Non-destructive data processing method based on deep-sea profiling buoy

Abstract

The present invention provides a lossless data processing method based on a deep-sea profile buoy, which can be applied to the field of data processing technology. The lossless data processing method based on a deep-sea profile buoy includes: performing word vectorization processing on a profile data sequence to obtain a word embedding vector sequence; repeatedly performing the following probability prediction operations until the prediction probability of the last profile data in the profile data sequence is obtained: inputting the first k word embedding vectors into a preliminary probability prediction model to obtain a preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector; inputting the preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector and the first k word embedding vectors into the probability prediction model to obtain a prediction probability of the profile data corresponding to the k+1th word embedding vector; based on the prediction probability sequence of the profile data sequence, adaptively encoding the profile data sequence to obtain a compressed sequence.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, in particular to lossless data processing technology, and more particularly to a lossless data processing method based on a deep-sea profiling buoy. Background Art

[0002] Ocean observation technology plays an important role in the development and exploration of the ocean. With the development of technology, there are more and more means of ocean monitoring. Deep-sea profiling buoys are an important means of ocean monitoring. They carry a variety of sensors to measure a variety of ocean profile data during the profile movement, playing an irreplaceable role in ocean monitoring.

[0003] Since the storage capacity of deep-sea profiling buoy data is limited, and the satellite communication method for transmitting deep-sea profiling buoy data is charged by data volume, it is necessary to use data compression algorithms to compress deep-sea profiling buoy data according to the data characteristics of deep-sea profiling buoys to reduce communication costs and improve data transmission efficiency. In addition, in order to ensure the accuracy of data compression, lossless data compression algorithms are used to compress deep-sea profiling buoy data.

[0004] Lossless data compression algorithm refers to the premise that the data is not distorted, and the redundant information of the data is extracted, and it is ensured that after the data is decompressed, the reconstructed data is exactly the same as the original data. Therefore, compared with the lossy compression algorithm, the compression ratio of lossless compression will be correspondingly lower.

[0005] At present, data compression based on deep-sea profiling buoys is mostly concentrated on traditional compression algorithms, such as run-length coding compression algorithm, Huffman coding compression algorithm, dictionary-based lossless data compression algorithm, and sliding window-based lossless data compression algorithm. Although traditional compression algorithms are relatively simple, their compression efficiency is low.

[0006] With the development of machine learning and deep learning, probabilistic prediction models and algorithms have developed rapidly. Since the essence of lossless data compression is a probabilistic statistical model, the application of machine learning and deep learning in data compression has achieved good results. However, due to the fact that the gradients of models such as recurrent neural networks, gated recurrent unit algorithms, and long short-term memory networks are easy to disappear or explode, the calculation is time-consuming, the fitting speed is slow, and the dependencies of long links cannot be captured. In addition, the traditional transformer network model is too complex and difficult to transplant, the model parameters are too many and sensitive, the training cycle is long, the performance is excessive, the hardware performance requirements are high, and too much compression ratio needs to be sacrificed during training. Therefore, when using traditional machine learning and deep learning models to compress deep-sea profile buoy data, there are problems such as the inability to capture the dependencies of long links, lack of contextual relevance, and large and complex models. Summary of the invention

[0007] In view of the above problems, the present invention provides a non-destructive data processing method and device based on deep-sea profiling buoy.

[0008] According to a first aspect of the present invention, there is provided a lossless data processing method based on a deep-sea profiling buoy, comprising: performing word vectorization processing on a profiling data sequence to obtain a word embedding vector sequence, wherein the profiling data sequence is obtained using a deep-sea profiling buoy and is used to characterize the state of the ocean at different time periods; repeatedly performing the following probability prediction operation until the predicted probability of the last profiling data in the profiling data sequence is obtained, wherein the predicted probability includes the probability that the last profiling data corresponds to a target character in a dictionary, wherein the target character is determined by using the respective probabilities of a plurality of characters in the dictionary for characterizing the last profiling data; inputting the first k word embedding vectors into the word embedding vector sequence; Input into the preliminary probability prediction model to obtain the preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector, wherein the above preliminary probability prediction model is generated based on multi-head self-attention; input the preliminary prediction probability of the profile data corresponding to the above k+1th word embedding vector and the above first k word embedding vectors into the probability prediction model to obtain the prediction probability of the profile data corresponding to the above k+1th word embedding vector, wherein the above probability prediction model is generated based on a multi-layer perceptron, k is greater than 1 and less than K, and K is the number of word embedding vectors in the word embedding vector sequence; based on the prediction probability sequence of the above profile data sequence, the above profile data sequence is adaptively encoded to obtain a compressed sequence.

[0009] According to an embodiment of the present invention, the above method also includes: taking a profile data subsequence in the above profile data sequence as a training sample, and determining the profile data set after the above training sample from the above profile data sequence as a label; inputting multiple sample word embedding vectors corresponding to the above profile data subsequence into the above preliminary probability prediction model to obtain the sample preliminary prediction probability for the above label; inputting the multiple sample word embedding vectors and the above sample preliminary prediction probability into the above probability prediction model to obtain the sample prediction probability; and adjusting the parameters in the above probability prediction model based on the above sample prediction probability and the above label.

[0010] According to an embodiment of the present invention, the preliminary probability prediction model includes: a multi-head attention layer, an intermediate processing layer, a first linear transformation branch, a second linear transformation branch, and a preliminary probability output layer. The method of inputting the first k word embedding vectors into the preliminary probability prediction model to obtain the preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector includes: inputting the first k word embedding vectors into the multi-head attention layer to obtain the attention output result; inputting the attention output result into the intermediate processing layer to obtain the intermediate processing result; inputting the intermediate processing result into the first linear transformation branch and the second linear transformation branch respectively to obtain the first branch transformation result and the second branch transformation result; and inputting the first branch transformation result and the second branch transformation result into the preliminary probability output layer to obtain the preliminary prediction probability.

[0011] According to an embodiment of the present invention, the above-mentioned first linear transformation branch includes a first linear transformation layer; the above-mentioned second linear transformation branch includes in sequence: a second linear transformation layer, an activation layer and a third linear transformation layer; the above-mentioned intermediate processing layer includes in sequence: a dimension flipping unit, a slicing unit and a flattening unit.

[0012] According to an embodiment of the present invention, the above-mentioned probability prediction model includes a first splicing layer, a first probability prediction branch, a second probability prediction branch, a third probability prediction branch, a second splicing layer, and a probability prediction layer; the above-mentioned preliminary prediction probability of the profile data corresponding to the above-mentioned k+1th word embedding vector and the above-mentioned first k word embedding vectors are input into the probability prediction model to obtain the prediction probability of the profile data corresponding to the above-mentioned k+1th word embedding vector, including: inputting the above-mentioned preliminary prediction probability and the above-mentioned first k word embedding vectors into the above-mentioned first splicing layer to obtain a first splicing result; inputting the above-mentioned first splicing result into the above-mentioned first probability prediction branch, the above-mentioned second probability prediction branch and the third probability prediction branch respectively to obtain a first probability prediction result, a second probability prediction result and a third probability prediction result; inputting the above-mentioned first probability prediction result, the above-mentioned second probability prediction result and the above-mentioned third probability prediction result into the second splicing layer to obtain a second splicing result; inputting the above-mentioned second splicing result into the above-mentioned probability prediction layer to obtain the above-mentioned prediction probability of the profile data corresponding to the above-mentioned k+1th word embedding vector.

[0013] According to an embodiment of the present invention, the first probability prediction branch sequentially includes a first fully connected layer, a first fully connected module, a second fully connected module and a fourth linear transformation layer.

[0014] According to an embodiment of the present invention, the above-mentioned first splicing result is input into the above-mentioned first probability prediction branch to obtain the first probability prediction result, including: inputting the above-mentioned first splicing result into the above-mentioned first fully connected layer to obtain the first fully connected output result; inputting the above-mentioned first fully connected output result into the second fully connected layer of the above-mentioned first fully connected module to obtain the second fully connected output result; inputting the above-mentioned second fully connected output result and the above-mentioned first fully connected output result into the first addition layer of the above-mentioned first fully connected module to obtain the first addition result; inputting the above-mentioned first addition result into the third fully connected layer of the above-mentioned second fully connected module to obtain the third fully connected output result; inputting the above-mentioned third fully connected output result and the above-mentioned second fully connected output result into the second addition layer of the above-mentioned second fully connected module to obtain the second addition result; inputting the above-mentioned second addition result into the above-mentioned fourth linear transformation layer to obtain the above-mentioned first probability prediction result.

[0015] According to an embodiment of the present invention, the second probability prediction branch includes a fifth linear layer; and the third probability prediction branch includes a fourth fully connected layer and a sixth linear layer in sequence.

[0016] According to an embodiment of the present invention, the above method further comprises: dividing the profile data sequence set according to a predetermined step length to obtain a plurality of the above profile data sequences.

[0017] According to an embodiment of the present invention, the above-mentioned word vectorization processing of the profile data sequence to obtain the word embedding vector sequence includes: based on a dictionary, performing character conversion on the initial profile data sequence to obtain the profile data sequence in the above-mentioned predetermined format; and performing word vectorization processing on the above-mentioned profile data sequence using the embedding matrix of the embedding layer to obtain the above-mentioned word embedding vector sequence.

[0018] The second aspect of the present invention provides a lossless data processing device based on a deep-sea profile buoy, comprising: an embedding module, used to perform word vectorization processing on a profile data sequence to obtain a word embedding vector sequence, wherein the profile data sequence is obtained using a deep-sea profile buoy and is used to characterize the state of the ocean at different time periods; a probability prediction module, used to repeatedly perform probability prediction operations until the predicted probability of the last profile data in the profile data sequence is obtained, wherein the predicted probability includes the probability that the last profile data corresponds to a target character in a dictionary, wherein the target character is determined using the probabilities of each of the multiple characters in the dictionary for characterizing the last profile data; wherein the probability prediction module includes a preliminary probability prediction submodule and a probability prediction submodule, the preliminary probability prediction The detection submodule is used to input the first k word embedding vectors into the preliminary probability prediction model to obtain the preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector, wherein the above preliminary probability prediction model is generated based on multi-head self-attention; the probability prediction submodule is used to input the preliminary prediction probability of the profile data corresponding to the above k+1th word embedding vector and the above first k word embedding vectors into the probability prediction model to obtain the prediction probability of the profile data corresponding to the above k+1th word embedding vector, wherein the above probability prediction model is generated based on a multi-layer perceptron, k is greater than 1 and less than K, and K is the number of word embedding vectors in the word embedding vector sequence; and an adaptive coding module is used to adaptively encode the above profile data sequence based on the prediction probability sequence of the above profile data sequence to obtain a compressed sequence.

[0019] The non-destructive data processing method based on deep-sea profiling buoy provided in this application includes at least the following technical effects:

[0020] By using a multi-head self-attention layer to build a preliminary probability prediction model, focusing on contextual relevance, it is possible to learn multiple types of relationships and features in parallel, improving the generalization ability of the probability prediction model and its ability to capture complex patterns. Different attention heads can focus on different attention patterns, increasing the flexibility and diversity of the preliminary probability prediction model. In addition, based on the preliminary prediction probability obtained using the preliminary probability prediction model, the probability prediction model is used to further determine the prediction probability, thereby improving the accuracy of the prediction probability.

[0021] Therefore, compared with traditional compression algorithms, the lossless data processing method based on deep-sea profiling buoys of the present application at least partially overcomes the technical problems such as the large storage space occupied by long-term observation data of deep-sea profiling buoys, long satellite communication time and high cost, the inability to capture long link dependencies, lack of contextual relevance, and large and complex models, thereby achieving the technical effects of low compression rate, fast speed, strong targeting and short model training cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The above contents and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings.

[0023] Figure 1 A diagram showing an application scenario of a non-destructive data processing method based on a deep-sea profiling buoy according to an embodiment of the present invention.

[0024] Figure 2 A flow chart of a non-destructive data processing method based on a deep-sea profiling buoy according to an embodiment of the present invention is shown.

[0025] Figure 3 A schematic diagram of a non-destructive data processing method based on a deep-sea profiling buoy according to an embodiment of the present invention is shown.

[0026] Figure 4 A schematic structural diagram of a preliminary probability prediction model according to a specific embodiment of the present invention is shown.

[0027] Figure 5 A schematic diagram of the structure of a probability prediction model according to a specific embodiment of the present invention is shown.

[0028] Figure 6 A schematic diagram of a compression process according to a specific embodiment of the present invention is shown.

[0029] Figure 7 A schematic diagram of a decompression process according to a specific embodiment of the present invention is shown.

[0030] Figure 8 A structural block diagram of a non-destructive data processing device based on a deep-sea profiling buoy according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0031] Below, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of concepts of the present invention.

[0032] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0033] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0034] When using expressions such as "at least one of A, B, and C, etc.", they should generally be interpreted according to the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0035] When using machine learning and deep learning to compress data from deep-sea profiling buoys, traditional machine learning and deep learning models such as recurrent neural networks, gated recurrent unit algorithms, and long short-term memory networks have certain defects, such as gradients that are prone to disappearance or explosion, time-consuming calculations, slow fitting speeds, and inability to capture long-link dependencies. Although traditional transformer neural networks have strong representation and parallel computing capabilities, they are too complex and difficult to transplant, have too many and sensitive model parameters, have long training cycles, have excessive performance, have high hardware performance requirements, and sacrifice too much compression ratio during the training process.

[0036] An embodiment of the present invention provides a lossless data processing method based on a deep-sea profiling buoy, comprising: performing word vectorization processing on a profile data sequence to obtain a word embedding vector sequence, wherein the profile data sequence is obtained using a deep-sea profiling buoy and is used to characterize the state of the ocean at different time periods; repeatedly performing the following probability prediction operations until the predicted probability of the last profile data in the profile data sequence is obtained, wherein the predicted probability includes the probability that the last profile data corresponds to a target character in a dictionary, wherein the target character is determined by using the probabilities of each of a plurality of characters in the dictionary for characterizing the last profile data: inputting the first k word embedding vectors into In the preliminary probability prediction model, the preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector is obtained, wherein the preliminary probability prediction model is generated based on multi-head self-attention; the preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector and the first k word embedding vectors are input into the probability prediction model to obtain the prediction probability of the profile data corresponding to the k+1th word embedding vector, wherein the probability prediction model is generated based on a multi-layer perceptron, k is greater than 1 and less than K, and K is the number of word embedding vectors in the word embedding vector sequence; based on the prediction probability sequence of the profile data sequence, the profile data sequence is adaptively encoded to obtain a compressed sequence.

[0037] According to an embodiment of the present invention, a probability prediction model based on a multi-layer perceptron combined with a multi-head self-attention is used to perform probability prediction on deep-sea profile data, and a block compression and training-while-compression method is introduced, and adaptive arithmetic coding is combined to achieve lossless compression of deep-sea profile data. The probability prediction model is mainly constructed using a multi-head self-attention layer, pays more attention to context relevance, and learns multiple types of relationships and features in parallel, which improves the generalization ability of the model and the ability to capture complex patterns. During compression, a multi-layer perceptron is introduced to achieve training-while-compression, which can dynamically update the model during the compression process so that it can gradually adapt to the distribution of data, thereby providing better probability prediction and improving compression efficiency. During compression, a predetermined step size for determining the model update frequency and the length of the encoding symbol block is also introduced to achieve block compression, which can achieve a balance between the update frequency and computational efficiency of the probability prediction model. Therefore, by adopting the above-mentioned technical means, the lossless data processing method based on the deep-sea profile buoy of the present application makes up for the problems of low compression ratio, slow fitting speed, and inability to capture the dependence of long links in other lossless data processing methods, and can achieve a high compression ratio of the deep-sea profile buoy.

[0038] Figure 1 A diagram showing an application scenario of a non-destructive data processing method based on a deep-sea profiling buoy according to an embodiment of the present invention.

[0039] like Figure 1 As shown, the application scenario 100 according to this embodiment may include a detection device 101, a terminal device 102, and a network 103. The network 103 is used to provide a medium for a communication link between the detection device 101 and the terminal device 102. The network 103 may include various connection types, such as wired and / or wireless communication links, etc., wherein the wireless communication link may include a satellite communication link. A user may use the terminal device 102 to interact with the detection device 101 through the network 103 to receive deep-sea profiling buoy data, etc.

[0040] The detection device 101 may be a system device with underwater detection function, such as a buoy. A variety of sensors may be installed on the detection device 101 to collect deep-sea profile buoy data, including underwater temperature, density, depth, etc. The terminal device 102 may be various electronic devices with a display screen, including but not limited to a desktop computer, etc.

[0041] It should be noted that the non-destructive data processing method based on deep-sea profiling buoy provided in the embodiment of the present application can generally be executed by the main control chip of the detection device 101.

[0042] For example, the profile data can be collected by the sensor of the detection device 101, and then the main control chip in the detection device 101 executes the lossless data processing method based on the deep-sea profile buoy provided in the embodiment of the present application to compress the profile data to obtain compressed data. After obtaining the compressed data, the detection device 101 sends the compressed data to the terminal device 102 through the network 103. In response to receiving the compressed data, the terminal device 102 decompresses the compressed data to obtain the original profile data, so that the user can view and use the profile data through the terminal device 102.

[0043] It should be understood that Figure 1 The number of detection devices, terminal devices and networks in the embodiment is only for illustration. Any number of detection devices, terminal devices and networks may be provided according to implementation requirements.

[0044] The following will be based on Figure 1 The scene described by Figure 2~Figure 6 The non-destructive data processing method based on deep-sea profiling buoy according to an embodiment of the invention is described in detail.

[0045] Figure 2 A flow chart of a non-destructive data processing method based on a deep-sea profiling buoy according to an embodiment of the present invention is shown.

[0046] like Figure 2 As shown, the non-destructive data processing method based on deep-sea profiling buoy of this embodiment includes operations S210 to S230.

[0047] In operation S210, word vectorization is performed on the profile data sequence to obtain a word embedding vector sequence.

[0048] According to an embodiment of the present invention, the profile data sequence is obtained by using a deep-sea profile buoy and is used to characterize the state of the ocean at different time periods. The word embedding vector sequence is obtained by word vectorization of the profile data sequence, wherein the profile data in the profile data sequence corresponds one-to-one to the word embedding vector in the word embedding vector sequence. Specifically, the profile data sequence can be converted into a word embedding vector by a word embedding layer.

[0049] According to an embodiment of the present invention, by converting a discrete sequence of profile data into a corresponding sequence of word embedding vectors, the semantic relationship between the profile data can be effectively captured, and discrete text data can be converted into a continuous vector that can be input into a neural network.

[0050] In operation S220, the probability prediction operation is repeatedly performed until the predicted probability of the last profile data in the profile data sequence is obtained.

[0051] According to an embodiment of the present invention, the predicted probability includes the probability that the last profile data corresponds to a target character in the dictionary. The target character is determined by using the probabilities of each of the multiple characters in the dictionary for representing the last profile data.

[0052] According to an embodiment of the present invention, a dictionary is obtained based on multiple different characters included in a profile data sequence. Specifically, the profile data can be read character by character to obtain all the characters included in the profile data, and character identifiers are configured for the characters respectively. The same characters are configured with the same character identifiers, and different characters are configured with different character identifiers. The more character identifiers there are, the more characters are included in the corresponding dictionary.

[0053] According to an embodiment of the present invention, the dictionary may include a first dictionary and a second dictionary, the first dictionary is used to represent the one-way correspondence between characters and character identifiers, and the second dictionary is used to represent the one-way correspondence between character identifiers and characters.

[0054] According to an embodiment of the present invention, the probability prediction operation includes operations S221 to S222.

[0055] In operation S221, the first k word embedding vectors are input into a preliminary probability prediction model to obtain a preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector.

[0056] According to an embodiment of the present invention, the preliminary probability prediction model is generated based on multi-head self-attention and is obtained through pre-training. Therefore, the preliminary probability prediction model can capture the complex dependencies between word embedding vectors, especially long-distance dependencies. Different heads can focus on different attention modes, which increases the flexibility and diversity of the preliminary probability prediction model.

[0057] According to an embodiment of the present invention, k is greater than 1 and less than K, where K is the number of word embedding vectors in the word embedding vector sequence.

[0058] In operation S222, the preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector and the first k word embedding vectors are input into the probability prediction model to obtain the prediction probability of the profile data corresponding to the k+1th word embedding vector.

[0059] According to an embodiment of the present invention, the probability prediction model is generated based on a multi-layer perceptron and is obtained through pre-training.

[0060] In operation S230, the profile data sequence is adaptively encoded based on the predicted probability sequence of the profile data sequence to obtain a compressed sequence.

[0061] According to an embodiment of the present invention, the adaptive coding may be adaptive arithmetic coding. When the profile data sequence is adaptively coded, the cumulative frequency table may be updated by the predicted probability sequence, and the profile data sequence may be adaptively coded according to the cumulative frequency table to obtain a compressed sequence, wherein the frequency in the cumulative frequency table may be used to characterize the frequency of occurrence of characters.

[0062] According to an embodiment of the present invention, since no probability prediction is performed on the initial k word embedding vectors in the profile data sequence, the initial k word embedding vectors have no corresponding prediction probabilities. When the initial k word embedding vectors are adaptively encoded, a uniform probability distribution can be adopted for the initial k word embedding vectors.

[0063] According to an embodiment of the present invention, by using a multi-head self-attention layer to build a preliminary probability prediction model, focusing on context relevance, it is possible to learn multiple types of relationships and features in parallel, thereby improving the generalization ability of the probability prediction model and the ability to capture complex patterns. Different attention heads can focus on different attention patterns, increasing the flexibility and diversity of the preliminary probability prediction model. In addition, on the basis of obtaining the preliminary prediction probability using the preliminary probability prediction model, the probability prediction model is used to further determine the prediction probability, thereby improving the accuracy of the prediction probability.

[0064] Figure 3 A schematic diagram of a non-destructive data processing method based on a deep-sea profiling buoy according to an embodiment of the present invention is shown.

[0065] like Figure 3 As shown, the preliminary probability prediction model 310 is obtained by pre-training the historical profile data 10. The probability prediction model 320 is obtained based on a multi-layer perceptron, and the number of branches, the number of fully connected layers, and the number of linear layers of the probability prediction model are not specifically limited.

[0066] According to an embodiment of the present invention, the profile data sequence 20 is subjected to the preliminary probability prediction model 310 and the probability prediction model 320 to obtain a prediction probability sequence. The probability prediction sequence is adaptively encoded to obtain a compressed sequence 30, and the probability prediction sequence is used to adjust the parameters of the probability prediction model 320 to achieve compression and training.

[0067] According to an embodiment of the present invention, the above-mentioned lossless data processing method based on deep-sea profiling buoy also includes: taking a profile data subsequence in a profile data sequence as a training sample, and determining the profile data set after the training sample from the profile data sequence as a label; inputting multiple sample word embedding vectors corresponding to the profile data subsequence into a preliminary probability prediction model to obtain a sample preliminary prediction probability for the label; inputting multiple sample word embedding vectors and the sample preliminary prediction probability into a probability prediction model to obtain a sample prediction probability; and adjusting the parameters in the probability prediction model based on the sample prediction probability and the label.

[0068] According to an embodiment of the present invention, the profile data sequence can be divided into multiple profile data subsequences through window sliding and time step, and training samples can be generated based on the profile data subsequences. Specifically, a sample identifier can be configured for each profile data in the profile data subsequence to obtain a sample identifier subsequence, and training samples can be obtained based on the sample identifier subsequence and the profile data subsequence, wherein the training samples can be two-dimensional arrays.

[0069] According to an embodiment of the present invention, after adjusting the parameters of the probability prediction model based on the sample prediction probability and label corresponding to the kth profile data subsequence, a probability prediction operation is performed on the k+1th profile data subsequence based on the probability prediction model after the parameter adjustment, thereby realizing dynamic updating of the probability prediction model during the compression process, so that the probability prediction model can gradually adapt to the distribution of the data.

[0070] According to an embodiment of the present invention, the profile data sequence is taken as a unit, and the profile data sequence is compressed and the probability prediction model is trained at the same time. Specifically, the probability prediction operation is repeatedly performed on the profile data subsequence until the prediction probability of the profile data after the profile data subsequence, that is, the sample prediction probability, is obtained, and the parameters of the probability prediction model are adjusted using the sample prediction probability and the label to realize the training of the probability prediction model; at the same time, when the prediction probability of the last profile data in the profile data sequence is obtained, a compressed sequence is obtained based on the prediction probability sequence of the profile data sequence.

[0071] According to an embodiment of the present invention, a multilayer perceptron is introduced during compression to achieve compression while training, which can not only capture the complex patterns and features of the input data and improve the performance of the model, but also dynamically update the probability prediction model during the compression process so that it can gradually adapt to the distribution of the data, thereby improving the accuracy of the prediction probability and improving the compression efficiency.

[0072] According to an embodiment of the present invention, before performing word vectorization processing on the profile data sequence, it also includes: dividing the profile data sequence set according to a predetermined step length to obtain multiple profile data sequences.

[0073] According to an embodiment of the present invention, a predetermined step size can be used to characterize the length of a profile data sequence. A profile data sequence set can be divided into multiple profile data sequences according to the predetermined step size. The above-mentioned lossless data processing method can be executed on each profile data sequence respectively, and block compression and prediction probability model training can be performed on the profile data sequence as a unit.

[0074] According to an embodiment of the present invention, when the total length of the profile data sequence set is determined, the smaller the predetermined step size, the greater the number of profile data sequences. Since the training of the prediction probability model is based on the profile data sequence, the smaller the predetermined step size, the more times the prediction model is trained, and the better the compression effect. However, the increase in the number of training times leads to an increase in the computational cost, so it is necessary to determine a suitable predetermined step size so as to maintain a balance between the probability prediction model update frequency and the computational efficiency.

[0075] According to an embodiment of the present invention, for a profile data sequence that is less than a preset step length, probability prediction operations and adaptive encoding are also performed to ensure compression of all profile data.

[0076] According to an embodiment of the present invention, a predetermined step size is introduced during compression to determine the probability prediction model update frequency and the length of the profile data sequence, thereby achieving block compression and achieving a balance between the probability prediction model update frequency and computational efficiency.

[0077] According to an embodiment of the present invention, word vectorization is performed on a profile data sequence to obtain a word embedding vector sequence, including: based on a dictionary, character conversion is performed on an initial profile data sequence to obtain a profile data sequence in a predetermined format; and word vectorization is performed on the profile data sequence using an embedding matrix of an embedding layer to obtain a word embedding vector sequence.

[0078] According to an embodiment of the present invention, the file format corresponding to the initial profile data sequence may be a text file format (txt), and the file format corresponding to the profile data sequence may be a binary file format (npy) for storing arrays.

[0079] According to an embodiment of the present invention, when performing character conversion on the initial profile data sequence, the characters in the initial profile data sequence can be converted into integer profile data character by character through a constructed dictionary, thereby obtaining a profile data sequence for subsequent reading and training.

[0080] According to an embodiment of the present invention, the profile data sequence is word vectorized using the embedding matrix of the embedding layer, including: multiplying the profile data sequence by the embedding matrix to obtain a word embedding vector sequence, wherein the size of the embedding matrix is ​​related to the alphabet size and the embedding dimension, the rows of the embedding matrix correspond one-to-one to the characters in the alphabet, and the embedding dimension is determined based on actual conditions.

[0081] According to an embodiment of the present invention, by converting the initial profile data sequence into a corresponding word embedding vector sequence, the semantic relationship between the data can be effectively captured, and the discrete text data can be converted into a continuous vector that can be input into a neural network.

[0082] According to an embodiment of the present invention, the first k word embedding vectors are input into a preliminary probability prediction model to obtain a preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector, including: inputting the first k word embedding vectors into a multi-head attention layer to obtain an attention output result; inputting the attention output result into an intermediate processing layer to obtain an intermediate processing result; inputting the intermediate processing result into a first linear transformation branch and a second linear transformation branch respectively to obtain a first branch transformation result and a second branch transformation result; and inputting the first branch transformation result and the second branch transformation result into a preliminary probability output layer to obtain a preliminary prediction probability.

[0083] According to an embodiment of the present invention, the preliminary probability prediction model includes: a multi-head attention layer, an intermediate processing layer, a first linear transformation branch, a second linear transformation branch, and a preliminary probability output layer.

[0084] According to an embodiment of the present invention, a multi-head attention layer may include a linear transformation unit, an attention calculation unit, and an attention splicing unit. The first k word embedding vectors are input into the multi-head attention layer, including: through the linear transformation unit, each word embedding vector is projected into a query matrix, a key matrix, and a value matrix through three independent linear transformations, and the query matrix, the key matrix, and the value matrix include a predetermined number of attention heads; through the attention calculation unit, based on the similarity between the query vector and the key vector corresponding to each attention head, the scaled dot product attention mechanism is used to obtain the attention weight of the attention head, and the attention weight and the value vector corresponding to the attention head are weighted and summed to obtain the attention result of the attention head; through the attention splicing unit, the attention results of multiple attention heads are spliced, and the spliced ​​attention results are linearly transformed to obtain the attention output result.

[0085] According to an embodiment of the present invention, the intermediate processing layer can be used to process the attention output results, so that the preliminary probability prediction model can understand the profile data sequence from different directions and scales, and improve the ability to capture complex patterns.

[0086] According to an embodiment of the present invention, the intermediate processing result is input into the first linear transformation branch to obtain the first branch transformation result, and the intermediate processing result is input into the second linear transformation branch to obtain the second branch transformation result.

[0087] According to an embodiment of the present invention, the preliminary probability output layer may include a preliminary splicing unit and a preliminary activation unit. Inputting the first branch transformation result and the second branch transformation result into the preliminary probability output layer includes: splicing the first branch transformation result and the second branch transformation result through the preliminary splicing unit to obtain a preliminary splicing result; and converting the preliminary splicing result into a preliminary prediction probability through the preliminary activation unit.

[0088] According to an embodiment of the present invention, the preliminary activation unit can be used to convert the probability prediction task into a classification task. Specifically, an activation function such as the log_softmax function is used to convert the preliminary splicing result into a preliminary prediction probability, wherein the preliminary prediction probability is a logarithmic probability distribution.

[0089] According to an embodiment of the present invention, the profile data sequence is predicted by a preliminary probability prediction model constructed based on multi-head attention, so that the preliminary probability prediction model can capture long-distance dependencies in the profile data sequence.

[0090] According to an embodiment of the present invention, the intermediate processing layer sequentially includes a dimension flipping unit, a slicing unit and a flattening unit.

[0091] According to an embodiment of the present invention, the dimension inversion unit is used to flip and slice the attention output results, thereby generating different input views to improve the generalization ability of the preliminary probability prediction model and reduce overfitting, and help the preliminary probability prediction model capture information at different positions in the attention output results.

[0092] According to an embodiment of the present invention, the slicing unit can obtain information from different time steps or positions, and can capture dependencies at different scales.

[0093] According to an embodiment of the present invention, the flattening unit can be used to convert a high-dimensional tensor into a two-dimensional tensor, which can introduce more diverse feature representations, enhance the expressive power of the model, and help the subsequent fully connected layer or classifier to further process the attention output results.

[0094] According to an embodiment of the present invention, the first linear transformation branch includes a first linear transformation layer, and the first linear transformation layer can perform linear transformation processing on the intermediate processing result to obtain a first branch transformation result. Specifically, the calculation process of the first linear transformation branch can be shown in the following formula (1):

[0095] y1=xA1 T +b1(1)

[0096] Among them, x is the intermediate processing result, y1 is the first branch transformation result, A1 is the weight matrix of the first linear transformation layer, and b1 is the bias parameter of the first linear transformation layer.

[0097] According to an embodiment of the present invention, the second linear transformation branch includes, in sequence: a second linear transformation layer, an activation layer, and a third linear transformation layer. The second linear transformation layer is used to map the intermediate processing result to a new space through linear transformation to obtain a first sub-transformation result. The activation layer is used to perform nonlinear mapping on the first sub-transformation result to obtain a second sub-transformation result. Specifically, the activation layer can use an activation function such as a GELU activation function. Compared with the commonly used ReLU activation function, the GELU activation function is smoother when processing values ​​less than zero, which can improve model performance. The third linear transformation layer is used to perform linear transformation processing on the second sub-transformation result to obtain a second branch transformation result. Specifically, the calculation process of the second linear transformation branch can be shown in the following formulas (2) to (3):

[0098] y2=GELU(xA2 T +b2) (2)

[0099] y3=y2A3 T +b3(3)

[0100] Among them, x is the intermediate processing result, y2 is the second sub-transformation result, GELU() represents the GELU activation function, A2 is the weight matrix of the second linear transformation layer, b2 is the bias parameter of the second linear transformation layer, y3 is the second branch transformation result, A3 is the weight matrix of the third linear transformation layer, and b3 is the bias parameter of the third linear transformation layer.

[0101] According to an embodiment of the present invention, by constructing two linear transformation branches, the accuracy of probability prediction performed by the preliminary prediction probability model is improved.

[0102] Figure 4 A schematic structural diagram of a preliminary probability prediction model according to a specific embodiment of the present invention is shown.

[0103] like Figure 4 As shown, the preliminary probability prediction model includes a multi-head attention layer 401, an intermediate processing layer 402, a first linear transformation layer 403, a second linear transformation layer 404, an activation layer 405, a third linear transformation layer 406, a preliminary splicing unit 407 and a preliminary activation unit 408. After the profile data sequence 20 is input into the embedding layer 330, a word embedding vector sequence is obtained, and after the word embedding vector sequence is input into the preliminary probability prediction model, a preliminary prediction probability 40 is output.

[0104] like Figure 4 As shown, the preliminary probability prediction model includes a first linear transformation branch and a second transformation branch, the first linear transformation branch includes a first linear transformation layer 403, and the second linear transformation branch includes a second linear transformation layer 404, an activation layer 405 and a third linear transformation layer 406.

[0105] According to an embodiment of the present invention, a preliminary prediction probability of profile data corresponding to the k+1th word embedding vector and the first k word embedding vectors are input into a probability prediction model to obtain a prediction probability of the profile data corresponding to the k+1th word embedding vector, including: inputting the preliminary prediction probability and the first k word embedding vectors into a first splicing layer to obtain a first splicing result; inputting the first splicing result into a first probability prediction branch, a second probability prediction branch and a third probability prediction branch respectively to obtain a first probability prediction result, a second probability prediction result and a third probability prediction result; inputting the first probability prediction result, the second probability prediction result and the third probability prediction result into a second splicing layer to obtain a second splicing result; inputting the second splicing result into the probability prediction layer to obtain a prediction probability of the profile data corresponding to the k+1th word embedding vector.

[0106] According to an embodiment of the present invention, the probability prediction model includes a first splicing layer, a first probability prediction branch, a second probability prediction branch, a third probability prediction branch, a second splicing layer, and a probability prediction layer.

[0107] According to an embodiment of the present invention, the preliminary prediction probability and the first k word embedding vectors are input into the first splicing layer, including: converting the preliminary prediction probability into an embedding vector, performing a dimension transformation on the embedding vector and splicing it with the first k word embedding vectors to obtain a first splicing result.

[0108] According to an embodiment of the present invention, by converting the preliminary prediction probability into an embedding vector, the semantic relationship between the data can be effectively captured, and the discrete text data can be converted into a continuous vector that can be input into a neural network.

[0109] According to an embodiment of the present invention, the first splicing result is input into the first probability prediction branch to obtain the first probability prediction result. The first splicing result is input into the second probability prediction branch to obtain the second probability prediction result. The first splicing result is input into the third probability prediction branch to obtain the third probability prediction result.

[0110] According to an embodiment of the present invention, the second splicing layer is used to splice the first probability prediction result, the second probability prediction result and the third probability prediction result to obtain a second splicing result.

[0111] According to an embodiment of the present invention, the probability output layer may include a final linear unit and an activation unit. The final linear unit is used to achieve the final dimensional change of the second splicing result, and the activation unit is used to convert the second splicing result after the dimensional change into a probability distribution to obtain a predicted probability.

[0112] According to an embodiment of the present invention, by constructing a probability prediction module based on a multi-layer perceptron, the accuracy of the predicted probability is improved, and the accuracy of the compressed sequence is further improved.

[0113] According to an embodiment of the present invention, the first probability prediction branch includes a first fully connected layer, a first fully connected module, a second fully connected module and a fourth linear transformation layer in sequence; the second probability prediction branch includes a fifth linear layer; and the third probability prediction branch includes a fourth fully connected layer and a sixth linear layer in sequence.

[0114] According to an embodiment of the present invention, the first splicing result is input into the first probability prediction branch to obtain the first probability prediction result, including: inputting the first splicing result into the first fully connected layer to obtain the first fully connected output result; inputting the first fully connected output result into the second fully connected layer of the first fully connected module to obtain the second fully connected output result; inputting the second fully connected output result and the first fully connected output result into the first summation layer of the first fully connected module to obtain the first summation result; inputting the first summation result into the third fully connected layer of the second fully connected module to obtain the third fully connected output result; inputting the third fully connected output result and the second fully connected output result into the second summation layer of the second fully connected module to obtain the second summation result; inputting the second summation result into the fourth linear transformation layer to obtain the first probability prediction result.

[0115] According to an embodiment of the present invention, the first fully connected layer is used to calculate the first concatenation result to obtain a first fully connected output result. Specifically, the calculation process of the first fully connected layer is shown in the following formula (4):

[0116] q1=ReLU(pW1+c1) (4)

[0117] Among them, q1 is the first fully connected output result, p is the first concatenation result, ReLU() represents the ReLU activation function, W1 is the weight parameter of the first fully connected layer, and c1 is the bias parameter of the first fully connected layer.

[0118] According to an embodiment of the present invention, the first fully connected module includes a second fully connected layer and a first summing layer. The second fully connected layer is used to calculate the first fully connected output result to obtain a second fully connected output result. Specifically, the calculation process of the second fully connected layer is shown in the following formula (5):

[0119] q2=ReLU(ReLU(q1)W2+c2)W3+c3(5)

[0120] Among them, q2 is the second fully connected output result, q1 is the first fully connected output result, W2 and W3 are the weight parameters of the second fully connected layer, and c2 and c3 are the bias parameters of the second fully connected layer.

[0121] According to an embodiment of the present invention, the first summing layer is used to sum the first fully connected output result q1 and the second fully connected output result q2 to obtain a first summing result q3.

[0122] According to an embodiment of the present invention, the second fully connected module includes a third fully connected layer and a second summing layer. The third fully connected layer is used to calculate the first summing result to obtain a third fully connected output result. Specifically, the calculation process of the third fully connected layer is shown in the following formula (6):

[0123] q4=ReLU(ReLU(q3)W4+c4)W5+c5(6)

[0124] Among them, q4 is the third fully connected output result, q3 is the first addition result, W4 and W5 are the weight parameters of the third fully connected layer, and c4 and c5 are the bias parameters of the third fully connected layer.

[0125] According to an embodiment of the present invention, the second addition layer is used to add the first addition result q3 and the third fully connected output result q4 to obtain a second addition result q5.

[0126] According to an embodiment of the present invention, the second addition result q5 is input to the fourth linear transformation layer to obtain a first probability prediction result.

[0127] According to an embodiment of the present invention, the fifth linear layer is used to perform a linear transformation on the first splicing result p to obtain a second probability prediction result.

[0128] According to an embodiment of the present invention, the fourth fully connected layer is used to perform a linear transformation on the first concatenation result to obtain a fourth fully connected output result. Specifically, the calculation process of the fourth fully connected layer is shown in the following formula (7):

[0129] q6=ReLU(ReLU(pW6+c6)W7+c7(7)

[0130] Among them, q6 is the fourth fully connected output result, p is the first splicing result, W6 and W7 are the weight parameters of the fourth fully connected layer, and c6 and c7 are the bias parameters of the fourth fully connected layer.

[0131] According to an embodiment of the present invention, the sixth linear layer is used to perform a linear transformation on the fourth fully connected output result q6 to obtain a third probability prediction result.

[0132] According to an embodiment of the present invention, by constructing a complex fully connected layer, the accuracy of the first probability prediction result is improved, and by setting multiple probability prediction branches, the accuracy of the probability prediction result is further improved.

[0133] Figure 5 A schematic diagram of the structure of a probability prediction model according to a specific embodiment of the present invention is shown.

[0134] like Figure 5As shown, the probability prediction model includes a first concatenation layer 501, a first fully connected layer 502, a second fully connected layer 503, a first summation layer 504, a third fully connected layer 505, a second summation layer 506, a fourth linear layer 507, a fourth fully connected layer 508, a sixth linear layer 509, a fifth linear layer 510, a first concatenation layer 511, a final linear unit 512 and an activation unit 513. After the profile data sequence 20 is input into the embedding layer 330, a word embedding vector sequence is obtained, and after the word embedding vector sequence is input into the preliminary probability prediction model 310, a preliminary prediction probability is obtained, and the preliminary prediction probability and the word embedding vector sequence are input into the probability prediction model, and a prediction probability sequence 50 is output.

[0135] like Figure 5 As shown, the probability prediction model includes a first probability prediction branch, a second probability prediction branch, and a third probability prediction branch. The first probability prediction branch includes a first fully connected layer 502, a second fully connected layer 503, a first sum layer 504, a third fully connected layer 505, a second sum layer 506, and a fourth linear layer 507. The second probability prediction branch includes a fifth linear layer 510. The third probability prediction branch includes a fourth fully connected layer 508 and a sixth linear layer 509.

[0136] According to an embodiment of the present invention, based on the predicted probability sequence of the profile data sequence, the profile data sequence is adaptively encoded to obtain a compressed sequence, including: determining the number of iterations according to a predetermined step size, and initializing the starting index of each profile data sequence to facilitate traversal of the data; initializing the storage file and the bit output stream object, setting the initial probability distribution and cumulative frequency table of the characters; adaptively encoding the profile data sequence, and writing the compressed sequence to the storage file.

[0137] According to an embodiment of the present invention, a storage file and a bit output stream object are initialized, and an initial probability distribution and a cumulative frequency table of characters are set, including: creating a compressed data storage file for each profile data sequence, creating a bit output stream object for writing data for each storage file, and creating an adaptive encoder for compression for each bit output stream object; initializing a uniform initial probability distribution for each character in the alphabet; creating a cumulative frequency table, wherein the cumulative frequency table can be used to determine the encoding interval of the character, the cumulative frequency table can be represented by an array, and the cumulative frequency in the cumulative frequency table is calculated by the probability distribution.

[0138] According to an embodiment of the present invention, the profile data sequence is adaptively encoded, including: adaptively encoding the first profile data sequence using an initial probability distribution to obtain a first compressed sequence, which can initialize the state of the adaptive encoder and avoid subsequent boundary problems; traversing each example in the first profile data sequence and the time step of the example, taking the minimum value between the time step and the number of iterations to avoid exceeding the data range; and adaptively encoding other profile data sequences.

[0139] According to an embodiment of the present invention, other profile data sequences are adaptively encoded, including: using a cumulative frequency table and a numerical value space to adaptively encode the profile data sequence and calculate the loss; using a predicted probability sequence and a dictionary to update the cumulative frequency table and the numerical value space; and using the updated cumulative frequency table and the numerical value space to adaptively encode the next profile data sequence.

[0140] According to an embodiment of the present invention, the predicted probability sequence and the dictionary are used to update the cumulative frequency table and the numerical value space, including: using the predicted probability sequence to update the probability distribution of each character in the dictionary; using the probability distribution of each character after the update to update the cumulative frequency table and the numerical value space.

[0141] Figure 6 A schematic diagram of a compression process according to a specific embodiment of the present invention is shown.

[0142] like Figure 6 As shown, the initial profile data sequence set 60 is character-converted to obtain a profile data sequence set 70. The profile data sequence set 70 is segmented to obtain a profile data sequence 20. Specifically, the profile data sequence 20 includes characters S0, ..., characters Sk-1, ..., characters Sn-1 and characters Sn.

[0143] According to an embodiment of the present invention, uniform probability distribution calculation is performed for characters S0, ..., and Sk-1 to obtain S0 probabilities corresponding to character S0, ..., and Sk-1 probabilities corresponding to character Sk-1. Characters S0, ..., and Sk-1 are input into a preliminary probability prediction model 310. The preliminary probability prediction model 310 outputs preliminary prediction probabilities, and the preliminary prediction probabilities are input into a probability prediction model 320 to obtain Sk prediction probabilities corresponding to character Sk, and so on, until Sn prediction probabilities corresponding to character Sn are obtained, that is, a prediction probability sequence 50 is obtained. The prediction probability sequence 50 is adaptively encoded to obtain a compressed sequence 30.

[0144] According to an embodiment of the present invention, a storage file for storing a compressed sequence can be sent to a terminal device. The terminal device uses the same model and parameters as in the compression process to perform cyclic judgment and decoding on the compressed sequence in the storage file and output an initial profile data sequence.

[0145] According to an embodiment of the present invention, the first compressed sequence is adaptively decoded, including: using an initial probability distribution to adaptively decode the first compressed sequence to obtain a first profile data sequence.

[0146] According to an embodiment of the present invention, a compressed sequence is adaptively decoded, including: using a cumulative frequency table and a numerical value space to adaptively decode the compressed sequence; using a predicted probability sequence and a dictionary to update the cumulative frequency table and the numerical value space; and using the updated cumulative frequency table and the numerical value space to adaptively decode the next compressed sequence.

[0147] According to an embodiment of the present invention, a predicted probability sequence and a dictionary are used to update a cumulative frequency table and a numerical value space, including: using the predicted probability sequence to update the probability distribution of each character in the dictionary, wherein the predicted probability sequence can be obtained using an initial probability prediction model and a probability prediction model with the same parameters as those in the compression process; and using the probability distribution of each character after the update to update the cumulative frequency table and the numerical value space.

[0148] Figure 7 A schematic diagram of a decompression process according to a specific embodiment of the present invention is shown.

[0149] like Figure 7 As shown, the compressed sequence 30 is adaptively decoded to obtain a predicted probability sequence 50. According to the predicted probability sequence 50, the profile data sequence 20 is obtained, and then the profile data sequence set 70 is obtained, and the profile data sequence set 70 is character converted to obtain an initial profile data sequence set 60.

[0150] The comparison results of compressed file size (Byte) under different deep learning methods are shown in Table 1. The comparison results of compression rate (%) under different deep learning methods are shown in Table 2. The comparison results of compression ratio (%) under different deep learning methods are shown in Table 3. The comparison results of compression time (s) under different deep learning methods are shown in Table 4. The comparison results of compressed file size (Byte) under different compression formats are shown in Table 5.

[0151] As shown in Tables 1 to 5 below, compared with the gated recurrent unit algorithm and the long short-term memory network, the lossless data processing method based on deep-sea profiling buoy of the present application has an average compression rate reduced by 11.1% on a single batch of 4000m profile data, an average compression ratio increased by 11.2%, and an average compression time shortened by 2.62 times; the compression rate on a single batch of 1000m profile data was reduced by 8.6% on average, the compression ratio increased by 8.6% on average, and the compression time was shortened by 1.47 times on average; compared with the encoder of the transformer network model, the compression rate on a single batch of 4000m profile data was reduced by 8.9%, the compression ratio increased by 9%, and the compression time was shortened by 2.77 times; the compression rate on a single batch of 1000m profile data was reduced by 7.4%, the compression ratio increased by 7.4% on average, and the compression time was shortened by 1.65 times.

[0152] Table 1

[0153]

[0154] Table 2

[0155]

[0156] Table 3

[0157]

[0158] Table 4

[0159]

[0160] Table 5

[0161]

[0162] The lossless data processing method based on deep-sea profiling buoys of the present application has achieved good results on single batches of 4000m and 1000m profile data of deep-sea profiling buoys, and has achieved a significant reduction in compression rate compared to mainstream deep learning compression algorithms, thus improving the compression ratio. The characteristic of BeiDou-3 satellite data transmission is that 228 bytes are transmitted per minute. The lossless data processing method based on deep-sea profiling buoys of the present application can greatly reduce the data transmission time of deep-sea profiling buoys, and has strong practicality.

[0163] In addition, the lossless data processing method based on deep-sea profiling buoys of the present application reduces the compressed file size by an average of 2.89 times on a single batch of 4000m profile data compared to popular compressors; and reduces the compressed file size by an average of 1.61 times on a single batch of 1000m profile data. Experiments show that the lossless data processing method based on deep-sea profiling buoys of the present application has stronger compression performance on deep-sea profiling buoy data than popular compressors.

[0164] Based on the above-mentioned non-destructive data processing method based on deep-sea profiling buoy, the present invention also provides a non-destructive data processing device based on deep-sea profiling buoy. Figure 8 The device is described in detail.

[0165] Figure 8 A structural block diagram of a non-destructive data processing device based on a deep-sea profiling buoy according to an embodiment of the present invention is shown.

[0166] like Figure 8 As shown, the lossless data processing device 800 based on deep-sea profiling buoy of this embodiment includes an embedding module 810, a probability prediction module 820 and an adaptive coding module 830.

[0167] The embedding module 810 is used to perform word vectorization processing on the profile data sequence to obtain a word embedding vector sequence, wherein the profile data sequence is obtained using a deep-sea profile buoy and is used to characterize the state of the ocean at different time periods. In one embodiment, the embedding module 810 can be used to perform the operation S210 described above, which will not be repeated here.

[0168] The probability prediction module 820 is used to repeatedly perform the probability prediction operation until the predicted probability of the last profile data in the profile data sequence is obtained, wherein the predicted probability includes the probability that the last profile data corresponds to the target character in the dictionary, wherein the target character is determined by using the probabilities of each of the multiple characters in the dictionary for representing the last profile data. In one embodiment, the probability prediction module 820 can be used to perform the operation S220 described above, which will not be repeated here.

[0169] The adaptive encoding module 830 is used to adaptively encode the profile data sequence based on the predicted probability sequence of the profile data sequence to obtain a compressed sequence. In one embodiment, the adaptive encoding module 830 can be used to perform the operation S230 described above, which will not be described in detail here.

[0170] According to an embodiment of the present invention, the probability prediction module 820 includes a preliminary probability prediction submodule 821 and a probability prediction submodule 822 .

[0171] The preliminary probability prediction submodule 821 is used to input the first k word embedding vectors into the preliminary probability prediction model to obtain the preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector, wherein the preliminary probability prediction model is generated based on multi-head self-attention. In one embodiment, the preliminary probability prediction submodule can be used to perform the operation S221 described above, which will not be repeated here.

[0172] The probability prediction submodule 822 is used to input the preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector and the first k word embedding vectors into the probability prediction model to obtain the prediction probability of the profile data corresponding to the k+1th word embedding vector, wherein the probability prediction model is generated based on a multi-layer perceptron, k is greater than 1 and less than K, and K is the number of word embedding vectors in the word embedding vector sequence. In one embodiment, the probability prediction submodule can be used to perform the operation S222 described above, which will not be repeated here.

[0173] According to an embodiment of the present invention, the non-destructive data processing device 800 based on the deep-sea profiling buoy further includes a sample division module, a sample preliminary prediction module, a sample prediction module and a parameter adjustment module.

[0174] The sample division module is used to use the profile data subsequence in the profile data sequence as training samples, and determine the profile data set after the training samples from the profile data sequence as labels.

[0175] The sample preliminary prediction module is used to input multiple sample word embedding vectors corresponding to the profile data subsequence into the preliminary probability prediction model to obtain the sample preliminary prediction probability for the label.

[0176] The sample prediction module is used to input multiple sample word embedding vectors and sample preliminary prediction probabilities into the probability prediction model to obtain the sample prediction probability.

[0177] The parameter adjustment module is used to adjust the parameters in the probability prediction model based on the sample prediction probability and label.

[0178] According to an embodiment of the present invention, the preliminary probability prediction model includes: a multi-head attention layer, an intermediate processing layer, a first linear transformation branch, a second linear transformation branch, and a preliminary probability output layer. The preliminary probability prediction submodule includes an attention unit, a processing unit, a transformation unit, and a preliminary prediction unit.

[0179] The attention unit is used to input the first k word embedding vectors into the multi-head attention layer to obtain the attention output result.

[0180] The processing unit is used to input the attention output result into the intermediate processing layer to obtain the intermediate processing result.

[0181] The transformation unit is used to input the intermediate processing result into the first linear transformation branch and the second linear transformation branch respectively, to obtain the first branch transformation result and the second branch transformation result.

[0182] The preliminary prediction unit is used to input the first branch transformation result and the second branch transformation result into the preliminary probability output layer to obtain a preliminary prediction probability.

[0183] According to an embodiment of the present invention, the first linear transformation branch includes a first linear transformation layer; the second linear transformation branch includes in sequence: a second linear transformation layer, an activation layer and a third linear transformation layer; the intermediate processing layer includes in sequence a dimension flipping unit, a slicing unit and a flattening unit.

[0184] According to an embodiment of the present invention, the probability prediction model includes a first splicing layer, a first probability prediction branch, a second probability prediction branch, a third probability prediction branch, a second splicing layer, and a probability prediction layer. The probability prediction submodule includes a first splicing unit, a branch prediction unit, a second splicing unit, and a probability prediction unit.

[0185] The first concatenation unit is used to input the preliminary prediction probability and the first k word embedding vectors into the first concatenation layer to obtain a first concatenation result.

[0186] The branch prediction unit is used to input the first splicing result into the first probability prediction branch, the second probability prediction branch and the third probability prediction branch respectively to obtain the first probability prediction result, the second probability prediction result and the third probability prediction result.

[0187] The second splicing unit is used to input the first probability prediction result, the second probability prediction result and the third probability prediction result into the second splicing layer to obtain a second splicing result.

[0188] The probability prediction unit is used to input the second splicing result into the probability prediction layer to obtain the predicted probability of the profile data corresponding to the k+1th word embedding vector.

[0189] According to an embodiment of the present invention, the first probability prediction branch sequentially includes a first fully connected layer, a first fully connected module, a second fully connected module and a fourth linear transformation layer.

[0190] According to an embodiment of the present invention, the branch prediction unit further includes a first fully connected subunit, a second fully connected subunit, a first summing subunit, a third fully connected subunit, a second summing subunit and a prediction subunit.

[0191] The first fully connected subunit is used to input the first concatenation result into the first fully connected layer to obtain a first fully connected output result.

[0192] The second fully-connected subunit is used to input the first fully-connected output result into the second fully-connected layer of the first fully-connected module to obtain a second fully-connected output result.

[0193] The first summing subunit is used to input the second fully connected output result and the first fully connected output result into the first summing layer of the first fully connected module to obtain a first summing result.

[0194] The third fully-connected subunit is used to input the first addition result into the third fully-connected layer of the second fully-connected module to obtain a third fully-connected output result.

[0195] The second addition subunit is used to input the third fully-connected output result and the second fully-connected output result into the second addition layer of the second fully-connected module to obtain a second addition result.

[0196] The prediction subunit is used to input the second addition result into the fourth linear transformation layer to obtain a first probability prediction result.

[0197] According to an embodiment of the present invention, the second probability prediction branch includes a fifth linear layer; and the third probability prediction branch includes a fourth fully connected layer and a sixth linear layer in sequence.

[0198] According to an embodiment of the present invention, the non-destructive data processing device 800 based on the deep-sea profiling buoy further includes a segmentation module.

[0199] The segmentation module is used to segment the profile data sequence set according to a predetermined step length to obtain multiple profile data sequences.

[0200] According to an embodiment of the present invention, the embedding module 810 includes a character conversion submodule and a vectorization submodule.

[0201] The character conversion submodule is used to perform character conversion on the initial profile data sequence based on a dictionary to obtain a profile data sequence in a predetermined format.

[0202] The vectorization submodule is used to perform word vectorization processing on the profile data sequence using the embedding matrix of the embedding layer to obtain a word embedding vector sequence.

[0203] According to an embodiment of the present invention, any multiple modules of the embedding module 810, the probability prediction module 820 and the adaptive coding module 830 can be combined into one module for implementation, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module. According to an embodiment of the present invention, at least one of the embedding module 810, the probability prediction module 820 and the adaptive coding module 830 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or implemented in any one of the three implementation methods of software, hardware and firmware or in any appropriate combination of any of them. Alternatively, at least one of the embedding module 810, the probability prediction module 820 and the adaptive coding module 830 can be at least partially implemented as a computer program module, and when the computer program module is run, the corresponding function can be executed.

[0204] According to an embodiment of the present invention, the program code for executing the computer program provided by the embodiment of the present invention can be written in any combination of one or more programming languages. Specifically, these computing programs can be implemented using high-level process and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, Java, C++, python, "C" language or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, partially on the remote computing device, or entirely on the remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (e.g., using an Internet service provider to connect through the Internet).

[0205] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0206] It will be appreciated by those skilled in the art that the features described in the various embodiments of the present invention may be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, without departing from the spirit and teachings of the present invention, the features described in the various embodiments of the present invention may be combined and / or combined in various ways. All of these combinations and / or combinations fall within the scope of the present invention.

[0207] The embodiments of the present invention are described above. However, these embodiments are only for the purpose of illustration, and are not intended to limit the scope of the present invention. Although each embodiment is described above, it does not mean that the measures in each embodiment cannot be used in combination advantageously. Without departing from the scope of the present invention, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present invention.

Claims

1. A non-destructive data processing method based on deep-sea profiling buoy, characterized in that: include: Performing word vectorization processing on the profile data sequence to obtain a word embedding vector sequence, wherein the profile data sequence is obtained using a deep-sea profile buoy and is used to characterize the state of the ocean at different time periods; Repeat the following probability prediction operation until the predicted probability of the last profile data in the profile data sequence is obtained, wherein the predicted probability includes the probability that the last profile data corresponds to a target character in the dictionary, wherein the target character is determined by using the probabilities of each of the multiple characters in the dictionary for representing the last profile data: Input the first k word embedding vectors into a preliminary probability prediction model to obtain a preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector, wherein the preliminary probability prediction model is generated based on multi-head self-attention; Inputting the preliminary predicted probability of the profile data corresponding to the k+1th word embedding vector and the first k word embedding vectors into a probability prediction model to obtain the predicted probability of the profile data corresponding to the k+1th word embedding vector, wherein the probability prediction model is generated based on a multilayer perceptron, k is greater than 1 and less than K, and K is the number of word embedding vectors in the word embedding vector sequence; Based on the predicted probability sequence of the profile data sequence, the profile data sequence is adaptively encoded to obtain a compressed sequence.

2. The method according to claim 1, characterized in that Also includes: Using a profile data subsequence in the profile data sequence as a training sample, and determining profile data set after the training sample from the profile data sequence as a label; Inputting a plurality of sample word embedding vectors corresponding to the profile data subsequence into the preliminary probability prediction model to obtain a preliminary prediction probability of the sample for the label; Inputting a plurality of the sample word embedding vectors and the sample preliminary prediction probability into the probability prediction model to obtain the sample prediction probability; Based on the sample prediction probability and the label, the parameters in the probability prediction model are adjusted.

3. The method according to claim 1, characterized in that The preliminary probability prediction model includes: a multi-head attention layer, an intermediate processing layer, a first linear transformation branch, a second linear transformation branch, and a preliminary probability output layer; The first k word embedding vectors are input into the preliminary probability prediction model to obtain the preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector, including: Input the first k word embedding vectors into the multi-head attention layer to obtain the attention output result; Inputting the attention output result into the intermediate processing layer to obtain an intermediate processing result; Inputting the intermediate processing result into the first linear transformation branch and the second linear transformation branch respectively to obtain a first branch transformation result and a second branch transformation result; and The first branch transformation result and the second branch transformation result are input into the preliminary probability output layer to obtain the preliminary prediction probability.

4. The method according to claim 3, characterized in that: The first linear transformation branch comprises a first linear transformation layer; The second linear transformation branch includes, in sequence: a second linear transformation layer, an activation layer, and a third linear transformation layer; The intermediate processing layer includes a dimension flipping unit, a slicing unit and a flattening unit in sequence.

5. The method according to claim 1, characterized in that The probability prediction model includes a first splicing layer, a first probability prediction branch, a second probability prediction branch, a third probability prediction branch, a second splicing layer, and a probability prediction layer; The step of inputting the preliminary predicted probability of the profile data corresponding to the k+1th word embedding vector and the first k word embedding vectors into the probability prediction model to obtain the predicted probability of the profile data corresponding to the k+1th word embedding vector comprises: Inputting the preliminary prediction probability and the first k word embedding vectors into the first concatenation layer to obtain a first concatenation result; Inputting the first splicing result into the first probability prediction branch, the second probability prediction branch and the third probability prediction branch respectively to obtain a first probability prediction result, a second probability prediction result and a third probability prediction result; Inputting the first probability prediction result, the second probability prediction result and the third probability prediction result into a second splicing layer to obtain a second splicing result; The second splicing result is input into the probability prediction layer to obtain the predicted probability of the profile data corresponding to the k+1th word embedding vector.

6. The method according to claim 5, characterized in that: The first probability prediction branch sequentially includes a first fully connected layer, a first fully connected module, a second fully connected module and a fourth linear transformation layer; The second probability prediction branch includes a fifth linear layer; The third probability prediction branch sequentially includes a fourth fully connected layer and a sixth linear layer.

7. The method according to claim 6, characterized in that: The step of inputting the first splicing result into the first probability prediction branch to obtain a first probability prediction result includes: Inputting the first concatenation result into the first fully connected layer to obtain a first fully connected output result; Inputting the first fully connected output result into the second fully connected layer of the first fully connected module to obtain a second fully connected output result; Inputting the second fully connected output result and the first fully connected output result into a first summing layer of the first fully connected module to obtain a first summing result; Inputting the first addition result into the third fully connected layer of the second fully connected module to obtain a third fully connected output result; Inputting the third fully connected output result and the second fully connected output result into the second summing layer of the second fully connected module to obtain a second summing result; The second addition result is input into the fourth linear transformation layer to obtain the first probability prediction result.

8. The method according to claim 1, characterized in that Also includes: The profile data sequence set is segmented according to a predetermined step length to obtain a plurality of the profile data sequences.

9. The method according to claim 1, characterized in that: The word embedding vector sequence is obtained by performing word vectorization processing on the profile data sequence, including: Based on the dictionary, character conversion is performed on the initial profile data sequence to obtain a profile data sequence in a predetermined format; and The profile data sequence is word-vectorized using the embedding matrix of the embedding layer to obtain the word embedding vector sequence.

10. A non-destructive data processing device based on deep-sea profiling buoy, characterized in that: include: An embedding module, used for performing word vectorization processing on a profile data sequence to obtain a word embedding vector sequence, wherein the profile data sequence is obtained using a deep-sea profile buoy and is used to characterize the state of the ocean at different time periods; A probability prediction module, used for repeatedly performing probability prediction operations until the predicted probability of the last profile data in the profile data sequence is obtained, wherein the predicted probability includes the probability that the last profile data corresponds to a target character in a dictionary, wherein the target character is determined by using the probabilities of each of a plurality of characters in the dictionary for representing the last profile data; wherein the probability prediction module includes a preliminary probability prediction submodule and a probability prediction submodule; The preliminary probability prediction submodule is used to input the first k word embedding vectors into the preliminary probability prediction model to obtain the preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector, wherein the preliminary probability prediction model is generated based on multi-head self-attention; The probability prediction submodule is used to input the preliminary prediction probability of the profile data corresponding to the k+1th word embedding vector and the first k word embedding vectors into the probability prediction model to obtain the prediction probability of the profile data corresponding to the k+1th word embedding vector, wherein the probability prediction model is generated based on a multi-layer perceptron, k is greater than 1 and less than K, and K is the number of word embedding vectors in the word embedding vector sequence; The adaptive coding module is used to adaptively encode the profile data sequence based on the predicted probability sequence of the profile data sequence to obtain a compressed sequence.

Images