BLS-based intelligent decision support method for navtex message context

Abstract

The application discloses a BLS-based NAVTEX message context intelligent decision support method, applies data cleaning, word type reduction and reference stop word operations to NAVTEX message data, and proposes a message context label to mark the categories of the message, which specifically shows that the message content contained in the navigation warning is analyzed, and six categories of message context labels are summarized.Through the application, effective information in the original message data can be accurately extracted, and strong support is provided for subsequent message context classification tasks and navigation decision support, so that the trained message features are more concise and representative.The application is based on BLS and is improved, the accuracy of the NAVTEX message context classification is improved, intelligent navigation decision support services are provided for the mariners, the working strength can be effectively reduced and the working efficiency can be improved, meanwhile, the safety navigation performance of the ship is further enhanced, and the function of the navigation aid equipment is enriched.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of natural language processing and artificial intelligence, and particularly relates to a BLS-based NAVTEX message context intelligent decision support method. BACKGROUND

[0002] During the process of marine transportation, most ships are equipped with the global maritime distress and safety system. The NAVTEX system contained in the system involves a large number of navigation warnings, weather warnings, weather forecasts and other text information related to navigation safety emergency information. Therefore, it is crucial to accurately master and use the NAVTEX message for ship navigation decision support. However, the NAVTEX message has not yet been fully integrated into the ship navigation auxiliary system, and has not played an effective role in ship driving management, decision support and other practical scenarios. This is because the related technology of ship intelligent driving is still in the theoretical research stage and has not been implemented in practical applications. At the same time, the shipping industry is relatively limited in its own development compared to other industries, and is still relatively backward in terms of technology and management, making it still highly dependent on seafarers. In the process of using the NAVTEX system, after receiving the NAVTEX message, the seafarers must manually read and analyze it before making corresponding navigation decisions. This not only increases the work intensity of seafarers, but also affects their work efficiency. In addition, the accuracy of manual information reading is lower than that of automation, which further affects the safety of ship navigation.

[0003] In view of the above problems, the current research on intelligent classification of NAVTEX message content and providing context decision support services for seafarers still has a lot of development space. In the current research on processing NAVTEX messages, only preliminary stop words are removed in the preprocessing of message data, without further processing such as word type restoration. The text classification model part only applies traditional machine learning models, and only classifies the message content without further research on context intelligent decision support. Therefore, a method is needed to classify the context of NAVTEX messages and provide decision support to seafarers, improve work efficiency and the accuracy of message context classification, and further enhance the safety 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 context intelligent decision support method. The application of BLS to the field of shipping can greatly improve the accuracy of NAVTEX message context classification and provide intelligent navigation decision support services for seafarers, effectively reducing the work intensity and ensuring the safe navigation of ships.

[0005] In order to achieve the above object, the technical means adopted by the present application are as follows: the BLS-based NAVTEX message context intelligent decision support 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 during the ship navigation process, and the NAVTEX message data preprocessing includes four parts of data cleaning, word type restoration, introduction of stop words and category labeling, and the operation process is as follows:

[0008] S11, data cleaning

[0009] The original message is subjected to data cleaning to eliminate invalid messages in the message data. And through the method of regular expression, the non-text part in the message is filtered out to obtain the message data set D={D1, D2,..., D i ,...,D m}, wherein D i represents the ith message in the set, and m is the number of message samples.

[0010] S12, word type restoration

[0011] The stem part of the word in the ith message D i is extracted, and the affix is removed to restore the word to the original form.

[0012] S13, introduction of stop words

[0013] The message words are split by using the space between them; by introducing stop words, some function words, phrases and punctuation marks in the message are removed to reduce the content that does not need to be analyzed in the word library, thereby obtaining a set of message feature word sets D'={D'1, D'2,..., D' i ,...,D' m}, wherein D' i represents the feature word set of the ith message.

[0014] S14, category labeling

[0015] According to the corresponding action to be taken after analyzing the description content in the NAVTEX message, the context label is classified and labeled, and is set to 6 context categories, which are prohibition of entry, vigilance, give way, keep careful lookout, check the planned route, and other behaviors. Among them, each message feature word set corresponds to at least one label, and the label matrix is recorded as L={L1, L2,..., L i ,...,L m}, wherein L is a vector space of mxc dimensions, and L iIt is a class c label among the 6 types of message context labels, and 1≤c≤6.

[0016] S2. Weighted processing of message feature words

[0017] Based on the feature word set D', D' is weighted using vectorization and weight correction methods, as follows:

[0018] S21. Vectorize the feature word set of the message.

[0019] The feature set D' of the i-th message i Using message feature vector X i It means, i.e., X i ={(x i1 ,w i1 ),(x i2 ,w i2 ),…,(x ij ,w ij ),…,(x in ,w in )}, X i It can be viewed as an n-dimensional feature vector composed of binary feature sets. Where x ij w represents the j-th feature term of the feature vector of the i-th message; ij It is the characteristic term x ij The corresponding weights represent the importance of the feature terms in the message; n represents the feature dimension, where each dimension represents the message feature vector X. i Let n be a feature term in the feature word set D', and n be the total number of feature words remaining after removing the number of duplicate feature words from the total number of feature words in the feature word set D'.

[0020] S22. Adjust the weights of message feature items.

[0021] Once the feature dimensions in the message data are determined, the message feature vector X will be... i Simplified to:

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

[0023] This represents the feature weight vector of the i-th message. Correcting the message feature weights involves the following steps:

[0024] S221. Solving for the term frequency of message feature terms.

[0025] TF(x) of message features ij ,X i ) represents the characteristic term xij The frequency of the feature vector X i in the message, the calculation formula is as follows:

[0026]

[0027] Wherein, The feature item x ij The number of times of occurrence in the message feature vector X i The total number of times of occurrence of all feature items in X i

[0028] S222, solve the inverse document frequency of message feature item

[0029] The inverse document frequency of message feature item IDF(x ij ) represents the frequency of feature item x ij in all message feature vectors, and the calculation formula is as follows:

[0030]

[0031] Wherein, X i The number of message feature vectors in which the feature x ij appears.

[0032] S223, solve the weight of message feature item

[0033] The result obtained by multiplying the results obtained in steps S2213 and S222 is the weight of feature item x ij in X i , and the calculation formula is as follows:

[0034]

[0035] Wherein, the calculation formula of the normalization factor E is as follows:

[0036]

[0037] S3, sort and encode the message context label

[0038] Introduce One-Hot encoding to convert category variables into a form convenient for classification model, the steps are as follows:

[0039] S31, integer encoding of message context label

[0040] Sort and integer encode 6 message context labels, respectively assign integer 0 to 5 to the 6 context categories described in step S14. That is, map the context label matrix L to the integer label matrix L' = {L'1, L'2,..., L'​​i ..., L' m}, wherein L' i is an integer array representing the corresponding context category, and 1≤c≤6.

[0041] S32, One-Hot encoding is performed on the integer label

[0042] Each integer array is represented as a binary vector. According to the six types of integer values, the length of the binary vector is determined to be 6; the index of each integer in each integer array corresponding to the ordered position is marked as 1, and the remaining positions are marked as 0, thereby obtaining a One-Hot label matrix:

[0043]

[0044] wherein and at least one position in the One-Hot label vector is 1, and the remaining positions are all 0.

[0045] S4, establishing a context intelligent decision support model

[0046] The network structure of the BLS includes four parts: an input sample layer, a feature mapping layer, an enhanced node layer, and a solution output layer weight. The output layer of the BLS is improved to process the message context intelligent decision support problem, and the training process of the context intelligent decision support model is as follows:

[0047] S41, taking the message feature item weight matrix as input

[0048] In the input sample layer, the message feature item weight matrix is input to the BLS network, and the input matrix X' is represented as follows:

[0049]

[0050] 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 the message feature vector X i .

[0051] S42, performing feature mapping on the input message feature item weight matrix

[0052] The BLS network performs feature mapping on the input message feature item weight matrix X', and maps X' into t groups of message feature mapping nodes through a feature mapping function φ g (*), wherein each group of mapping generates k nodes, and the calculation formula of the gth group of message feature mapping nodes Z g is as follows:

[0053]

[0054] where 1≤g≤t, Z g is mapped and concatenated as Z t = (Z1, Z2, …, Z t ) and the mapped message features are regarded as a layer, i.e., the feature mapping layer. Wherein, φ g (*) is a linear mapping function, is a randomly generated weight matrix, is its corresponding bias term.

[0055] S43, generating message enhancement nodes

[0056] The t groups of message feature mapping nodes Z g obtained in step S42 are connected with the enhancement node layer, and a weight coefficient matrix and a bias term are randomly generated. s (*) is a nonlinear function, Z t is mapped into p groups of message enhancement nodes with randomly generated weights, i.e., the enhancement node layer is obtained. Let q nodes be generated in each group, and H s represents the s-th group of message enhancement nodes, and the calculation formula is as follows:

[0057]

[0058] H s is concatenated to obtain H p = (H1, H2, …, H p ), wherein

[0059] S44, solving the connection weight matrix

[0060] Z t obtained in step S42 and H p obtained in step S43 are concatenated as:

[0061] A = (Z t | H p )

[0062] A obtained by concatenating the feature mapping layer and the enhancement node layer is directly connected to the output end, which is called the output layer. Therefore, the output probability matrix is:

[0063] Y = AW

[0064] wherein W is the connection weight matrix of the output layer, then:

[0065] W = A -1 Y

[0066] In the formula A -1 Describe the pseudo-inverse of matrix A by solving A -1 The ridge regression approximation solution is used to calculate A. -1 This leads to the connection weight matrix W, at which point the model training is complete. The calculation formula is as follows:

[0067]

[0068] Where I represents the identity matrix, A T Let A be the transpose of A.

[0069] S45, Multi-label Output

[0070] The probability matrix Y output by BLS is calculated as follows:

[0071]

[0072] in, 6 represents the number of message context categories, Y i Let y be the output probability vector obtained after the i-th message sample is processed by the classification model. ij These are the probability values ​​for the corresponding message context categories. In the BLS output layer, for each Y... i y in ij The homogenization process is performed using the following formula:

[0073]

[0074] To make its output a uniform matrix Y', it is represented as follows:

[0075]

[0076] At this point, each output uniformized vector Y' i There is y' i1 +y' i2 +…+y' i6 =1. Then, by setting a decision function... To apply y' to each normalized vector ij The filtering formula is as follows:

[0077]

[0078] Where θ is the threshold, when y' ij When it is greater than the set θ, then f(y') ij ) = 1 When y' ij When it is less than the set θ, then f(y′) ij ) = 0 The final multi-label output matrix of BLS is represented as follows:

[0079]

[0080] wherein, represents the output vector corresponding to the i-th message, and at least one position in the output vector is 1, and the rest are 0. According to the One-Hot label matrix obtained in step S32 comparison and determination of the message context category are performed, so as to obtain the message context classification result.

[0081] Compared with the prior art, the present application has the following beneficial effects:

[0082] 1. The present application applies data cleaning, word type restoration, and reference stop word operation to the NAVTEX message data, and proposes a message context label to label the category of the message, which specifically represents that the message content contained in the navigation warning is analyzed, and six types of message context labels are summarized. Through the present application, effective information in the original message data can be accurately extracted, and strong support is provided for subsequent message context classification tasks and navigation decision support, so that the trained message features are more concise and representative.

[0083] 2. The present application is based on BLS and is improved, which improves the accuracy of NAVTEX message context classification, provides intelligent navigation decision support service for mariners, can effectively reduce the work intensity and improve the work efficiency, and further enhances the safe navigation performance of the ship and enriches the function of the navigation aid equipment. BRIEF DESCRIPTION OF DRAWINGS

[0084] Figure 1 is the overall flowchart of the intelligent decision support of the NAVTEX message context of the present application.

[0085] Figure 2 is a summary diagram of the present application for sorting and encoding the message context label.

[0086] Figure 3 is the flowchart of the intelligent decision support of the message context based on BLS of the present application. DETAILED DESCRIPTION

[0087] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0088] As shown in the figure, a BLS-based NAVTEX message context intelligent decision support method has the following steps: Figure 1

[0089] S1, preprocessing the NAVTEX message data. First, the NAVTEX message data received in the ship navigation process is data cleaned to obtain a message data set D={D1, D2,..., D i ,...,D m}, then a set of message feature word sets D'={D'1, D'2,..., D' i ,…,D′ m} is obtained through word type reduction and introduction of stop words, and finally a label matrix L={L1, L2,..., Li , ...,L m} is obtained through category labeling.

[0090] S2, weighted processing of message feature words. Vectorization is used to process D' i to obtain X i ={(x i1 ,w i1 ),(x i2 ,w i2 ),…,(x ij ,w ij ),…,(x in ,w in )}, and after the feature dimension in the message data is determined, X i is simplified to X' i ={w i1 ,w i2 ,…,w ij ,…,w in}, representing the feature item weight vector of the ith message. The message feature item weight w ij is corrected by solving the word frequency, inverse document frequency and weight three processes to obtain w ij .

[0091] S3, sorting and encoding the message context label. As shown in the figure, a BLS-based NAVTEX message context intelligent decision support method has the following steps: Figure 2 ​As shown, 6 categories of message context labels are sorted and integer coded, mapping the context label matrix L to an integer label matrix L' = {L'1, L'2,..., L'6}, where L' is a set of integers representing the corresponding context categories. One-Hot encoding is performed on L' to obtain a One-Hot label matrix i ,...,L' m} where at least one position in each row is 1 and the rest are 0. i

[0092] S4, establish a context intelligent decision support model. As shown in Figure 3 , the network structure of BLS mainly includes four parts: input sample layer, feature mapping layer, enhanced node layer, and output layer. The message feature item weight matrix X' is input to the BLS network at the input layer. In the feature mapping layer, X' is mapped to t groups of message feature mapping nodes through the feature mapping function φ g (*) and connected as Z t =(Z1, Z2,..., Z t ). Z t is connected to the enhanced node layer, and a weight coefficient matrix and a bias term are randomly generated. Z s is mapped to p groups of message enhanced nodes with randomly generated weights through a nonlinear function ξ t (*) and connected as H p =(H1, H2,..., H p ). Z t and H p are connected to obtain A = (Z t |H p ). The output probability matrix of the output layer is Y = AW. Where 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 the connection weight matrix W is obtained, and the calculation formula is where I represents the identity matrix, and A T is the transpose matrix of A. For the output layer, the probability value y i in each output probability vector Y ij corresponding to the message context category is uniformly processed, so that it outputs a uniform matrix Y', and each output uniformization vector Y' i has y' i1 +y'​​​​i2 +…+y' i6 = 1. Then, by setting a decision function to filter y' in each normalized vector, the final multi-label output matrix of BLS is obtained ij The context category of the message is determined by comparing the two results, thereby obtaining intelligent decision support for the message context.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not limited thereto; 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 cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.​​​

Claims

1. A BLS-based NAVTEX message context intelligent decision support method, characterized in that: The method comprises the following steps: S1, preprocessing NAVTEX message data The NAVTEX message data is the relevant maritime safety information data received during ship navigation, and the preprocessing of the NAVTEX message data comprises four parts of data cleaning, word type reduction, introduction of stop words and category labeling, and the operation process is as follows: S11, data cleaning The original message is data cleaned, and the invalid message in the message data is eliminated; and the non-text part in the message is filtered through the method of regular expression to obtain a message data set wherein represents the th message in the set, is the number of message samples; S12, word type reduction extracting the first a packet stemming the words to remove affixes and reduce the words to their original forms; S13, introduction of stop words The message is split by the space between the words; some function words, phrases and punctuation marks in the message are removed by introducing stop words to reduce the content in the word library that does not need to be analyzed, so as to obtain a set of message feature words wherein indicates the feature word set of the th message S14, category labeling The context label is classified and marked according to corresponding actions to be taken after analyzing the content described in the NAVTEX message. Six context categories are set, which are forbidden entry, keep alert, give way, keep careful lookout, check planned route, and other behaviors. Each message feature word set corresponds to at least one category of label. The label matrix is denoted as , wherein , that is L belongs to m For the category label in the six categories of message context labels, and ; S2, weighted processing of message feature words According to the feature word set , the method of vectorization and weight correction is used to weighted processing, the steps are as follows: S21, vectorization of the feature word set of the message The first characteristic word set of the message The message characteristic vector is expressed as , i.e. , is regarded as a binary characteristic group composed of dimensional characteristic vectors; wherein, represents the first characteristic word of the first message characteristic vector; is the first characteristic item of the first characteristic word; represents the characteristic dimension, any dimension represents a characteristic item in the message characteristic vector , and is the total number of characteristic words remaining after removing the number of repeated characteristic words from the number of all characteristic words in the characteristic word set S22, correction of the feature item weight of the message When the feature dimension in the packet data is determined, the packet feature vector is simplified as: representing the first characteristic item weight vector of the packet; the step of correcting the packet characteristic item weight comprises the following steps: S221, solving the word frequency of the message feature item word frequency of the message feature item representative feature item frequency in the message feature vector The calculation formula is as follows: wherein, is a feature term in a message feature vector occurs, is the total number of occurrences of all feature terms in S222, solving the inverse document frequency of the message feature item inverse document frequency of a message feature item representative feature item the frequency of occurrence in all message feature vectors, which is calculated as follows: wherein, representing all characteristic vectors of the message characteristics occurring in the message characteristics S223, solving the weight of the message feature item The result obtained in step S2213 is multiplied by the result obtained in step S222 to obtain the characteristic term In the weight in the formula is calculated as follows: where the normalization factor The formula for calculating the normalization factor is as follows: S3, sorting and encoding the message context label The One-Hot encoding is introduced to convert the category variable into a form convenient for the classification model, and the steps are as follows: S31, integer encoding of the message context label The 6 categories of packet context labels are sequenced and integer coded, and integers 0 to 5 are sequentially assigned to the 6 context categories as described in step S14; that is, the context label matrix is mapped to the integer label matrix wherein , is an integer array representing the corresponding context category, and ; S32, One-Hot encoding of the integer label Each integer is represented as a binary vector; according to the 6 integer values in common, the length of the binary vector is determined to be 6; the index of each integer in each integer array corresponding to the sorting position is marked as 1, and the remaining positions are marked as 0, and then the One-Hot label matrix is obtained: wherein , , and at least one position in the One-Hot label vector is 1 and the rest are 0; S4, establishing a context intelligent decision support model The network structure of BLS comprises four parts: input sample layer, feature mapping layer, enhanced node layer and solving output layer weight; the output layer of BLS is improved to process the message context intelligent decision support problem, and the context intelligent decision support model training process is as follows: S41, taking the message feature item weight matrix as input In the input sample layer, the packet feature item weight matrix is input to the BLS network, and the input matrix is represented as follows: wherein, is the number of message samples, is the feature dimension, is the feature item is the weight in the message feature vector ​ S42, feature mapping of the input message feature item weight matrix The BLS network will perform feature mapping on the input packet feature item weight matrix , through a feature mapping function , , , , , The calculation formula of the first group of packet feature mapping nodes is as follows: wherein, , wherein is performed and represented as and ; and the mapped packet features are regarded as a layer, i.e., a feature mapping layer; wherein, is a linear mapping function, is a randomly generated weight matrix, is a corresponding bias term thereof; S43, generating message enhanced nodes The S42 step obtained Group message feature mapping node And the enhanced node layer is connected, and the weight coefficient matrix is randomly generated And the bias term , through the nonlinear function Map To Group random weight message enhancement node, that is, the enhanced node layer is obtained; suppose that each group of mapping generates Node, Represent the first Group message enhancement node, the calculation formula is as follows: Will Performing a parallel operation indicates obtaining... ,in ; S44, solving the connection weight matrix respectively, obtained from step S42 and obtained from step S43 and are represented by: The characteristic mapping layer and the enhanced node layer are connected in parallel to obtain Directly connected to the output end, referred to as the output layer; therefore, the probability matrix of the output is: wherein, is the connection weight matrix of the output layer, then we have: In the formula Representation matrix The pseudo-inverse matrix is ​​obtained by solving... Calculate the ridge regression approximation solution This leads to the connection weight matrix. The model training is now complete. The calculation formula is as follows: wherein denotes the identity matrix, is the transpose matrix; S45, multi-label output The probability matrix output by the BLS Y The calculation is as follows: in, 6 represents the number of message context categories. For the first The corresponding output probability vector obtained after the message sample is processed by the classification model. These are the probability values ​​for the corresponding message context categories; in the BLS output layer, for each In The homogenization process is performed using the following formula: to a uniform matrix is represented as follows: At this time, each output uniformization vector has ; and then by setting a decision function to screen each uniformization vector , the formula is as follows: in, As a threshold, when Greater than the setting At that time, Right now ;when Smaller than the setting At that time, Right now The final multi-label output matrix of BLS is then... It is expressed as follows: in, Representing the The output vector corresponding to each message. Furthermore, at least one position in the output vector is 1, and all other positions are 0; based on the One-Hot label matrix obtained in step S32 The message context category is determined by comparison, thus obtaining the message context classification result.

Images