Maritime communication data compression method and system based on dual-mode satellite fusion

Abstract

The maritime communication data compression method and system based on dual-mode satellite fusion provided by the embodiment of the present invention realizes effective compression of maritime communication messages by deeply understanding and analyzing the communication messages to be compressed. First, by obtaining the communication messages to be compressed and performing text knowledge vector mining, the semantic information of each text unit can be accurately captured, thereby improving the accuracy of text processing. Secondly, an optimized paragraph distribution attention model is adopted to strengthen the vector representation of the text unit according to its position in the message, thereby further improving the efficiency of text processing. Then, by comparing the feature similarity between the current text unit and the past text unit, a corresponding text compression weight is determined for each text unit, making the compression process more targeted. Finally, based on these text compression weights, the compression decision features of the message are determined, which effectively guides the compression process of the message and realizes efficient and targeted message compression.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for compressing maritime communication data based on dual-mode satellite fusion. Background Art

[0002] In the field of modern maritime communications, dual-mode satellite fusion technology (such as the dual-channel use of ChinaSat 6D VAST satellite channel and high-throughput satellite channel) is being widely used. This technology can effectively improve the stability and reliability of communications and provide strong support for maritime communications. However, since maritime communication messages usually contain a large amount of data, how to achieve efficient data compression while ensuring the integrity of information transmission has become an important issue.

[0003] Maritime communication data compression technology is a method mainly used to solve this problem. However, how to further optimize this method to improve its performance in practical applications is still a technical problem that needs to be solved urgently. Summary of the invention

[0004] In order to improve the above problems, the present invention provides a method and system for compressing maritime communication data based on dual-mode satellite fusion.

[0005] In a first aspect, a method for compressing maritime communication data based on dual-mode satellite fusion is provided, which is applied to a maritime communication data compression system, and the method comprises:

[0006] Acquire a communication message to be compressed containing a target maritime communication session, and perform text knowledge vector mining on each maritime service communication text unit contained in the communication message to be compressed to obtain a corresponding text unit knowledge vector; wherein the communication message to be compressed is a communication message in an application scenario of a dual-mode satellite fusion channel, and the dual-mode satellite fusion channel includes a ChinaSat 6DVAST satellite channel and a high-throughput satellite channel;

[0007] By completing the optimized paragraph distribution attention model, according to the paragraph distribution labels of each of the maritime business communication text units in the communication message to be compressed, the corresponding text unit knowledge vectors are respectively enhanced in paragraph distribution to obtain the corresponding current text unit distribution enhancement vector containing the paragraph distribution label;

[0008] Based on the text feature similarity between each current text unit distribution reinforcement vector and the previously stored past text unit distribution reinforcement vector, the text compression weight corresponding to each of the maritime business communication text units is determined respectively; wherein each past text unit distribution reinforcement vector is obtained based on the compressed communication message mining, and each past text unit distribution reinforcement vector includes the paragraph distribution label of the corresponding maritime business communication text unit, and the compressed communication message includes the compressed past maritime communication session;

[0009] Based on the obtained compression weights of each text, a message compression decision feature of the communication message to be compressed is determined.

[0010] Optionally, before completing the optimized paragraph distribution attention model, respectively performing paragraph distribution enhancement on the corresponding text unit knowledge vector according to the paragraph distribution label of each maritime business communication text unit in the communication message to be compressed, and obtaining the corresponding current text unit distribution enhancement vector containing the paragraph distribution label, the method further includes:

[0011] Based on a number of communication message training examples and their corresponding prior learning annotations, the paragraph distribution attention model to be optimized is debugged in a loop until the loop completion requirements are met, thereby obtaining the optimized paragraph distribution attention model; wherein each prior learning annotation is used to characterize the possibility that each maritime business communication text unit in the corresponding communication message training example is a text unit to be compressed;

[0012] During a debugging process, the following steps are performed:

[0013] For a communication message training example, performing text knowledge vector mining on each maritime service communication text unit of the communication message training example to obtain a corresponding text unit knowledge vector training example;

[0014] Through the paragraph distribution attention model applied in this round, based on the paragraph distribution labels of each maritime business communication text unit in the one communication message training sample, the corresponding text unit knowledge vector training sample is respectively subjected to paragraph distribution reinforcement, and the corresponding current text unit distribution reinforcement vector containing the paragraph distribution label is obtained;

[0015] Obtaining a message compression decision feature of the communication message training example according to the current text unit distribution enhancement vectors corresponding to each maritime business communication text unit of the communication message training example;

[0016] Based on the differences between several obtained message compression decision features and the corresponding prior learning annotations, the model variables of the paragraph distributed attention model applied in this round are improved.

[0017] Optionally, before determining the text compression weights corresponding to the respective maritime business communication text units based on the text feature similarities between the respective current text unit distribution reinforcement vectors and the previously stored past text unit distribution reinforcement vectors, the method further includes:

[0018] Through the second text knowledge vector mining model, text knowledge vector mining is performed on each maritime business communication text unit contained in each compressed communication message to obtain an initial text unit knowledge vector set, wherein the initial text unit knowledge vector set includes initial text unit knowledge vectors corresponding to each maritime business communication text unit of each compressed communication message;

[0019] According to the initial text unit knowledge vector set, several communication message training examples are respectively annotated to obtain corresponding prior learning annotations, and according to the several communication message training examples and the corresponding prior learning annotations, the communication message disassembly network is cyclically debugged until the cycle completion requirements are met; wherein the communication message disassembly network includes a first text knowledge vector mining model and the paragraph distributed attention model, and the first text knowledge vector mining model and the second text knowledge vector mining model share model variables;

[0020] Based on the first text knowledge vector mining model in the debugged communication message disassembly network, text knowledge vector mining is performed on each compressed communication message respectively, and paragraph distribution reinforcement is performed on the mined text unit knowledge vector through the paragraph distribution attention model in the debugged communication message disassembly network to obtain a past text unit distribution reinforcement vector set, which includes the previously stored past text unit distribution reinforcement vectors.

[0021] Optionally, the step of annotating a plurality of communication message training examples based on the initial text unit knowledge vector set to obtain corresponding prior learning annotations includes:

[0022] For the several communication message training examples, the following steps are implemented:

[0023] For a communication message training example, using the second text knowledge vector mining model, respectively perform text knowledge vector mining on each maritime service communication text unit in the communication message training example to obtain a corresponding text unit knowledge vector training example;

[0024] Determine the subordinate learning annotations of the corresponding maritime business communication text unit based on the text feature similarity between the obtained training examples of each text unit knowledge vector and the initial text unit knowledge vector set, each subordinate learning annotation is used to characterize: the possibility that the corresponding maritime business communication text unit is a text unit to be compressed;

[0025] Based on the obtained subordinate learning annotations, a priori learning annotations of the one communication message training example are obtained.

[0026] Optionally, based on the first text knowledge vector mining model in the debugged communication message disassembly network, text knowledge vector mining is performed on each of the compressed communication messages respectively, and paragraph distribution enhancement is performed on the mined text unit knowledge vector through the paragraph distribution attention model in the debugged communication message disassembly network to obtain a past text unit distribution enhancement vector set, including:

[0027] According to the first text knowledge vector mining model, text knowledge vector mining is performed on each maritime business communication text unit of each compressed communication message, and paragraph distribution enhancement is performed on the text unit knowledge vector obtained by mining through the paragraph distribution attention model in the debugged communication message disassembly network to obtain an enhanced knowledge vector set, wherein the enhanced knowledge vector set includes the enhanced knowledge vectors corresponding to each maritime business communication text unit of each compressed communication message;

[0028] According to the text feature similarity between every two enhanced knowledge vectors in the enhanced knowledge vector set, part of the enhanced knowledge vectors are screened from the enhanced knowledge vector set to obtain the past text unit distribution enhanced vector set.

[0029] Optionally, the step of selecting part of the enhanced knowledge vectors from the enhanced knowledge vector set according to the text feature similarity between every two enhanced knowledge vectors in the enhanced knowledge vector set to obtain the enhanced vector set of past text unit distribution includes:

[0030] Selecting at least one enhanced knowledge vector from the enhanced knowledge vector set, and adjusting the enhanced vector set of past text unit distribution;

[0031] Circularly processing the past text unit distribution reinforcement vector set;

[0032] In each cycle, the following steps are performed:

[0033] For each enhanced knowledge vector in the enhanced knowledge vector set, a corresponding knowledge vector tuple is determined, each knowledge vector tuple comprising a enhanced knowledge vector and a past text unit distribution enhancement vector in the past text unit distribution enhancement vector set having the highest text feature similarity with the one enhanced knowledge vector;

[0034] For each past text unit distribution reinforcement vector included in each knowledge vector binary, the reinforced knowledge vector having the smallest text feature similarity with each past text unit distribution reinforcement vector is added to the past text unit distribution reinforcement vector set.

[0035] Optionally, by completing the optimized paragraph distribution attention model, according to the paragraph distribution labels of each of the maritime business communication text units in the communication message to be compressed, the corresponding text unit knowledge vectors are respectively subjected to paragraph distribution enhancement, and the corresponding current text unit distribution enhancement vector containing the paragraph distribution label is obtained, including:

[0036] Through the optimized paragraph distribution attention model, the following steps are implemented for each maritime business communication text unit:

[0037] For a maritime business communication text unit, according to a first paragraph distribution label of the maritime business communication text unit in the communication message to be compressed, obtaining a corresponding first paragraph distribution variable;

[0038] Obtaining a corresponding second paragraph distribution variable according to a second paragraph distribution label of the one maritime business communication text unit in the communication message to be compressed;

[0039] According to the text unit knowledge vector of the maritime business communication text unit, the first paragraph distribution variable and the second paragraph distribution variable, a corresponding current text unit distribution reinforcement vector is obtained.

[0040] Optionally, performing text knowledge vector mining on each maritime service communication text unit contained in the communication message to be compressed to obtain a corresponding text unit knowledge vector includes:

[0041] Using a recursive strategy, mining text knowledge vectors for the communication message to be compressed;

[0042] In each recursive round, the following steps are performed:

[0043] Based on the pre-configured reversible operator, the incoming vector of this round of recursion is reversibly processed to obtain a reversible knowledge vector; wherein, in the first round of recursion, the incoming vector is the communication message to be compressed, and in the non-first round of recursion, the incoming vector is the output vector of the previous round of recursion;

[0044] Performing feature integration based on the reversible knowledge vector and the input vector to obtain an aggregated vector;

[0045] If this round of recursion is the last round of recursion, the aggregated vector is used as the output vector of this round of recursion;

[0046] If the current round of recursion is not the last round of recursion, average pooling processing based on the vector dimension is performed on the aggregated vector, and the obtained average pooling vector is used as the output vector of the current round of recursion.

[0047] Optionally, the determining of the text compression weights corresponding to the respective maritime business communication text units based on the text feature similarity between the respective current text unit distribution reinforcement vectors and the previously stored past text unit distribution reinforcement vectors includes:

[0048] For each current text unit distribution reinforcement vector, the following steps are implemented:

[0049] For a current text unit distribution reinforcement vector, determining a past text unit distribution reinforcement vector having the highest text feature similarity to the current text unit distribution reinforcement vector;

[0050] According to the difference between the past text unit distribution enhancement vector and the current text unit distribution enhancement vector, a text compression weight of the maritime business communication text unit corresponding to the current text unit distribution enhancement vector is determined.

[0051] Optionally, the determining of the text compression weights corresponding to the respective maritime business communication text units based on the text feature similarity between the respective current text unit distribution reinforcement vectors and the previously stored past text unit distribution reinforcement vectors includes:

[0052] For each current text unit distribution reinforcement vector, the following steps are implemented:

[0053] For a current text unit distribution reinforcement vector, determining a past text unit distribution reinforcement vector having the highest text feature similarity to the current text unit distribution reinforcement vector;

[0054] Determine, from the previously stored past text unit distribution reinforcement vectors, a number of past text unit distribution reinforcement vectors whose text feature similarity with the past text unit distribution reinforcement vector with the highest text feature similarity is not greater than a preset text feature similarity threshold;

[0055] Based on the past text unit distribution reinforcement vector with the highest text feature similarity and the differences between each of the several past text unit distribution reinforcement vectors and the current text unit distribution reinforcement vector, the text compression weight of the maritime business communication text unit corresponding to the current text unit distribution reinforcement vector is determined.

[0056] Optionally, performing text knowledge vector mining on each maritime service communication text unit contained in the communication message to be compressed to obtain a corresponding text unit knowledge vector includes:

[0057] According to the target subject keyword of the target maritime communication session, a corresponding target reference communication message is obtained from a plurality of reference communication messages stored in advance;

[0058] Performing text matching on the communication message to be compressed according to the set word vector in the target reference communication message to obtain a text matching result;

[0059] According to a preset message set in the target reference communication message, obtaining a local message set to which the target maritime communication session belongs from the text matching result;

[0060] Text knowledge vector mining is performed on each maritime business communication text unit contained in the local message set to obtain a corresponding text unit knowledge vector.

[0061] In a second aspect, a maritime communication data compression system is provided, comprising a processor and a memory communicating with each other, wherein the processor is used to retrieve a computer program from the memory and implement the method described in the first aspect by running the computer program.

[0062] According to a third aspect, a computer-readable storage medium is provided, on which a computer program is stored, and the computer program implements the method described in the first aspect when running.

[0063] The maritime communication data compression method and system based on dual-mode satellite fusion provided by the embodiment of the present invention realizes effective compression of maritime communication messages by deeply understanding and analyzing the communication messages to be compressed. First, by obtaining the communication messages to be compressed and mining the text knowledge vector, the semantic information of each text unit can be accurately captured, thereby improving the accuracy of text processing. Secondly, the optimized paragraph distribution attention model is adopted to strengthen the vector representation of the text unit according to its position in the message, thereby further improving the efficiency of text processing. Then, by comparing the feature similarity between the current text unit and the past text unit, the corresponding text compression weight is determined for each text unit, making the compression process more targeted. Finally, based on these text compression weights, the compression decision features of the message are determined, which effectively guides the compression process of the message and realizes efficient and targeted message compression. Overall, the technical solution not only improves the transmission efficiency of maritime communication, but also reduces the communication cost, and has important practical value in the field of maritime communication. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0065] Figure 1 A flow chart of a method for compressing maritime communication data based on dual-mode satellite fusion provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0066] In the submarine cable operation scenario, the stability of communication and the integrity of data are crucial. To meet this requirement, the embodiment of the present invention can adopt an innovative dual-mode satellite fusion channel technology, which combines the ChinaSat 6D VAST satellite channel and the high-throughput satellite channel to ensure that the data at the operation site can be transmitted back to the rear in a timely and complete manner.

[0067] In this process, an efficient data compression method was designed, especially for offshore meteorological monitoring, hydrological monitoring, toxic and combustible gas and other environmental monitoring data, as well as human vital signs monitoring data. The core of this method is to intelligently determine the compression weight of each data segment through text knowledge vector mining and paragraph distribution reinforcement, so as to maximize the compression efficiency while ensuring the integrity of data transmission.

[0068] First, the communication messages to be compressed containing the target maritime communication session are obtained. These messages cover all the key monitoring data of the submarine cable operation surface, such as meteorological and hydrological information, environmental monitoring data, and vital signs data of operators. Next, the text knowledge vectors of each maritime business communication text unit in these messages are deeply mined to extract key text features and form text unit knowledge vectors.

[0069] To further optimize the compression process, an innovative paragraph distribution attention model is introduced. This model assigns a specific paragraph distribution label to each text unit based on its position and context in the message. These labels not only reflect the importance of the content of the text unit, but also reflect their role in the overall message structure. In this way, the feature representation of each text unit can be more accurately enhanced, and a text unit distribution enhancement vector containing the paragraph distribution label can be obtained.

[0070] These reinforcement vectors are then compared with the previously stored text unit distribution reinforcement vectors to find the similarity of text features between them. This comparison not only takes into account the text content, but also the position and role of the text unit in the message. In this way, an accurate compression weight can be assigned to each maritime business communication text unit.

[0071] Finally, based on these carefully calculated text compression weights, efficient message compression decision features were developed. This not only greatly improved the efficiency and accuracy of data compression, but also ensured that key data from the submarine cable operation site could be transmitted back to the rear in real time and in full, providing a strong guarantee for the safety and efficiency of offshore operations.

[0072] In some other application scenarios, the technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. The embodiments of the present invention and the technical features in the embodiments can be combined with each other without conflict.

[0073] Figure 1 A method for compressing maritime communication data based on dual-mode satellite fusion is shown, which is applied to a maritime communication data compression system. The method includes the following steps 110 to 140.

[0074] Step 110: obtain a communication message to be compressed that contains a target maritime communication session, and perform text knowledge vector mining on each maritime business communication text unit contained in the communication message to be compressed to obtain a corresponding text unit knowledge vector; wherein the communication message to be compressed is a communication message in an application scenario of a dual-mode satellite fusion channel, and the dual-mode satellite fusion channel includes a ChinaSat 6DVAST satellite channel and a high-throughput satellite channel.

[0075] Among them, the target maritime communication session refers to a communication dialogue conducted at sea. For example, communication between two ships or communication between a ship and a land base station. The communication message to be compressed is a communication message that needs data compression. For example, a ship sends a message containing a large amount of data to another ship, which may need to be compressed to reduce transmission time and bandwidth usage. The maritime business communication text unit is the text information related to the specific business contained in the communication message to be compressed in maritime communication. For example, it may include content about navigation direction, weather conditions, cargo information, etc. Text knowledge vector mining refers to extracting useful knowledge from maritime business communication text units by using natural language processing technology and converting this knowledge into a vector form that can be understood and processed by machines. The text unit knowledge vector is a point in a multidimensional space with specific mathematical properties (such as length, direction) obtained after text knowledge vector mining, which can represent the semantic information of a specific text unit. In the dual-mode satellite fusion channel application scenario, two different satellite communication technologies (ChinaSat 6DVAST satellite channel and high-throughput satellite channel) are used to provide communication services. ChinaSat 6DVAST satellite channel is a communication method using ChinaSat 6D satellite. VAST stands for Very Small Aperture Terminal, which refers to a satellite communication system with a small aperture of the receiving antenna of the satellite ground station. High-throughput satellite channel refers to the channel for communication via high-throughput satellite (HTS). HTS is a type of satellite that can provide greater bandwidth than traditional satellites, so that more communication capacity can be provided under the same frequency resources.

[0076] In some other specific examples, text unit knowledge vectors, also known as word vectors, are vectors that convert words, phrases, or sentences in a text into vectors in a high-dimensional space. These vectors can capture the semantic information of the text unit, and similar text units will be mapped to similar positions in this high-dimensional space.

[0077] In the scenario of maritime communication messages, we first need to convert each piece of information in the message (for example, the name of the ship, its geographical location, its speed, etc.) into a corresponding word vector. The following is a simple example:

[0078] For example, there is a simple vocabulary: {"ship": 1, "ocean": 2, "speed": 3, "position": 4}, where each word is assigned a unique numeric ID. Then, each word can be represented by a four-dimensional vector, for example, "ship" can be represented as [1, 0, 0, 0], "ocean" can be represented as [0, 1, 0, 0], and so on. This is the most basic one-hot encoding method.

[0079] However, One-hot encoding cannot express the relationship between words. To solve this problem, more complex models such as Word2Vec or GloVe can be used to generate word vectors. These models can generate word vectors based on the context of the word, so that words with similar semantics are closer in the vector space.

[0080] For example, after training, the word vector of "ship" is [0.1, 0.3, 0.2, 0.4], and the word vector of "ocean" is [0.2, 0.3, 0.1, 0.2]. Then, when a maritime communication message containing "ship" and "ocean" is received, the message can be converted into a combination of two word vectors, that is, [[0.1, 0.3, 0.2, 0.4], [0.2, 0.3, 0.1, 0.2]].

[0081] In this way, these word vectors can be used to perform various operations, such as determining the topic of a message by calculating the cosine similarity between word vectors, or using machine learning algorithms to classify messages.

[0082] Step 120, by completing the optimized paragraph distribution attention model, according to the paragraph distribution labels of each maritime business communication text unit in the communication message to be compressed, the corresponding text unit knowledge vector is respectively subjected to paragraph distribution reinforcement, so as to obtain the corresponding current text unit distribution reinforcement vector containing the paragraph distribution label.

[0083] Among them, the optimized paragraph distribution attention model is a machine learning model that focuses on the information distribution of different paragraphs in the message. Through training, this model can understand and grasp the importance of each paragraph, helping to improve the quality of the compressed message. The paragraph distribution label is a representation of the position of each text unit in the entire message (that is, in which paragraph). For example, if a text unit appears in the third paragraph of the message, its paragraph distribution label may be "3". Paragraph distribution reinforcement is a process through which the paragraph distribution attention model and paragraph distribution label can be used to strengthen the text unit knowledge vector. In other words, through paragraph distribution reinforcement, the model can better understand the position and importance of the text unit in the entire message. The current text unit distribution reinforcement vector is the text unit knowledge vector obtained after paragraph distribution reinforcement. It contains the content information of the text unit and the paragraph distribution information of the unit in the message.

[0084] The following is a simple example. Assume that there are three paragraphs in the communication message to be compressed, and each paragraph contains some maritime business communication text units. The optimized paragraph distribution attention model may find that the information in the second paragraph is the most important to the overall message. Therefore, these text units in the second paragraph will be given a higher weight (that is, "2" as the paragraph distribution label). Next, paragraph distribution reinforcement is performed, that is, this weight is used to strengthen the text unit knowledge vector in the second paragraph. Finally, the current text unit distribution reinforcement vector containing the paragraph distribution label is obtained, which contains both the original text information and the position and importance information of the text unit in the message.

[0085] Step 130: Determine the text compression weights corresponding to the maritime business communication text units based on the text feature similarities between the current text unit distribution enhancement vectors and the previously stored past text unit distribution enhancement vectors.

[0086] Among them, each past text unit distribution enhancement vector is obtained based on the compressed communication message mining, and each past text unit distribution enhancement vector contains the paragraph distribution label of the corresponding maritime business communication text unit, and the compressed communication message contains the compressed past maritime communication session.

[0087] Among them, the previously stored past text unit distribution reinforcement vector is a vector mined from the compressed communication message, and each vector contains the paragraph distribution label of the corresponding maritime business communication text unit. These vectors can help understand the importance and priority of each text unit in the previous communication session, and provide a reference for the subsequent text compression. Text feature similarity refers to the similarity between two or more texts in terms of semantics, theme, style, etc. In this step, the current text unit distribution reinforcement vector is compared with the past vector to evaluate the similarity between them. The text compression weight is determined according to the result of text feature similarity, and the text compression weight corresponding to each maritime business communication text unit is determined. The higher the weight, the higher the priority of the text unit during compression. The compressed communication messages are past maritime communication sessions that have been compressed, which contain rich business communication text units. These messages are used to generate past text unit distribution reinforcement vectors. Past maritime communication sessions that have completed compression refer to maritime communication sessions that have completed compression processing in the application scenario of dual-mode satellite fusion channels.

[0088] Example: Assume that there is a maritime communication message to be compressed, which contains three text units A, B and C. First, the past text unit distribution reinforcement vector is obtained from the compressed communication message. Then, by comparing the feature similarity between the current text unit and the past text unit, the compression weight corresponding to each text unit is determined. For example, if text unit A has the highest similarity with the past text unit, it may get the highest compression weight, indicating that A has the highest priority when compressing the message.

[0089] Step 140: Determine the message compression decision feature of the communication message to be compressed based on the obtained compression weights of each text.

[0090] In step 140, the message compression decision feature is a set of features determined according to the compression weight of each text unit, which will be used to decide how to compress the communication message to be compressed. These features may include but are not limited to the importance, location information, content similarity, etc. of the text unit.

[0091] For example, a communication message to be compressed contains three paragraphs, each of which has multiple text units. In step 130, the compression weight of each text unit has been obtained. These weights may be based on many factors, such as the content of the text unit, its position in the message, its similarity with past text units, etc.

[0092] In step 140, these weights can be used to determine message compression decision features. For example, if a text unit has a high compression weight, most of the information of the text unit may be retained during the compression process. Conversely, if a text unit has a low compression weight, the information of the text unit may be deleted or simplified during the compression process. In this way, the message compression decision features can guide how to more effectively compress the message to keep the most important information while reducing the size of the message as much as possible.

[0093] In step 130 and step 140, the message compression decision feature is determined based on the text compression weight. Each maritime business communication text unit will obtain a compression weight according to the text feature similarity between it and the distribution enhancement vector of the past text unit.

[0094] For example, there is a maritime communication message to be compressed, which contains three text units A, B and C. Each text unit has been enhanced by paragraph distribution to generate the current text unit distribution enhancement vector. Then, the stored past text unit distribution enhancement vectors are obtained from the database, and their text feature similarity with the current text unit distribution enhancement vector is compared. Assuming that the features of text unit A are most similar to past text units, it may obtain the highest compression weight, such as 0.6; text unit B has the second highest similarity and obtains a weight of 0.3; and text unit C has the lowest similarity to past text units, so it obtains the lowest weight, such as 0.1.

[0095] Next, in step 140, these compression weights are used to determine the message compression decision features. In other words, these weights are used to determine which text units should be compressed first and which can be processed later. In this example, because A has the highest weight, it will be compressed first; followed by B, which has the second highest weight, and finally C.

[0096] In this way, not only can the effectiveness of message compression be guaranteed (the most important information is compressed first), but also the efficiency of compression can be improved (ineffective trials and errors are avoided).

[0097] Below, a specific maritime communication scenario is used as an example to illustrate the technical solutions of the above steps 110 to 140.

[0098] In step 110, a communication message to be compressed is received, which comes from the communication in the dual-mode satellite fusion channel application scenario, and contains the target maritime communication session. The message contains various maritime business communication text units, such as: "Sea wind force 7, waves 3 meters", "Ship speed 12 knots, heading 30 degrees north east", etc. Then, through text knowledge vector mining, each text unit is analyzed and understood, and the semantic information of the text is extracted and converted into the corresponding text unit knowledge vector.

[0099] In step 120, the optimized paragraph distribution attention model is used to strengthen the text unit knowledge vector according to the paragraph distribution label of each maritime business communication text unit in the communication message to be compressed. For example, if a text unit often appears at the beginning of the message, the knowledge vector of the text unit will be strengthened. In this way, the current text unit distribution strengthening vector containing the paragraph distribution label can be obtained.

[0100] In step 130, the text feature similarity between the current text unit distribution enhancement vector and the previously stored past text unit distribution enhancement vector is compared. The past text unit distribution enhancement vector is obtained by mining the compressed communication message. For example, if a current text unit is very similar to a past text unit in semantics, then the text unit will be assigned a higher text compression weight.

[0101] In step 140, the message compression decision features of the communication message to be compressed are determined according to the text compression weights of each text unit. These decision features will be used to guide the message compression process to achieve efficient compression of the maritime communication message.

[0102] In summary, by deeply understanding and analyzing the communication messages to be compressed, effective compression of maritime communication messages is achieved. First, by obtaining the communication messages to be compressed and mining the text knowledge vector, the semantic information of each text unit can be accurately captured, which improves the accuracy of text processing. Secondly, the optimized paragraph distributed attention model is adopted to strengthen the vector representation of the text unit according to its position in the message, which further improves the efficiency of text processing. Then, by comparing the feature similarity between the current text unit and the past text unit, the corresponding text compression weight is determined for each text unit, making the compression process more targeted. Finally, based on these text compression weights, the compression decision features of the message are determined, which effectively guides the compression process of the message and realizes efficient and targeted message compression. Overall, this technical solution not only improves the transmission efficiency of maritime communications, but also reduces the communication cost, and has important practical value in the field of maritime communications.

[0103] In some optional embodiments, before the paragraph distribution attention model is completed and the corresponding text unit knowledge vectors are respectively paragraph distribution strengthened according to the paragraph distribution labels of each maritime business communication text unit in the communication message to be compressed, and the corresponding current text unit distribution strengthening vector containing the paragraph distribution label is obtained, the method further includes: based on a number of communication message training examples and their corresponding prior learning annotations, the paragraph distribution attention model to be optimized is loop-debugged until the loop completion requirements are met, so as to obtain the completed optimized paragraph distribution attention model.

[0104] Among them, each prior learning annotation is used to characterize the possibility that each maritime business communication text unit in the corresponding communication message training sample is a text unit that needs to be compressed.

[0105] During a debugging process, the following steps are implemented: for a communication message training sample, text knowledge vector mining is performed on each maritime business communication text unit of the communication message training sample to obtain a corresponding text unit knowledge vector training sample; through the paragraph distribution attention model applied in this round, based on the paragraph distribution label of each maritime business communication text unit in the communication message training sample, paragraph distribution reinforcement is performed on the corresponding text unit knowledge vector training sample to obtain the corresponding current text unit distribution reinforcement vector containing the paragraph distribution label; according to the current text unit distribution reinforcement vector corresponding to each maritime business communication text unit of the communication message training sample, the message compression decision feature of the communication message training sample is obtained; based on the difference between the obtained several message compression decision features and the corresponding prior learning annotations, the model variables of the paragraph distribution attention model applied in this round are improved.

[0106] The main idea of ​​the above technical solution is to obtain the optimized model by cyclically debugging the paragraph distributed attention model to be optimized based on several communication message training examples and their prior learning annotations.

[0107] The following is a specific example: there is a series of communication message training examples, such as: ["Sea wind force 7, waves 3 meters", "Ship speed 12 knots, heading 30 degrees north east"], etc., as well as their corresponding prior learning annotations, which represent the possibility that each maritime business communication text unit in the corresponding communication message training examples is a text unit that needs to be compressed.

[0108] During a debugging process, a communication message training example is selected, for example, "Wind force 7 at sea, waves 3 meters". Then, text knowledge vector mining is performed on each maritime business communication text unit of this training example to obtain the corresponding text unit knowledge vector training example.

[0109] Next, through the paragraph distribution attention model applied in this round, the corresponding text unit knowledge vector training examples are enhanced according to the paragraph distribution labels of each maritime business communication text unit in the one communication message training example, and the current text unit distribution enhancement vector containing the paragraph distribution label is obtained.

[0110] Then, the message compression decision feature of the communication message training sample is obtained according to the current text unit distribution reinforcement vector corresponding to each maritime business communication text unit of the communication message training sample.

[0111] Finally, based on the differences between several obtained message compression decision features and the corresponding prior learning annotations, the model variables of the paragraph distributed attention model applied in this round are improved.

[0112] The above steps will be repeated until the loop completion requirements are met (such as reaching the preset maximum number of iterations or the model performance can no longer be significantly improved), and finally the optimized paragraph distributed attention model is obtained. Then, this model can be used to compress the communication messages to be compressed.

[0113] In some other possible embodiments, before determining the text compression weights corresponding to each of the maritime business communication text units based on the text feature similarities between each current text unit distribution reinforcement vector and the previously stored past text unit distribution reinforcement vectors, the method also includes steps 210 to 230.

[0114] Step 210: Through the second text knowledge vector mining model, text knowledge vector mining is performed on each maritime business communication text unit contained in each compressed communication message to obtain an initial text unit knowledge vector set, wherein the initial text unit knowledge vector set includes the initial text unit knowledge vector corresponding to each maritime business communication text unit of each compressed communication message.

[0115] Step 220: annotate several communication message training examples respectively according to the initial text unit knowledge vector set to obtain corresponding prior learning annotations, and perform loop debugging on the communication message disassembly network according to the several communication message training examples and their corresponding prior learning annotations until the loop completion requirements are met.

[0116] The communication message disassembly network includes a first text knowledge vector mining model and the paragraph distribution attention model, and the first text knowledge vector mining model and the second text knowledge vector mining model share model variables.

[0117] Step 230: Based on the first text knowledge vector mining model in the debugged communication message disassembly network, text knowledge vector mining is performed on each of the compressed communication messages respectively, and paragraph distribution reinforcement is performed on the mined text unit knowledge vectors through the paragraph distribution attention model in the debugged communication message disassembly network to obtain a past text unit distribution reinforcement vector set, wherein the past text unit distribution reinforcement vector set includes the previously stored past text unit distribution reinforcement vectors.

[0118] First, step 210 involves using the second text knowledge vector mining model to deeply understand and analyze each maritime business communication text unit contained in each compressed communication message. For example, for a compressed maritime communication message, such as "Ship speed 12 knots, north by east 30 degrees", through this model, the initial text unit knowledge vector of each text unit (such as "Ship speed 12 knots" and "North by east 30 degrees") can be obtained to form an initial text unit knowledge vector set.

[0119] Next, in step 220, several communication message training examples are annotated according to the initial text unit knowledge vector set. These annotations serve as prior learning annotations to help better understand and annotate the text. Then, these training examples are input into the communication message disassembly network together with their prior learning annotations for loop debugging. This network includes a first text knowledge vector mining model and a paragraph distribution attention model, both of which share model variables. The debugging will end only when the network output meets the loop completion requirements.

[0120] Finally, in step 230, based on the debugged communication message disassembly network, the text knowledge vectors of each compressed communication message are mined again, and the debugged paragraph distribution attention model is used to strengthen the mined text unit knowledge vectors. In this way, a set of past text unit distribution strengthening vectors is obtained, including the previously stored past text unit distribution strengthening vectors.

[0121] In general, by adding steps 210-230, this technical solution makes the compression of maritime communication messages more accurate and efficient, further improving the quality and efficiency of maritime communications.

[0122] In some alternative embodiments, the method of annotating several communication message training samples based on the initial text unit knowledge vector set to obtain corresponding prior learning annotations includes: for the several communication message training samples, the following steps are implemented: for one communication message training sample, using the second text knowledge vector mining model, respectively performing text knowledge vector mining on each maritime business communication text unit in the one communication message training sample to obtain a corresponding text unit knowledge vector training sample; based on the text feature similarity between each obtained text unit knowledge vector training sample and the initial text unit knowledge vector set, determining the subordinate learning annotations of the corresponding maritime business communication text unit, each subordinate learning annotation is used to characterize the possibility that the corresponding maritime business communication text unit is a text unit to be compressed; based on the obtained subordinate learning annotations, obtaining the prior learning annotations of the one communication message training sample.

[0123] In this technical solution, the communication message training samples will be annotated with the initial text unit knowledge vector set to obtain the corresponding prior learning annotations. This is a typical supervised learning method. By labeling the training data, the model can learn the mapping relationship between input and output.

[0124] For example, there already exists an initial set of text unit knowledge vectors, which contains semantic information of various maritime business communication text units, such as vector representations of wind force, waves, ship speed and heading, etc.

[0125] Then, there are several communication message training examples, such as: "Sea wind force 7, waves 3 meters", "Ship speed 12 knots, heading 30 degrees north east", etc. For each training example, the second text knowledge vector mining model will be used to analyze each maritime business communication text unit contained in it to obtain the corresponding text unit knowledge vector training example.

[0126] Next, the obtained text unit knowledge vector training examples are compared with the initial text unit knowledge vector set, and the subordinate learning annotations of each maritime business communication text unit are determined according to the text feature similarity between them. These annotations characterize the possibility of each text unit being a text unit to be compressed.

[0127] Finally, the prior learning annotations of the communication message training example are obtained based on all the subordinate learning annotations obtained.

[0128] In summary, this technical solution can effectively generate prior learning annotations through in-depth understanding and analysis of the initial text unit knowledge vector set and communication message training examples, which is of great help to the training of the model. In this way, the model can be continuously optimized during the training process to improve its accuracy and stability, thereby improving the effect of maritime communication message compression and achieving more efficient communication transmission.

[0129] Under some other possible design ideas, based on the first text knowledge vector mining model in the debugged communication message disassembly network, text knowledge vector mining is performed on each compressed communication message respectively, and the paragraph distribution of the mined text unit knowledge vector is enhanced through the paragraph distribution attention model in the debugged communication message disassembly network to obtain a past text unit distribution enhanced vector set, including: according to the first text knowledge vector mining model, text knowledge vector mining is performed on each maritime business communication text unit of each compressed communication message respectively, and the paragraph distribution of the mined text unit knowledge vector is enhanced through the paragraph distribution attention model in the debugged communication message disassembly network to obtain an enhanced knowledge vector set, the enhanced knowledge vector set includes the enhanced knowledge vectors corresponding to each maritime business communication text unit of each compressed communication message; according to the text feature similarity between every two enhanced knowledge vectors in the enhanced knowledge vector set, some enhanced knowledge vectors are selected from the enhanced knowledge vector set to obtain the past text unit distribution enhanced vector set.

[0130] The above technical solution mainly involves the processing of compressed communication messages, including mining text knowledge vectors, strengthening vectors and screening some vectors to obtain a set of strengthened vectors of past text unit distribution.

[0131] For example, there is a series of compressed communication messages, such as "Ship speed 12 knots, 30 degrees north by east" and "Sea waves 3 meters, wind force 7". First, the debugged first text knowledge vector mining model is used to mine text knowledge vectors for each maritime business communication text unit in each compressed communication message. Then, the debugged paragraph distributed attention model is used to enhance the mined text unit knowledge vectors, thereby obtaining an enhanced knowledge vector set.

[0132] Then, based on the similarity of text features between these enhanced knowledge vectors, some vectors are selected from the enhanced knowledge vector set. For example, if two enhanced knowledge vectors are very similar in semantics, they may be selected into the enhanced vector set of past text unit distribution.

[0133] Finally, the obtained enhanced vector set of past text unit distribution will be used as the basis for determining text compression weights, further improving the accuracy and efficiency of maritime communication message compression.

[0134] In general, this technical solution improves the effect of message compression by deeply understanding and analyzing compressed messages, and finely mining, strengthening and screening knowledge vectors, thereby effectively optimizing the quality and efficiency of maritime communications.

[0135] In some preferred embodiments, according to the text feature similarity between every two reinforced knowledge vectors in the reinforced knowledge vector set, some reinforced knowledge vectors are selected from the reinforced knowledge vector set to obtain the past text unit distribution reinforcement vector set, including: selecting at least one reinforced knowledge vector from the reinforced knowledge vector set and adjusting the past text unit distribution reinforcement vector set; looping the past text unit distribution reinforcement vector set; wherein, in each loop processing, the following steps are implemented: for each reinforced knowledge vector in the reinforced knowledge vector set, corresponding knowledge vector tuples are determined respectively, each knowledge vector tuple contains a reinforced knowledge vector and a past text unit distribution reinforcement vector in the past text unit distribution reinforcement vector set with the highest text feature similarity to the one reinforced knowledge vector; for each past text unit distribution reinforcement vector contained in each knowledge vector tuple, the reinforced knowledge vector with the smallest text feature similarity to each past text unit distribution reinforcement vector is added to the past text unit distribution reinforcement vector set.

[0136] The key to the above technical solution is to select some enhanced knowledge vectors from the enhanced knowledge vector set and construct an enhanced vector set of past text unit distribution. The basis for selection is the similarity of text features between every two enhanced knowledge vectors.

[0137] The following is a specific example: suppose there is a set of enhanced knowledge vectors, which may come from various maritime communication messages, such as the enhanced knowledge vectors of "wind force 7", "waves 3 meters", etc.

[0138] First, at least one enhanced knowledge vector needs to be selected from this set and added to the enhanced vector set of past text unit distribution. Then, this enhanced vector set of past text unit distribution is processed cyclically.

[0139] In each loop processing, for each enhanced knowledge vector in the enhanced knowledge vector set, a corresponding knowledge vector tuple is determined. Each tuple includes an enhanced knowledge vector and a vector in the enhanced vector set of past text unit distribution that has the highest text feature similarity with the enhanced knowledge vector.

[0140] Next, for each past text unit distribution reinforcement vector in the knowledge vector binary, the reinforced knowledge vector with the smallest text feature similarity is added to the past text unit distribution reinforcement vector set.

[0141] In this way, a more representative set of past text unit distribution reinforcement vectors is constructed through cyclic processing.

[0142] In summary, this technical solution provides an effective method for screening and updating the past text unit distribution enhancement vector set, which can better capture and retain valuable information and improve the compression effect of maritime communication messages. This can not only improve the compression quality, but also save communication resources to a certain extent, which has important practical value for maritime communications.

[0143] In some exemplary embodiments, by completing the optimized paragraph distribution attention model, according to the paragraph distribution labels of each of the maritime business communication text units in the communication message to be compressed, the corresponding text unit knowledge vectors are respectively subjected to paragraph distribution reinforcement, and the corresponding current text unit distribution reinforcement vector containing the paragraph distribution label is obtained, including: by completing the optimized paragraph distribution attention model, for each of the maritime business communication text units, the following steps are implemented: for a maritime business communication text unit, according to the first paragraph distribution label of the maritime business communication text unit in the communication message to be compressed, the corresponding first paragraph distribution variable is obtained; according to the second paragraph distribution label of the maritime business communication text unit in the communication message to be compressed, the corresponding second paragraph distribution variable is obtained; according to the text unit knowledge vector, the first paragraph distribution variable and the second paragraph distribution variable of the maritime business communication text unit, the corresponding current text unit distribution reinforcement vector is obtained.

[0144] In this technical solution, the optimized paragraph distribution attention model is mainly used to deeply process and analyze each maritime business communication text unit to obtain the current text unit distribution reinforcement vector containing the paragraph distribution label.

[0145] For example, suppose there is a communication message to be compressed: "Ship speed 12 knots, 30 degrees north east. Waves 3 meters, wind force 7." There are four maritime business communication text units in this message: "Ship speed 12 knots", "30 degrees north east", "waves 3 meters" and "wind force 7".

[0146] For each text unit, such as "Ship speed 12 knots", firstly, according to its first paragraph distribution label in the communication message to be compressed, the corresponding first paragraph distribution variable is obtained through the optimized paragraph distribution attention model. Then, similarly, according to the second paragraph distribution label of the text unit in the communication message to be compressed, the corresponding second paragraph distribution variable is obtained.

[0147] Then, according to the text unit knowledge vector, the first paragraph distribution variable and the second paragraph distribution variable of the text unit, the corresponding current text unit distribution enhancement vector is obtained. This enhancement vector not only contains the semantic information of the text unit, but also contains its position information in the message, so as to better preserve the context and structure of the original message.

[0148] By repeating the above steps, corresponding current text unit distribution enhancement vectors can be obtained for all maritime business communication text units in the communication message to be compressed.

[0149] In general, this technical solution effectively improves the accuracy and efficiency of communication message compression by carefully processing and analyzing each maritime business communication text unit, thereby further optimizing the quality and efficiency of maritime communications.

[0150] In another example, horizontal paragraph distribution tags and vertical paragraph distribution tags are two types of tags used to locate and describe the positions of text units when processing text data.

[0151] Horizontal paragraph distribution label (first paragraph distribution label): It is usually used to indicate the relative position of a text unit in a given paragraph or sentence. For example, in a maritime communication message, such as "Wind force 7, waves 3 meters", "Wind force 7" may be assigned horizontal paragraph distribution label 1, and "waves 3 meters" may be assigned horizontal paragraph distribution label 2. These labels can help the model understand the sequential relationship between text units, so as to better understand and process text information.

[0152] Vertical paragraph distribution label (second paragraph distribution label): It is usually used to indicate the relative position of a text unit in multiple paragraphs or sentences. For example, if there are multiple maritime communication messages, such as: "Wind force 7, waves 3 meters", "Ship speed 12 knots, heading 30 degrees north east". Then, the message "Wind force 7, waves 3 meters" may be assigned vertical paragraph distribution label 1, and the message "Ship speed 12 knots, heading 30 degrees north east" may be assigned vertical paragraph distribution label 2. These labels can help the model understand the sequential relationship between different messages, so as to better process and understand information across messages.

[0153] In practical applications, horizontal paragraph distribution labels and vertical paragraph distribution labels are often used together to fully and accurately express the position information of text units in the message. In this way, the model can make full use of this position information to better understand and process the message, further improving the effect of maritime communication message compression.

[0154] In some preferred embodiments, text knowledge vector mining is performed on each maritime business communication text unit contained in the communication message to be compressed to obtain a corresponding text unit knowledge vector, including: adopting a recursive strategy to mine text knowledge vectors for the communication message to be compressed; in each round of recursion, the following steps are implemented: based on a pre-configured reversible operator, reversibly processing the incoming vector of this round of recursion to obtain a reversible knowledge vector; wherein, in the first round of recursion, the incoming vector is the communication message to be compressed, and in the non-first round of recursion, the incoming vector is the output vector of the previous round of recursion; feature integration is performed based on the reversible knowledge vector and the incoming vector to obtain an aggregated vector; if this round of recursion is the last round of recursion, the aggregated vector is used as the output vector of this round of recursion; if this round of recursion is not the last round of recursion, the aggregated vector is average pooled based on the vector dimension, and the obtained average pooled vector is used as the output vector of this round of recursion.

[0155] The above technical solution mainly mines text knowledge vectors of compressed communication messages through a recursive strategy to obtain corresponding text unit knowledge vectors.

[0156] Let's use a specific example to understand this process: suppose there is a communication message to be compressed: "Ship speed 12 knots, 30 degrees north east. Waves 3 meters, wind force 7."

[0157] First, the communication message to be compressed (i.e., the incoming vector of the first round of recursion) is reversibly processed using a pre-configured reversible operator to obtain a reversible knowledge vector. Then, the features of the reversible knowledge vector and the communication message to be compressed are integrated to obtain an aggregated vector.

[0158] Next, in the non-first round of recursion, the reversible operator is used again to perform reversible processing on the output vector of the previous round of recursion (that is, the aggregation vector of the previous round), and the features are continued to be integrated with the original input vector to obtain a new aggregation vector.

[0159] This process will continue until the predetermined number of recursive layers is reached. In the last round of recursion, the final aggregate vector is used as the output vector of this round of recursion, that is, the required text unit knowledge vector is obtained. In the non-last round of recursion, the aggregate vector will be average pooled based on the vector dimension, and the average pooled vector obtained will be used as the input of the next round of recursion.

[0160] In summary, this technical solution uses recursive strategies and reversible operators to perform deep mining and integrate text features, making the obtained text unit knowledge vector more expressive and robust. This not only improves the accuracy of maritime communication message compression, but also enhances the model's adaptability to complex and changeable communication message environments, thereby improving the quality and efficiency of maritime communications.

[0161] In addition, reversible processing can be understood as a special convolution processing, which can retain the main information of the original data while also achieving data dimensionality reduction.

[0162] Take the maritime business communication message as an example, such as "wind force 7, wave 3 meters, ship speed 12 knots, heading 30 degrees north east". First, based on the pre-configured reversible operator, the communication message to be compressed is reversibly processed to obtain a reversible knowledge vector.

[0163] Then, based on this reversible knowledge vector and the input vector (the communication message to be compressed in the first round of recursion, and the output vector of the previous round of recursion in the non-first round of recursion), feature integration is performed to obtain an aggregated vector.

[0164] If the current round of recursion is the last round, the aggregated vector is used as the output vector of the current round of recursion; if it is not the last round, the aggregated vector is average pooled based on the vector dimension, and the obtained average pooled vector is used as the output vector of the current round of recursion. Average pooling helps to further reduce the dimension of the vector, thereby achieving data compression.

[0165] This process will continue until all rounds of recursive processing are completed. Finally, a text knowledge vector that has been deeply mined and processed will be obtained.

[0166] In summary, this technical solution, by using recursive strategies and reversible operators, can effectively perform deep mining and processing on compressed communication messages, improve the quality of text knowledge vectors, and at the same time achieve data compression, further optimizing the efficiency of maritime communications.

[0167] In some optional embodiments, the text compression weights corresponding to each of the maritime business communication text units are determined based on the text feature similarities between each current text unit distribution reinforcement vector and a previously stored past text unit distribution reinforcement vector, including: for each of the current text unit distribution reinforcement vectors, the following steps are implemented: for a current text unit distribution reinforcement vector, a past text unit distribution reinforcement vector having the highest text feature similarity with the current text unit distribution reinforcement vector is determined; based on the difference between the past text unit distribution reinforcement vector and the current text unit distribution reinforcement vector, the text compression weight of the maritime business communication text unit corresponding to the current text unit distribution reinforcement vector is determined.

[0168] In this technical solution, the text compression weight corresponding to each maritime business communication text unit is determined mainly by utilizing the text feature similarity between the current text unit distribution reinforcement vector and the previously stored past text unit distribution reinforcement vector.

[0169] To explain with a specific example: suppose there is a series of current text unit distribution reinforcement vectors, which may come from marine communication messages to be compressed, such as "wind force 7", "waves 3 meters", "ship speed 12 knots", "30 degrees north east", etc.

[0170] First, for each current text unit distribution reinforcement vector (for example, the vector of "wind force 7"), the past text unit distribution reinforcement vector with the highest text feature similarity is found. The "text feature similarity" here can be measured by calculating the cosine similarity or Euclidean distance of two vectors.

[0171] Then, based on the difference between the past text unit distribution reinforcement vector and the current text unit distribution reinforcement vector, the text compression weight of the maritime business communication text unit corresponding to the current text unit distribution reinforcement vector is determined. This "difference" can be determined by comparing the values ​​of the two vectors in different dimensions, or by a more complex machine learning model.

[0172] In summary, this technical solution gives a reasonable compression weight to each maritime business communication text unit by carefully comparing and analyzing the similarities and differences between the current text unit and the past text units, thereby achieving more accurate and efficient message compression. This not only improves the compression quality, but also helps save communication resources, which has important practical value for maritime communications.

[0173] In some preferred embodiments, the text compression weights corresponding to the maritime business communication text units are determined based on the text feature similarities between each current text unit distribution reinforcement vector and the previously stored past text unit distribution reinforcement vectors, including: for each current text unit distribution reinforcement vector, the following steps are implemented: for a current text unit distribution reinforcement vector, determine the past text unit distribution reinforcement vector with the highest text feature similarity to the current text unit distribution reinforcement vector; from the previously stored past text unit distribution reinforcement vectors, determine several past text unit distribution reinforcement vectors whose text feature similarities with the past text unit distribution reinforcement vector with the highest text feature similarity are not greater than a preset text feature similarity threshold; based on the past text unit distribution reinforcement vector with the highest text feature similarity and the differences between each of the several past text unit distribution reinforcement vectors and the current text unit distribution reinforcement vector, determine the text compression weight of the maritime business communication text unit corresponding to the current text unit distribution reinforcement vector.

[0174] In this technical solution, the text compression weight corresponding to each maritime business communication text unit is determined mainly based on the text feature similarity between the current text unit distribution reinforcement vector and the past text unit distribution reinforcement vector.

[0175] For example, suppose there is a set of previously stored past text unit distribution enhancement vectors, such as vector representations of "wind force 7", "waves 3 meters", etc. Now, a new communication message to be compressed is obtained: "Ship speed 12 knots, north by east 30 degrees." Each text unit of this new message (such as "Ship speed 12 knots" and "north by east 30 degrees") is processed to obtain the corresponding current text unit distribution enhancement vector.

[0176] Next, for each current text unit distribution reinforcement vector, the most similar past text unit distribution reinforcement vector will be found in the past text unit distribution reinforcement vector set. Then, from this past vector set, some past text unit distribution reinforcement vectors whose text feature similarity with the current text unit distribution reinforcement vector is not greater than a preset threshold will be screened out.

[0177] Finally, according to the difference between the selected past text unit distribution reinforcement vectors and the current text unit distribution reinforcement vector, the text compression weight of the maritime business communication text unit corresponding to the current text unit distribution reinforcement vector is determined.

[0178] In summary, this technical solution determines a suitable compression weight for each maritime business communication text unit by comparing the similarity of text features between new messages and past messages and considering the differences between them. This ensures that when compressing messages, the main information of the original message can be retained as much as possible, while achieving an ideal compression effect, thereby effectively improving the quality and efficiency of maritime communications.

[0179] Under some other possible design ideas, text knowledge vector mining is performed on each maritime business communication text unit contained in the communication message to be compressed to obtain a corresponding text unit knowledge vector, including: obtaining a corresponding target reference communication message from a plurality of reference communication messages stored in advance according to the target subject keyword of the target maritime communication session; performing text matching on the communication message to be compressed according to the set word vector in the target reference communication message to obtain a text matching result; obtaining a local message set to which the target maritime communication session belongs from the text matching result according to a preset message set in the target reference communication message; and performing text knowledge vector mining on each maritime business communication text unit contained in the local message set to obtain a corresponding text unit knowledge vector.

[0180] In this technical solution, according to the target topic keywords of the target maritime communication session, the target reference communication message is obtained from the pre-stored reference communication message, and then the compressed communication message is text matched according to the set word vector to further obtain the text matching result. Then, by comparing the preset message set and the text matching result, the local message set belonging to the target maritime communication session is determined, and finally the text knowledge vector is mined for the local message set to obtain the corresponding text unit knowledge vector.

[0181] Let's use a specific example to explain: for example, if the subject keyword of the target maritime communication session is "wind power", then firstly all messages related to "wind power" will be found from the pre-stored reference communication messages as the target reference communication messages.

[0182] Next, a set word vector model (such as Word2Vec or BERT, etc.) is used to perform text matching on the compressed communication message to find information related to "wind power" and obtain text matching results.

[0183] Then, based on the preset message set in the target reference communication message, a qualified local message set is screened out from the text matching results. These messages all belong to the target maritime communication session and are related to "wind power".

[0184] Finally, the text knowledge vector mining is performed on each maritime business communication text unit contained in this local message set to obtain the knowledge vector corresponding to each text unit.

[0185] In summary, this technical solution can effectively filter and extract relevant message information by accurately locating and matching target topic keywords, and further convert them into high-dimensional knowledge vectors. This method can not only greatly improve the efficiency and quality of message compression, but also help better understand and analyze the intrinsic semantic structure of maritime communication messages, thereby optimizing the overall communication effect.

[0186] In an embodiment of the present invention, the marine communication data compression technology based on dual-mode satellite fusion mainly utilizes efficient data compression algorithms and methods to reduce the amount of transmitted data, increase the transmission speed, and reduce network congestion to meet the access needs of offshore operating vessels to the Internet and power integrated data network private network services. Examples may involve the following aspects.

[0187] (1) Data compression technology: The goal of this technology is to reduce the amount of data that needs to be transmitted over the satellite link. It can include various types of compression, such as lossless compression and lossy compression. Lossless compression can completely restore the original data after decompression and is suitable for applications that require data integrity. Lossy compression will lose some information, but can provide a higher compression rate and is suitable for the transmission of multimedia data such as images and videos;

[0188] (2) High-efficiency coding technology: Coding technologies such as H.265 / HEVC can achieve high-quality video streaming while maintaining a relatively low bit rate, thereby reducing the bandwidth requirements of satellite links. This technology is particularly suitable for applications such as video surveillance, distance education, and live TV.

[0189] (3) TCP / IP optimization: Since satellite communications have high latency and packet loss rates, the TCP / IP protocol needs to be optimized to improve the efficiency of data transmission. For example, the performance of satellite links can be improved by increasing the TCP window size and using fast retransmission and recovery.

[0190] (4) Network protocol fusion: In dual-mode satellite fusion communications, multiple network protocols (such as TCP, UDP, HTTP, etc.) can be integrated to achieve more efficient data transmission. For example, TCP traffic with a small amount of data can be combined with UDP traffic with a large amount of data and transmitted through the same satellite link, thereby improving bandwidth utilization.

[0191] These technologies are designed to improve the communication efficiency and stability of offshore vessels, meet their business needs such as public network mobile communications, telephone calls, security control and surveillance video backhaul, data collection, etc., and ensure the safety and smooth progress of offshore operations.

[0192] It can be seen that the above technical scheme can perform targeted compression on the communication message data of offshore workers in the scenario of dual-channel application of ChinaSat 6DVAST satellite channel and high-throughput satellite channel based on dual-mode satellite fusion communication technology, thereby improving compression efficiency and ensuring the retention of key information, so as to improve the transmission timeliness of maritime communication data.

[0193] In the actual implementation process, dual-channel operation of the ChinaSat 6D VAST satellite channel and the high-throughput satellite channel can be realized to meet the access needs of submarine cable operating vessels to the Internet and power integrated data network private network services when performing tasks at sea.

[0194] This process requires the use of intermediate frequency distribution technology to solve the connection and signal distribution problems of the uplink and downlink. At the same time, the local oscillator of the low noise block converter (LNB) should be controlled by software to switch channels, which requires the intermediate frequency distributor to receive the local oscillator control signal of the LNB and prevent this control signal from being absorbed by the modems of other ports. In this way, different satellite modems can be controlled by software to work on the corresponding local oscillator of the LNB, thereby realizing time-sharing access of two satellite modems.

[0195] In addition, by using LTE BBU micro base stations combined with high throughput as the backhaul channel, access solutions based on China Telecom and China Unicom communication standards and technologies were studied. At the same time, the use of MESH self-organizing wireless broadband networking and LTE PRRU sub-base station expansion methods to achieve public network fixed-point coverage of multiple ships. In this way, public network communications can be introduced to each submarine cable operation ship to meet the operator's public network mobile phone communications, telephones, and security control and deployment video backhaul, data collection and other business needs, thereby meeting the communication backhaul needs of marine operation safety control.

[0196] On the basis of the above, a maritime communication data compression system is provided, comprising a processor and a memory communicating with each other, wherein the processor is used to retrieve a computer program from the memory and implement the above method by running the computer program.

[0197] Based on the above, a computer-readable storage medium is provided, on which a computer program is stored, and the computer program implements the above method when running.

[0198] The above are only embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for compressing maritime communication data based on dual-mode satellite fusion, characterized in that: Applied to a marine communication data compression system, the method comprises: Acquire a communication message to be compressed containing a target maritime communication session, and perform text knowledge vector mining on each maritime service communication text unit contained in the communication message to be compressed to obtain a corresponding text unit knowledge vector; wherein the communication message to be compressed is a communication message in an application scenario of a dual-mode satellite fusion channel, and the dual-mode satellite fusion channel includes a ChinaSat 6DVAST satellite channel and a high-throughput satellite channel; By completing the optimized paragraph distribution attention model, according to the paragraph distribution labels of each of the maritime business communication text units in the communication message to be compressed, the corresponding text unit knowledge vectors are respectively enhanced in paragraph distribution to obtain the corresponding current text unit distribution enhancement vector containing the paragraph distribution label; Based on the text feature similarity between each current text unit distribution reinforcement vector and the previously stored past text unit distribution reinforcement vector, the text compression weight corresponding to each of the maritime business communication text units is determined respectively; wherein each past text unit distribution reinforcement vector is obtained based on the compressed communication message mining, and each past text unit distribution reinforcement vector includes the paragraph distribution label of the corresponding maritime business communication text unit, and the compressed communication message includes the compressed past maritime communication session; Determining the message compression decision feature of the communication message to be compressed based on the obtained compression weights of each text; Before the optimized paragraph distribution attention model is completed and the corresponding text unit knowledge vectors are respectively subjected to paragraph distribution enhancement according to the paragraph distribution labels of each maritime business communication text unit in the communication message to be compressed, and the corresponding current text unit distribution enhancement vector containing the paragraph distribution label is obtained, the method further includes: Based on a number of communication message training examples and their corresponding prior learning annotations, the paragraph distribution attention model to be optimized is debugged in a loop until the loop completion requirements are met, thereby obtaining the optimized paragraph distribution attention model; wherein each prior learning annotation is used to characterize the possibility that each maritime business communication text unit in the corresponding communication message training example is a text unit to be compressed; During a debugging process, the following steps are performed: For a communication message training example, performing text knowledge vector mining on each maritime service communication text unit of the communication message training example to obtain a corresponding text unit knowledge vector training example; Through the paragraph distribution attention model applied in this round, based on the paragraph distribution labels of each maritime business communication text unit in the one communication message training sample, the corresponding text unit knowledge vector training sample is respectively subjected to paragraph distribution reinforcement, and the corresponding current text unit distribution reinforcement vector containing the paragraph distribution label is obtained; Obtaining a message compression decision feature of the communication message training example according to the current text unit distribution enhancement vectors corresponding to each maritime business communication text unit of the communication message training example; Based on the differences between several obtained message compression decision features and the corresponding prior learning annotations, the model variables of the paragraph distributed attention model applied in this round are improved.

2. The method according to claim 1, characterized in that Before respectively determining the text compression weights corresponding to the respective maritime business communication text units based on the text feature similarities between the respective current text unit distribution reinforcement vectors and the previously stored past text unit distribution reinforcement vectors, the method further includes: Through the second text knowledge vector mining model, text knowledge vector mining is performed on each maritime business communication text unit contained in each compressed communication message to obtain an initial text unit knowledge vector set, wherein the initial text unit knowledge vector set includes initial text unit knowledge vectors corresponding to each maritime business communication text unit of each compressed communication message; According to the initial text unit knowledge vector set, several communication message training examples are respectively annotated to obtain corresponding prior learning annotations, and according to the several communication message training examples and the corresponding prior learning annotations, the communication message disassembly network is cyclically debugged until the cycle completion requirements are met; wherein the communication message disassembly network includes a first text knowledge vector mining model and the paragraph distributed attention model, and the first text knowledge vector mining model and the second text knowledge vector mining model share model variables; Based on the first text knowledge vector mining model in the debugged communication message disassembly network, text knowledge vector mining is performed on each compressed communication message respectively, and paragraph distribution reinforcement is performed on the mined text unit knowledge vector through the paragraph distribution attention model in the debugged communication message disassembly network to obtain a past text unit distribution reinforcement vector set, which includes the previously stored past text unit distribution reinforcement vectors.

3. The method according to claim 2, characterized in that The method of annotating a plurality of communication message training examples according to the initial text unit knowledge vector set to obtain corresponding prior learning annotations includes: for the plurality of communication message training examples, the following steps are implemented: for a communication message training example, using the second text knowledge vector mining model, respectively performing text knowledge vector mining on each maritime business communication text unit in the communication message training example to obtain a corresponding text unit knowledge vector training example; based on the text feature similarity between each obtained text unit knowledge vector training example and the initial text unit knowledge vector set, determining the subordinate learning annotation of the corresponding maritime business communication text unit, each subordinate learning annotation is used to characterize: the possibility that the corresponding maritime business communication text unit is a text unit to be compressed; based on each obtained subordinate learning annotation, obtaining the prior learning annotation of the communication message training example; Among them, based on the first text knowledge vector mining model in the debugged communication message disassembly network, text knowledge vector mining is performed on each compressed communication message respectively, and the paragraph distribution of the mined text unit knowledge vector is enhanced through the paragraph distribution attention model in the debugged communication message disassembly network to obtain a past text unit distribution enhanced vector set, including: according to the first text knowledge vector mining model, text knowledge vector mining is performed on each maritime business communication text unit of each compressed communication message respectively, and the paragraph distribution of the mined text unit knowledge vector is enhanced through the paragraph distribution attention model in the debugged communication message disassembly network to obtain an enhanced knowledge vector set, the enhanced knowledge vector set includes the enhanced knowledge vectors corresponding to each maritime business communication text unit of each compressed communication message; according to the text feature similarity between every two enhanced knowledge vectors in the enhanced knowledge vector set, part of the enhanced knowledge vectors are selected from the enhanced knowledge vector set to obtain the past text unit distribution enhanced vector set; Among them, according to the text feature similarity between every two reinforced knowledge vectors in the reinforced knowledge vector set, some reinforced knowledge vectors are selected from the reinforced knowledge vector set to obtain the past text unit distribution reinforcement vector set, including: selecting at least one reinforced knowledge vector from the reinforced knowledge vector set to adjust the past text unit distribution reinforcement vector set; looping the past text unit distribution reinforcement vector set; wherein, in each loop processing, the following steps are implemented: for each reinforced knowledge vector in the reinforced knowledge vector set, respectively determine the corresponding knowledge vector tuple, each knowledge vector tuple contains a reinforced knowledge vector, and the past text unit distribution reinforcement vector in the past text unit distribution reinforcement vector set with the highest text feature similarity with the one reinforced knowledge vector; for each past text unit distribution reinforcement vector contained in each knowledge vector tuple, add the reinforced knowledge vector with the smallest text feature similarity with each past text unit distribution reinforcement vector to the past text unit distribution reinforcement vector set.

4. The method according to claim 1, characterized in that By completing the optimized paragraph distribution attention model, according to the paragraph distribution labels of each of the maritime business communication text units in the communication message to be compressed, the corresponding text unit knowledge vectors are respectively subjected to paragraph distribution reinforcement, and the corresponding current text unit distribution reinforcement vector containing the paragraph distribution label is obtained, including: Through the optimized paragraph distributed attention model, the following steps are implemented for each maritime business communication text unit: For a maritime business communication text unit, according to a first paragraph distribution label of the maritime business communication text unit in the communication message to be compressed, obtaining a corresponding first paragraph distribution variable; Obtaining a corresponding second paragraph distribution variable according to a second paragraph distribution label of the one maritime business communication text unit in the communication message to be compressed; According to the text unit knowledge vector of the maritime business communication text unit, the first paragraph distribution variable and the second paragraph distribution variable, a corresponding current text unit distribution reinforcement vector is obtained.

5. The method according to claim 1, characterized in that Performing text knowledge vector mining on each maritime service communication text unit contained in the communication message to be compressed to obtain a corresponding text unit knowledge vector includes: Using a recursive strategy, mining text knowledge vectors for the communication message to be compressed; In each recursive round, the following steps are performed: Based on the pre-configured reversible operator, the incoming vector of this round of recursion is reversibly processed to obtain a reversible knowledge vector; wherein, in the first round of recursion, the incoming vector is the communication message to be compressed, and in the non-first round of recursion, the incoming vector is the output vector of the previous round of recursion; Performing feature integration on the reversible knowledge vector and the input vector to obtain an aggregated vector; If this round of recursion is the last round of recursion, the aggregated vector is used as the output vector of this round of recursion; If the current round of recursion is not the last round of recursion, average pooling processing based on the vector dimension is performed on the aggregated vector, and the obtained average pooling vector is used as the output vector of the current round of recursion.

6. The method according to claim 1, characterized in that The determining of the text compression weights corresponding to the respective maritime business communication text units based on the text feature similarities between the respective current text unit distribution reinforcement vectors and the previously stored past text unit distribution reinforcement vectors includes: For each current text unit distribution reinforcement vector, the following steps are implemented: For a current text unit distribution reinforcement vector, determining a past text unit distribution reinforcement vector having the highest text feature similarity to the current text unit distribution reinforcement vector; According to the difference between the past text unit distribution enhancement vector and the current text unit distribution enhancement vector, a text compression weight of the maritime business communication text unit corresponding to the current text unit distribution enhancement vector is determined.

7. The method according to claim 1, characterized in that The determining of the text compression weights corresponding to the respective maritime business communication text units based on the text feature similarities between the respective current text unit distribution reinforcement vectors and the previously stored past text unit distribution reinforcement vectors includes: For each current text unit distribution reinforcement vector, the following steps are implemented: For a current text unit distribution reinforcement vector, determining a past text unit distribution reinforcement vector having the highest text feature similarity to the current text unit distribution reinforcement vector; Determine, from the previously stored past text unit distribution reinforcement vectors, a number of past text unit distribution reinforcement vectors whose text feature similarity with the past text unit distribution reinforcement vector with the highest text feature similarity is not greater than a preset text feature similarity threshold; Based on the past text unit distribution reinforcement vector with the highest text feature similarity and the differences between each of the several past text unit distribution reinforcement vectors and the current text unit distribution reinforcement vector, the text compression weight of the maritime business communication text unit corresponding to the current text unit distribution reinforcement vector is determined.

8. The method according to claim 1, characterized in that Performing text knowledge vector mining on each maritime service communication text unit contained in the communication message to be compressed to obtain a corresponding text unit knowledge vector includes: According to the target subject keyword of the target maritime communication session, a corresponding target reference communication message is obtained from a plurality of reference communication messages stored in advance; Performing text matching on the communication message to be compressed according to the set word vector in the target reference communication message to obtain a text matching result; According to a preset message set in the target reference communication message, a local message set to which the target maritime communication session belongs is obtained from the text matching result; Text knowledge vector mining is performed on each maritime business communication text unit contained in the local message set to obtain a corresponding text unit knowledge vector.

9. A marine communication data compression system, characterized in that: The maritime communication data compression system includes a processor and a memory that communicate with each other, and the processor is used to call a computer program from the memory and implement the method described in any one of claims 1-8 by running the computer program.

Images