A BLS-based NAVTEX message semantic automatic classification method

Abstract

The application discloses a kind of based on BLS's NAVTEX message semantic automatic classification method, to NAVTEX message data application data cleaning and word segmentation operation, and proposed semantic label to give message label category, specifically as the summary of the message content contained in navigation warning, defined 7 kinds of semantic labels.By the application, redundant and complex original message data can be effectively simplified and the foundation for subsequent classification tasks is laid, making the trained features more representative.The application extracts features using the vector space model on the preprocessed NAVTEX message data, which has the advantages of easy calculation and simple structure.Then, based on BLS, the semantic classification is performed, which has the advantages of short training time and good classification effect, effectively alleviating the labor intensity of the crew and improving the emergency response speed and efficiency of responding to safety information, thereby improving the safety navigation performance of the ship.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of natural language processing and artificial intelligence, and particularly relates to a NAVTEX message semantic automatic classification method based on BLS (Breadth Learning System). BACKGROUND

[0002] In order to solve a series of problems such as effectively implementing search and rescue tasks, protecting the safety of life and property at sea, and greatly ensuring global safe transportation and production development at sea, a global maritime distress and safety system is installed on a ship during sea transportation. The NAVTEX system in the system broadcasts messages containing a large amount of navigation warnings, weather warnings, weather forecasts and other text information related to emergency information for navigation safety. Therefore, it is particularly important to fully and effectively use the NAVTEX message during ship navigation. However, the current NAVTEX message is difficult to be efficiently applied to the actual scene of ship driving management. This is because the process of ship intelligent driving is slightly slower than the development of mobile intelligent navigation technology and other technologies that have emerged with the development of informationization and intellectualization. Even if the performance of part of the navigation aid equipment has been improved, it still cannot meet the needs of the mariners. For the NAVTEX system, after the crew receives a new NAVTEX message, it must be manually read and classified. This not only involves the problem of low work efficiency, but also the probability of error in the process of manual participation is greater than that in the automatic process.

[0003] In the existing research on the NAVTEX system, the focus is on improving the performance of the system from the communication technology, and the research on the semantic classification of the message content is not mature enough, and there is a lot of room for expansion. Although some researches have proposed that more attention should be paid to the processing of NAVTEX messages, but in the message data processing part, only the initial stop word processing is done, and no further word stemming processing such as word segmentation process is performed, and in the classification model construction part, only the traditional machine learning model is applied for classification. These situations may affect the reliability of the message semantic classification. At the same time, with the rapid development of artificial intelligence technology, more and more machine learning methods are applied to the field of shipping. Therefore, a method is needed to classify the semantic of the NAVTEX message and provide it to the mariners, so as to ensure that the maritime safety information they can obtain is more accurate, thereby improving the safety performance of ship navigation. SUMMARY

[0004] In order to solve the above problems existing in the prior art, the present application proposes a BLS-based NAVTEX message semantic automatic classification method, which applies relevant methods of machine learning to the shipping field, can greatly improve the accuracy of NAVTEX message semantic automatic classification, can effectively alleviate the labor intensity of the crew and improve the emergency response speed and efficiency of responding to maritime safety information, so as to improve the safe navigation performance of the ship.

[0005] In order to achieve the above-mentioned purpose, the technical means adopted by the present application are as follows: a BLS-based NAVTEX message semantic automatic classification method, wherein the BLS is the abbreviation of the width learning system, and the method comprises the following steps:

[0006] S1, preprocessing the NAVTEX message data

[0007] The NAVTEX message data is the relevant maritime safety information data received in the ship navigation process, and the NAVTEX message data preprocessing includes data cleaning, word segmentation processing and category labeling, and the specific steps are as follows:

[0008] S11, data cleaning

[0009] In the data cleaning stage, the invalid message in the message data is filtered out. And the non-text part in the message is filtered out by the method of regular expression, to obtain a message data set D={D1, D2,..., D i ,...,D m}, wherein D i represents the i-th message in the set, and m is the number of message samples.

[0010] S12, word segmentation processing

[0011] The word segmentation processing includes space splitting, stop word removal, punctuation, and stem extraction of the message D i . The message D i is split by the space between words; the content is simplified by removing stop words, function words and punctuation marks in the message D i ; the message D i is processed to extract the stem, to obtain a set of message word segmentation D'={D'1, D'2,..., D' i ,...,D' m}, wherein D' i represents the word segmentation set of the i-th message.

[0012] S13, category labeling

[0013] According to the analysis of the content described in the NAVTEX message, the semantic label of artificial classification annotation is set to 7 semantic categories, which are: maritime operation, navigation obstacle, navigation mark change, military exercise, piracy, cancellation of previous warning, and other information. Each message word set D' i corresponds to only one type of label, and the label matrix is denoted as L={L1, L2,..., L i ,...,L m}, wherein L belongs to an m×1 dimensional vector space, and L i is the i-th label in the 7 semantic labels of the message.

[0014] S2, feature extraction of message words

[0015] According to the message word set D', the method of vector space model is used for feature extraction, and a multi-dimensional vector space is constructed, and the steps are as follows:

[0016] S21, using feature term-weight matrix to represent the whole message data set

[0017] A set of message word sets D' is represented by a message feature matrix X, wherein the i-th message word set D' i is represented by a vector as follows:

[0018] X i ={(x i1 ,w i1 ),(x i2 ,w i2 ),…,(x ij ,w ij ),…,(x i ,w in )}

[0019] X i is regarded as an n-dimensional word vector composed of binary features. Wherein, x ij is the j-th feature item of the i-th message feature vector of the message feature vector X i ; w ij is the weight corresponding to the feature item x ij , which represents the importance of the feature item in the message; n represents the feature dimension, any dimension represents a feature item in the message feature vector X i , and n is the total number of remaining words after removing the repeated words in the word set D'.

[0020] S22, solving the weight corresponding to the message feature item

[0021] When the feature dimension in the collected message data is determined, the feature vector X iSimplify as:

[0022] X i = {w i1 ,w i2 ,…,w ij ,…,w in}

[0023] X' i The message feature item weight vector of the i-th message is called. The weight w ij is calculated using the following formula:

[0024]

[0025] wherein, TF(x ij ,X i ) represents the frequency of feature item x ij in the message feature vector X i , and the calculation formula is as follows:

[0026]

[0027] is the number of times that feature item x ij appears in the message feature vector X i , and is the total number of times that all feature items in X i appear. And IDF(x ij ) represents the inverse document frequency of feature item x ij , and the calculation formula is as follows:

[0028]

[0029] wherein, is the number of message feature vectors in which the feature item x ij appears in the message feature matrix X. The calculation formula of the normalization factor E is as follows:

[0030]

[0031] S3, sorting and encoding the message semantic label

[0032] Introducing One-Hot encoding to convert the category variable into a form convenient for the classification model to use, and the steps are as follows:

[0033] S31, integer encoding of the message semantic label

[0034] The 7 message semantic labels are sorted and integer encoded, and the integers 0 to 6 are assigned to the 7 semantic categories in turn. That is, the semantic label matrix L is mapped to the integer label matrix L' = {L'1, L'2,..., L'i ..., L' m}, wherein L' i is an integer representing the corresponding semantic category.

[0035] S32, One-Hot encoding of integer labels

[0036] A binary vector is created to represent each integer value. According to the common 7 integer values, the length of the binary vector is determined to be 7; the label encoded as 0 is represented by a binary vector [1, 0, 0, 0, 0, 0, 0], where the 0th index is marked as 1 and the rest are 0. The other 6 integer encodings are the same, except that the integer corresponds to the position of the index marked as 1, and the rest are 0. Thus, the One-Hot label matrix is obtained:

[0037]

[0038] wherein and only one position in the One-Hot label vector is 1 and the rest are 0.

[0039] S4, establishing a BLS-based classification model

[0040] The message feature item weight matrix obtained in step S2 is used as the input of the classification model, thereby realizing the establishment of the message semantic classification model. The network structure of BLS includes four parts: input sample layer, feature mapping layer, enhanced node layer and output layer. The algorithm process of each layer is as follows:

[0041] S41, listing the input matrix of the classification model

[0042] The input matrix X' of the message semantic classification model is as follows:

[0043]

[0044] wherein, m is the number of message samples, n is the feature dimension, w ij is the weight of the feature item x ij in X i .

[0045] S42, feature mapping of the input message feature item weight matrix

[0046] In the input sample layer, the message feature item weight matrix X' is input into the classification model, and the BLS network will perform feature mapping on X'. Through the feature mapping function φ g(*) is mapped to t groups of mapping nodes of different features, and k nodes are generated in each group of mapping, and the gth group of message feature mapping nodes Z g The calculation formula is as follows:

[0047]

[0048] Wherein, 1≤g≤t, Z g is mapped to Z t =(Z1,Z2,…,Z t ) and The mapped feature weight feature is regarded as a layer, that is, a feature mapping layer. Wherein, φg(*) is a linear mapping function, is a randomly generated weight matrix, and its corresponding bias term.

[0049] S43, generating message enhancement nodes

[0050] Connect the t groups of message feature mapping nodes Z g obtained in step S42 with the enhancement node layer, randomly generate the weight coefficient matrix and the bias term Map Z s to p groups of randomly generated message enhancement nodes through a nonlinear function ξ t (*) and perform parallel connection to represent H p =(H1,H2,…,H p ) and That is, the enhancement node layer is obtained. Suppose that q nodes are generated in each group of mapping, and H s represents the s th group of message enhancement nodes, and the calculation formula is as follows:

[0051]

[0052] S44, solving the connection weight matrix

[0053] Parallelly connect Z t obtained in step S42 and H p obtained in step S43, and then the following is obtained:

[0054] A=(Z t |H p )

[0055] Directly connect A obtained by parallelly connecting the feature mapping layer and the enhancement node layer to the output end, which is called the output layer. Therefore, the output matrix is:

[0056] Y=AW

[0057] Wherein, W is the connection weight matrix of the output layer, and then Y=AW

[0058]

[0059] In the formula A -1 Describe the pseudo-inverse of matrix A by solving A -1 Calculation of the ridge regression approximate solution A -1 This leads to the connection weight matrix W, at which point the entire model training is complete. The calculation formula is as follows:

[0060]

[0061] Where I represents the identity matrix, A T Let A be the transpose of A.

[0062] S45. List the output matrix of the classification model.

[0063] The output matrix Y of the classification model is as follows:

[0064]

[0065] Among them, Y i Let y be the output probability vector obtained by the classification model for the i-th message sample. ij This represents the probability value belonging to the corresponding message semantic category. Take each output probability vector Y. i The maximum value y in ij , to the maximum value y ij The corresponding position in the probability vector is represented as 1, and the rest are represented as 0, thus obtaining the final output matrix Y', which is the final message semantic classification result, as shown below:

[0066]

[0067] Among them, Y' i Y' represents the output vector corresponding to the i-th message. i = [0 / 1,0 / 1,0 / 1,0 / 1,0 / 1,0 / 1,0 / 1,0 / 1,0 / 1], and only one position in the output vector is 1, while the rest are 0. Based on the One-Hot label matrix obtained in step S32... By comparing the results, the semantic category of the corresponding message is determined, and the semantic classification result of the i-th message is finally output.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] 1. The application applies data cleaning and word segmentation operations to NAVTEX message data, and proposes semantic labels to label the categories of the messages. Specifically, the message content contained in the navigation warning is summarized, and 7 types of semantic labels are defined. Through the application, redundant and complex original message data can be simplified and processed, and a foundation for subsequent classification tasks is laid, making the trained features more representative.

[0070] 2. The application extracts features from the preprocessed NAVTEX message data using the vector space model, which has the advantages of easy calculation and simple structure. Then, based on BLS, the semantic classification is performed, which has the advantages of short training time and good classification effect, and is very suitable for application in the navigation scene of the application with high real-time requirement, which can effectively alleviate the labor intensity of the crew and improve the emergency response speed and efficiency of safety information, thereby improving the safety navigation performance of the ship. BRIEF DESCRIPTION OF DRAWINGS

[0071] Figure 1 is the overall flowchart of the automatic classification of NAVTEX message semantics of the application.

[0072] Figure 2 is a schematic diagram of the feature extraction of the message using the vector space model of the application.

[0073] Figure 3 is a schematic diagram of the sorting and encoding of the message semantic labels of the application.

[0074] Figure 4 is a flowchart of the message semantic classification based on BLS of the application. DETAILED DESCRIPTION

[0075] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0076] As shown in Figure 1 , a BLS-based automatic classification method for NAVTEX message semantics has the following steps:

[0077] S1, pre-processing the NAVTEX message data. The message D is pre-processed through data cleaning, word segmentation and category labeling to obtain the word segmentation set D' and the category label L.

[0078] S2, feature extraction of message word segmentation. As shown inFigure 2 As shown, D' is obtained by the method of vector space model for word segmentation feature extraction to obtain the message feature item weight X'. First, use the message feature matrix X to represent D', X is simplified as X', and finally the weight of the feature item in X' is solved by the weight formula.

[0079] S3, sort and encode the message semantic label. As shown in Figure 3 , 7 types of message semantic label are sorted and integer coded, and the integer label matrix L' is obtained by L mapping, and the One-Hot label matrix is obtained by One-Hot coding L'.

[0080] S4, establish a classification model based on BLS. As shown in Figure 4 , input X' into the trained BLS semantic classification model, and the BLS network maps the feature through the feature mapping function g (*) to map X' to t groups of different feature mapping nodes. Connect the t groups of nodes to obtain Z t =(Z1,Z2,…,Z t ), and regard the mapped feature item weight feature as a layer, that is, the feature mapping layer. Connect Z t and the enhanced node layer, and randomly generate the weight coefficient matrix and the bias term through the nonlinear function s (*) to map Z t to p groups of message enhanced nodes with randomly generated weights, and perform parallel connection to obtain H p =(H1,H2,…,H p ), that is, the enhanced node layer. Perform parallel connection of Z t and H p to obtain A=(Z t |H p ), and directly connect A to the output end, which is called the output layer, and the output matrix Y=AW. Wherein, W is the connection weight matrix of the output layer, then W=A -1 Y. In the formula, A -1 represents the pseudo-inverse matrix of matrix A, A -1 is calculated by solving the ridge regression approximate solution of A -1 , and then W is obtained, and the model training is completed. The calculation formula is where I represents the unit matrix, and A T is the transpose matrix of A. Finally, in the output matrix Y, the maximum value y i in each output probability vector Y ij is represented as 1, and the remaining positions are represented as 0, so as to obtain the final output matrix Y', and Y' is compared with , that is, the semantic classification result of the message is obtained.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions described in the foregoing embodiments can still be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for automatic semantic classification of NAVTEX messages based on BLS, wherein BLS is short for Breadth Learning System, characterized in that: The method includes the following steps: S1. Preprocess NAVTEX message data NAVTEX message data refers to maritime safety information received by ships during navigation. NAVTEX message data preprocessing includes data cleaning, word segmentation, and category labeling. The specific steps are as follows: S11, Data Cleaning During the data cleaning phase, invalid messages are filtered out from the message data; and non-text portions of the messages are filtered out using regular expressions, resulting in a message dataset. ,in Represents the first element in the set. Message, Number of message samples; S12, Word segmentation processing Word segmentation processing includes processing the message. Perform space splitting, stop word removal, punctuation removal, and stemming; message Split words using spaces between words; remove messages Simplified content regarding stop words, function words, and punctuation marks in the message; Stemming is performed to obtain a set of message segmentation words. ,in Indicates the first The word segmentation set of the message; S13, Category Labeling Based on the semantic tags manually categorized and labeled from the content described in NAVTEX messages, seven semantic categories were established: maritime operations, navigational obstructions, changes in navigational aids, military exercises, piracy, cancellation of previous warnings, and other information; each message was segmented into a set of words. It corresponds to only one type of label, and the label matrix is ​​denoted as . ,in ,Right now L belong m This is the first of the 7 types of message semantic tags. Class tags; S2. Extract features from message segmentation. Based on the message segmentation set The vector space model is used to extract features and construct a multi-dimensional vector space. The steps are as follows: S21. Represent the entire message dataset using a feature-weight matrix. A set of message segmentation sets Using message feature matrix To indicate, of which the first The word segmentation set of the message Represented as a vector: Will Viewed as being composed of binary feature sets dimensional word vectors; where, It is the message feature vector The The first message feature vector One feature term; It is a characteristic term The corresponding weight represents the importance of the feature in the message; Each dimension represents a feature vector of the message. One of the feature terms, and For word segmentation set The total number of words remaining after removing the number of repeated words from the total number of words in the dictionary; S22. Solve for the weights corresponding to the message feature terms. Once the feature dimensions in the collected message data are determined, the feature vector... Simplified to: Called the first The message feature weight vector of each message; where the weights are... Use the following formula to perform the calculation: in, Representative characteristic terms In message feature vector The frequency of occurrence is calculated using the following formula: For characteristic terms In message feature vector The number of times it appears in for The total number of occurrences of all feature terms; and Representative characteristic terms The inverse document frequency is calculated using the following formula: in, message feature matrix Feature items appear in Number of message feature vectors; normalization factor The calculation formula is as follows: S3. Sort and encode the message semantic tags. One-Hot encoding is introduced to transform categorical variables into a form that is easier for classification models to use. The steps are as follows: S31. Encode the message semantic tags with integers. The semantic tags of seven message categories are ordered and encoded with integers, with integers 0 to 6 assigned sequentially to the seven semantic categories; that is, the semantic tag matrix. Mapping to an integer label matrix ,in , An integer representing the corresponding semantic category; S32. Perform One-Hot encoding on the integer labels. Create a binary vector to represent each integer value; based on the fact that there are 7 types of integer values, the length of the binary vector is determined to be 7; the label encoding the integer as 0 is represented by the binary vector. This means that the 0th index is marked as 1, and the rest are 0; the other 6 types of integer encoding are similar, except that the index of the integer is marked as 1, and the rest are 0, thus obtaining the One-Hot label matrix: in , Furthermore, the One-Hot label vector has only one position that is 1, while all other positions are 0; S4. Establish a BLS-based classification model. The message feature weight matrix obtained in step S2 is used as the input to the classification model, thereby establishing a message semantic classification model. The BLS network structure consists of four parts: input sample layer, feature mapping layer, enhancement node layer, and output layer. The algorithm flow for each layer of BLS for message semantic classification is as follows: S41. List the input matrix of the classification model. Input matrix of message semantic classification model as follows: in, , Number of message samples For feature dimension, For characteristic terms exist Weights in; S42. Perform feature mapping on the input message feature term weight matrix. In the input sample layer, the message feature term weight matrix is... Inputting the classification model, the BLS network will... Perform feature mapping, through the feature mapping function. Weight matrix of message feature terms Mapped to Groups of mapping nodes with different characteristics, assuming each group of mappings generates The node, the Group message feature mapping node The calculation formula is as follows: in, ,Will Performing a parallel operation indicates obtaining... and The feature weights of the mapped features are considered as a single layer, i.e., the feature mapping layer; where, It is a linear mapping function. To randomly generate the weight matrix, Its corresponding bias term; S43, Generate message enhancement node The result obtained in step S42 Group message feature mapping node Connect to the enhancement node layer and randomly generate a weight coefficient matrix. and bias terms Through nonlinear functions Will Mapped to Grouping message enhancement nodes with randomly generated weights and connecting them in parallel is represented as follows: and This results in an enhanced node layer; assuming each mapping generates... 1 node Representing the The calculation formula for group message enhancement nodes is as follows: S44. Solve for the connection weight matrix. The results obtained from step S42 are respectively and obtained from step S43 By performing a parallel representation, we obtain: The feature mapping layer and the enhancement node layer are connected in parallel to obtain The layer directly connected to the output is called the output layer; therefore, the output matrix is: in, Let the connection weight matrix of the output layer be: In the formula Representation matrix The pseudo-inverse matrix is ​​obtained by solving... Calculation of ridge regression approximate solution This leads to the connection weight matrix. The model training is now complete. The calculation formula is as follows: in Represents the identity matrix. for The transpose of the matrix; S45. List the output matrix of the classification model. The output matrix of the classification model as follows: in, For the first The corresponding output probability vector obtained by the classification model for each message sample. This represents the probability value belonging to the corresponding message semantic category; each output probability vector is then used. The maximum value in , to the maximum value The corresponding position in the probability vector is represented as 1, and the rest are represented as 0, thus obtaining the final output matrix. The final message semantic classification result is represented as follows: in, Representing the The output vector corresponding to each message. Furthermore, the output vector contains only one position with a value of 1, while all other positions are 0; based on the One-Hot label matrix obtained in step S32... By comparing the results, the semantic category of the corresponding message is determined, and the final output is the first... The semantic classification results of the message.

Images