convolution kernels can be obtained in the convolution process, so only the convolution kernel with a prime number size is added. If a convolution kernel with a size of 3*3 is used, it means that the association relationship of the monitoring phrases in the first order and the second order and the third order, the second order and the third order and the fourth order, the third order and the fourth order and the fifth order, etc. is found, so that the association relationship of the monitoring phrases in the first order and the third order, the second order and the fourth order, the third order and the fifth order, etc. which is not easy to find by the 2*2 convolution kernel can be found.

[0111] The multi-scale feature map is down-sampled by a 2*2 convolution kernel to perform feature fusion, to obtain a second feature map.

[0112] Wherein, the second feature map represents the association relationship of the monitoring phrases in multiple orders.

[0113] Wherein, in the embodiment, the 2*2 convolution kernel is convolved multiple times, so as to achieve down-sampling. The size of the feature map obtained after the convolution of the largest prime number convolution kernel.

[0114] Optionally, the training method of the association phrase network comprises:

[0115] An adjusted phrase matrix is obtained, and the adjusted phrase matrix represents the historical adjusted phrase matrix of multiple civil aviation inspectors.

[0116] An association data is obtained, and the association data represents the probability that one monitoring phrase appears in the inspector behavior data and another monitoring phrase also appears in the historical multiple civil aviation inspectors.

[0117] The training adjusted phrase matrix is input into a first association neural network to obtain a first training feature vector.

[0118] Wherein, the first training feature vector and the first feature vector respectively represent the data in the training and the data in the detection.

[0119] The first training feature vector is taken as a first training prediction vector; the first training prediction vector represents a probability of a plurality of supervision word groups predicted by the first correlation neural network appearing after one supervision word group in the supervisor behavior data;

[0120] The labeled correlation data and the first training prediction vector are used to calculate a loss, and the first correlation neural network is trained.

[0121] In this embodiment, the loss is calculated by using a cross-entropy loss function.

[0122] The training adjustment word group matrix is input into the first correlation convolutional network to obtain a second training feature map.

[0123] The second training feature map and the second training feature map respectively represent data in the training and data in the detection.

[0124] The second training feature map is input into a second fully connected network to obtain a second training prediction vector; the second training prediction vector represents a probability of the plurality of supervision word groups predicted by the first correlation convolutional network appearing after one supervision word group.

[0125] The second fully connected network is a fully connected neural network (FCNN).

[0126] The second training prediction vector represents an output of the second fully connected network. The second training feature map is stretched into a one-dimensional vector from top to bottom and from left to right, and is input into the second fully connected network for classification to predict the probability of the plurality of supervision word groups appearing after one supervision word group.

[0127] The labeled correlation data and the second training prediction vector are used to calculate a loss, and the first correlation convolutional network is trained.

[0128] The loss is calculated by using a cross-entropy loss function.

[0129] The first training feature vector and the second training feature map are input into a fusion neural network to obtain a training supervision word group correlation probability vector;

[0130] The labeled correlation data and the training supervision word group correlation probability vector are used to calculate a loss, and the fusion neural network, the first correlation convolutional network and the first correlation neural network are trained.

[0131] The loss is calculated by using a cross-entropy loss function.

[0132] Optionally, based on the supervisor behavior data and n*(n-1) supervision word group correlation probability values, a graph structure is constructed to obtain a supervision connection matrix, including:

[0133] The supervision phrase graph of the supervisor behavior data is obtained by taking a supervision phrase as a vertex and a supervision phrase correlation probability value as an edge.

[0134] The supervision connection matrix is constructed by taking the vertices as rows and columns.

[0135] The supervision connection matrix is a connection matrix used to represent a graph structure.

[0136] If the supervision phrase correlation probability is greater than a correlation threshold, the supervision phrase correlation probability is retained; if the supervision phrase correlation probability is less than or equal to the correlation threshold, the supervision phrase correlation probability is set to 0.

[0137] In this embodiment, the correlation threshold is 0.5.

[0138] The supervision phrase correlation probability value is filled in the supervision connection matrix according to the supervision phrases corresponding to the rows and columns of the supervision connection matrix.

[0139] Optionally, the supervisor competence is obtained based on the plurality of supervision connection matrices by using the competence determination network, and the competence determination network comprises:

[0140] The competence determination network comprises a first determination network, a first competence neural network, a 2*n two-dimensional convolutional neural network, and a 2*2 two-dimensional convolution kernel.

[0141] The 2*n two-dimensional convolutional neural network (CNN) and the 2*2 two-dimensional convolution kernel.

[0142] The supervision connection matrix is input into the first determination network to detect the competence corresponding to the plurality of supervision phrases of the correlation relationship, and a first competence feature vector is obtained.

[0143] The first determination network is a convolutional neural network (CNN).

[0144] The supervision connection matrix is convolved with the 2*n two-dimensional convolution kernel in the column direction at a step of 1 to obtain a first correlation feature.

[0145] In the supervisor behavior data, an adjective corresponding to the supervision phrase is found as a supervision adjective.

[0146] The supervision adjective is matched with the adjectives stored in the supervision database to obtain a supervision degree value, and the supervision degree value represents the degree of competence determination of the supervision phrase.

[0147] The monitoring degree value is a value associated with the stored adjective in the database.

[0148] In this embodiment, the monitoring degree value is marked by 1, 2, and 4 to represent unqualified, qualified, and excellent, respectively. For example, the monitoring adjective “cannot” represents unqualified, “can” represents qualified, and “better” represents excellent.

[0149] According to the arrangement order of the vertices in the connection matrix, the monitoring degree values corresponding to the monitoring phrase are formed into a monitoring phrase degree set, and the monitoring phrase degree set is input into the first competence neural network to obtain a second competence feature vector.

[0150] The number of elements of the first competence feature vector is equal to the number of elements of the second competence feature vector.

[0151] The first competence feature vector and the second competence feature vector form a first detection feature map.

[0152] The length of the first detection feature map is 2, and the width is the number of elements of the first competence feature vector.

[0153] In the wide direction, a 2*2 two-dimensional convolution kernel is used to convolve the first detection feature map to obtain the monitor competence.

[0154] The competence discrimination network is trained using historical labeled competences of civil aviation monitors.

[0155] The competence is predicted by finding the relationship between the competence represented by the monitoring adjective corresponding to the monitoring phrase and the changes between the corresponding multiple monitoring phrases.

[0156] Optionally, the adjusted phrase matrix based on the n monitoring phrases includes:

[0157] According to the monitor behavior data, the order of the appearance of the n monitoring phrases is labeled to obtain an order matrix. The columns of the order matrix represent the n monitoring phrases, and the rows represent the time points. The values in the order matrix represent the order of the appearance of the monitoring phrase at the time point.

[0158] The monitoring phrases with the same order in the order matrix are rearranged to obtain an adjusted phrase matrix. The columns of the adjusted phrase matrix represent the order of the appearance of the monitoring phrases, and the rows represent the time points. The values in the adjusted phrase matrix represent the monitoring phrases corresponding to different orders of appearance at the time point.

[0159] Optionally, the n monitoring phrases are obtained by detecting the phrases related to the competence according to the monitor behavior data at the m time points, including:

[0160] obtaining a supervision database; the supervision database stores a plurality of phrases related to competency;

[0161] detecting the phrases in the supervision database in the inspector behavior data at the m time points as supervision phrases.

[0162] Embodiment 2

[0163] Based on the above-mentioned big data-based analysis method for civil aviation inspector competency, the embodiment of the present application further provides a big data-based analysis system for civil aviation inspector competency, which comprises an obtaining module, a phrase module, an association module, a graph structure module and a competency judgment module.

[0164] The obtaining module is used for obtaining inspector behavior data at m time points; the inspector behavior data represents sentences describing the behavior of civil aviation inspectors;

[0165] The phrase module is used for detecting phrases related to competency according to the inspector behavior data at the m time points, to obtain n supervision phrases.

[0166] The association module is used for obtaining n*(n-1) supervision phrase association probability values based on the n supervision phrases through an association phrase network; the supervision phrase association probability value represents the probability that one supervision phrase appears in the inspector behavior data and another supervision phrase also appears.

[0167] The association phrase network comprises a first association neural network, a first association convolution network and a fusion neural network.

[0168] The graph structure module is used for constructing a graph structure based on the inspector behavior data and the n*(n-1) supervision phrase association probability values, to obtain a supervision connection matrix; the supervision connection matrix represents a graph structure with the n supervision phrases as vertices and the supervision phrase association probability values as edges.

[0169] The competency judgment module is used for obtaining inspector competency based on the supervision connection matrix through a competency judgment network.

[0170] The algorithms and displays presented herein are not inherently related to any particular computer, virtual system, or other apparatus. Various general purpose systems can be used with programs in accordance with the teachings herein, or it can prove convenient to construct more specialized apparatus to perform the required method steps. The required structure for a variety of these systems will be apparent from the description above. In addition, the present application is not intended to be limited to any particular programming language. It will be appreciated that there are many programming languages that can be used to implement the teachings herein, and any such programming language can be used in connection with the teachings herein.

[0171] In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the application can be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

Claims

1. A big data-based analysis method for civil aviation inspector competence, characterized in that, The method comprises the following steps: obtaining monitor behavior data at m time points; the monitor behavior data represents a statement describing the behavior of a civil aviation monitor; detecting competency-related phrases based on the monitor behavior data at m time points to obtain n monitor phrases; the monitor phrases are nouns; obtaining n*(n-1) monitor phrase association probability values based on the n monitor phrases through an association phrase network; the monitor phrase association probability value represents the probability that one monitor phrase appears in the monitor behavior data and another monitor phrase also appears; the association phrase network comprises a first association neural network, a first association convolutional network and a fusion neural network; constructing a graph structure based on the monitor behavior data and the n*(n-1) monitor phrase association probability values to obtain a monitor connection matrix; the monitor connection matrix represents a graph structure with the n monitor phrases as vertices and the monitor phrase association probability values as edges; obtaining the competency of the monitor based on the monitor connection matrix through a competency discrimination network; obtaining the n*(n-1) monitor phrase association probability values based on the n monitor phrases through the association phrase network comprises: obtaining an adjustment phrase matrix based on the n monitor phrases; the adjustment phrase matrix represents the order in which the monitor phrases appear in the monitor behavior data; segmenting the adjustment phrase matrix by column to obtain k sequential phrase vectors; one sequential phrase vector corresponds to one order of appearance in the monitor behavior data; inputting the k sequential phrase vectors into the first association neural network to obtain a first feature vector; the first feature vector represents a feature vector of the probability of the appearance of one monitor phrase after a plurality of monitor phrases in the monitor behavior data; obtaining a second feature map based on the adjustment phrase matrix through the first association convolutional network; the second feature map represents a feature map of the probability of the appearance of one monitor phrase after a plurality of monitor phrases in the monitor behavior data; inputting the first feature vector and the second feature map into the fusion neural network to obtain the n*(n-1) monitor phrase association probability values; all neurons corresponding to one sequential phrase vector in the input layer of the first association neural network are connected to one neuron of the first layer of the hidden layer; all neurons of the first layer of the hidden layer in the first association neural network are connected to the first neuron of the second layer of the hidden layer; neurons other than the first [1, i-1] neurons in the first layer of the hidden layer of the first association neural network are connected to the i-th neuron of the second layer of the hidden layer; i ranges from 2 to k; obtaining the competency of the monitor based on the monitor connection matrix through the competency discrimination network comprises: the competency discrimination network comprises a first discrimination network, a first competency neural network, a convolutional neural network composed of a two-dimensional convolution of 2*n, and a convolutional neural network composed of a two-dimensional convolution kernel of 2*2; inputting the monitor connection matrix into the first discrimination network to detect the competency corresponding to the plurality of monitor phrases in the association relationship to obtain a first competency feature vector; convolve the monitor connection matrix and the two-dimensional convolution kernel of 2*n in the column direction with a step size of 1 to obtain a first association feature; In the inspector behavior data, find the corresponding adjective of the supervision phrase as the supervision adjective; Match the supervision adjective with the adjectives stored in the supervision database to obtain a supervision degree value; the supervision degree value represents the degree of supervision phrase pair in competence judgment; According to the arrangement order of the vertices in the connection matrix, the supervision degree value corresponding to the supervision phrase pair is formed into a supervision phrase degree set, and the supervision phrase degree set is input into the first competence neural network to obtain a second competence feature vector; The first competence feature vector and the second competence feature vector form a first detection feature map; In the wide direction, a 2*2 two-dimensional convolution kernel is used to convolve the first detection feature map to obtain the inspector competence.

2. The big data-based analysis method for civil aviation supervisor competence according to claim 1, characterized in that, The second feature map is obtained based on the adjustment phrase matrix through the first association convolution network, which includes: Obtain a multi-scale convolution kernel; the multi-scale convolution kernel includes a two-dimensional convolution kernel with a size of 2*2 and a size of a prime number; Using the multi-scale convolution kernel, the adjustment phrase matrix is convolved with a step size of 1 respectively to detect the association relationship between supervision phrases in different orders to obtain a multi-scale feature map; The multi-scale feature map is down-sampled by a 2*2 convolution kernel for feature fusion to obtain a second feature map.

3. The big data-based analysis method for civil aviation monitor's competence according to claim 1, characterized in that, The training method of the association phrase network includes: Obtain a training adjustment phrase matrix; the training adjustment phrase matrix represents the adjustment phrase matrix of a plurality of historical civil aviation inspectors; Obtain labeled association data; the labeled association data represents the probability that one supervision phrase appears in the inspector behavior data and another supervision phrase also appears in the statistical historical plurality of civil aviation inspectors; Input the first training feature vector into the first association neural network to obtain a first training feature vector; The first training feature vector is used as a first training prediction vector; the first training prediction vector represents the probability that a plurality of supervision phrases predicted by the first association neural network appear after one supervision phrase; Calculate the loss between the labeled association data and the first training prediction vector to train the first association neural network; Input the training adjustment phrase matrix into the first association convolution network to obtain a second training feature map; Input the second training feature map into the second full connection network to obtain a second training prediction vector; the second training prediction vector represents the probability that a plurality of supervision phrases predicted by the first association convolution network appear after one supervision phrase; Calculate the loss between the labeled association data and the second training prediction vector to train the first association convolution network; Input the first training feature vector and the second training feature map into the fusion neural network to obtain a training supervision phrase association probability vector; Calculate the loss between the labeled association data and the training supervision phrase association probability vector to train the fusion neural network, the first association convolution network and the first association neural network.

4. The big data-based analysis method for civil aviation supervisor competence according to claim 1, characterized in that, Based on the inspector behavior data and n*(n-1) supervision phrase association probability values, a graph structure is constructed to obtain a supervision connection matrix, which includes: The supervision phrase of the inspector behavior data is used as a vertex, and the supervision phrase association probability value is used as an edge to obtain a supervision phrase graph; The vertices are used as rows and columns respectively to construct a supervision connection matrix; If the supervision word group association probability is greater than the association threshold, the supervision word group association probability is retained; if the supervision word group association probability is less than or equal to the association threshold, the supervision word group association probability is set to 0. The supervision word group association probability value is filled in the supervision connection matrix according to the supervision word groups corresponding to the rows and columns of the supervision connection matrix.

5. The big data-based analysis method for civil aviation monitor's competence according to claim 1, characterized in that, The adjustment word group matrix is obtained based on the n supervision word groups, and includes: According to the supervision data of the inspectors, the order of the appearance of the n supervision word groups is marked to obtain an order matrix; the rows of the order matrix represent the n supervision word groups, and the columns represent the time points; the values in the order matrix represent the order of the appearance of the supervision word groups at the time points in the supervision data of the inspectors; The supervision word groups with the same order in the order matrix are rearranged to obtain an adjustment word group matrix; the rows of the adjustment word group matrix represent the order of the appearance of the supervision word groups, and the columns represent the time points; the values in the adjustment word group matrix represent the supervision word groups corresponding to different orders of appearance at the time points.

6. The big data-based analysis method for civil aviation monitor competence according to claim 1, characterized in that, The n supervision word groups are obtained by detecting the word groups related to the competency according to the supervision data of the inspectors at the m time points, and include: The supervision database is obtained; the supervision database stores a plurality of word groups related to the competency; The word groups in the supervision data of the inspectors at the m time points that are in the supervision database are detected as supervision word groups.

7. A big data-based civil aviation monitor competence analysis system, characterized in that, It includes: The supervision data of the inspectors at the m time points is obtained by the acquisition module; The supervision data of the inspectors represents the sentences describing the behavior of the civil aviation inspectors; The word group module is used to detect the word groups related to the competency according to the supervision data of the inspectors at the m time points, and obtain n supervision word groups; the supervision word groups are nouns; The association module is used to obtain n*(n-1) supervision word group association probability values based on the n supervision word groups through the association word group network; the supervision word group association probability value represents the probability that one supervision word group appears in the supervision data of the inspectors and another supervision word group also appears; The association word group network includes a first association neural network, a first association convolutional network, and a fusion neural network; The graph structure module is used to construct a graph structure based on the supervision data of the inspectors and the n*(n-1) supervision word group association probability values, and obtain a supervision connection matrix; the supervision connection matrix represents a graph structure with the n supervision word groups as vertices and the supervision word group association probability values as edges; The competency discrimination module is used to obtain the competency of the inspectors based on the supervision connection matrix through a competency discrimination network. The n*(n-1) supervision word group association probability values are obtained based on the n supervision word groups through the association word group network, and include: The adjustment word group matrix is obtained based on the n supervision word groups; the adjustment word group matrix represents the order of the appearance of the supervision word groups in the supervision data of the inspectors; The adjustment word group matrix is divided into k order word group vectors in columns; one order word group vector corresponds to one order of appearance in the supervision data of the inspectors; The k order word group vectors are input into the first association neural network to obtain first feature vectors; the first feature vectors represent the feature vectors of the probabilities of the appearance of the supervision word groups behind one supervision word group in the supervision data of the inspectors; The first association convolutional network is used to obtain a second feature map based on the adjusted phrase matrix, and the second feature map represents a feature map of probabilities of occurrence of a plurality of monitoring phrases after a monitoring phrase in the monitoring data of the supervisor; The first feature vector and the second feature map are input into the fusion neural network to obtain n*(n-1) monitoring phrase association probability values; All neurons corresponding to a sequential phrase vector in the input layer of the first association neural network are connected to a neuron in the first layer of the hidden layer; All neurons in the first layer of the hidden layer of the first association neural network are connected to the first neuron in the second layer of the hidden layer; In the first layer of the hidden layer of the first association neural network, neurons other than the first [1, i-1] neurons are connected to the i-th neuron in the second layer of the hidden layer; i is in the range of [2, k]; The competency discrimination network is used to obtain the competency of the supervisor based on the monitoring connection matrix, and the competency includes: The competency discrimination network includes a first discrimination network, a first competency neural network, a convolutional neural network composed of a 2*n two-dimensional convolution, and a convolutional neural network composed of a 2*2 two-dimensional convolution kernel; The monitoring connection matrix is input into the first discrimination network to detect the competency corresponding to a plurality of monitoring phrases in the association relationship, and a first competency feature vector is obtained; In the column direction, the monitoring connection matrix is convolved with the 2*n two-dimensional convolution kernel at a step of 1 to obtain a first association feature; In the monitoring data of the supervisor, a descriptive word corresponding to the monitoring phrase is found as a monitoring descriptive word; The monitoring descriptive word is matched with the descriptive words stored in the monitoring database to obtain a monitoring degree value, and the monitoring degree value represents the degree of monitoring phrase discrimination of the competency; According to the arrangement order of the vertices in the connection matrix, the monitoring degree values corresponding to the monitoring phrases are formed into a monitoring phrase degree set, and the monitoring phrase degree set is input into the first competency neural network to obtain a second competency feature vector; The first competency feature vector and the second competency feature vector are formed into a first detection feature map; In the wide direction, a 2*2 two-dimensional convolution kernel is used to convolve the first detection feature map to obtain the competency of the supervisor.

Images