A method and apparatus for determining a maritime channel blockage strength

Abstract

The application provides a method and device for determining offshore channel blocking strength, and belongs to the technical field of marine communication. The method provided by the application comprises the following steps: acquiring offshore frequency modulation channel data at a current sampling time to form a channel array comprising a plurality of time delay domain data points; performing feature extraction on the channel array based on a single-layer CNN to acquire the phase relationship between each time delay domain data point in the channel array and obtain first features corresponding to the channel array; and performing classification and analysis on the first features based on an LSTM to obtain a blocking degree category and a blocking degree value corresponding to the channel array; wherein, when analyzing the first features, the LSTM fits a distribution function corresponding to the channel array based on the first features, wherein only the blocking degree value is a variable to be fitted in the distribution function. The method and device for determining offshore channel blocking strength provided by the application are used to realize automatic identification of channel states in offshore communication.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of marine communication technology, in particular to a method and device for determining the intensity of channel occlusion at sea. BACKGROUND

[0002] In the field of marine wireless communication, ship-shore communication as a typical application scenario, the channel state is easily affected by the occlusion of marine obstacles, resulting in fluctuations, leading to shadow effects such as sharp decline in signal power and sudden change in propagation characteristics, which directly affects the stability and data transmission efficiency of the communication system.

[0003] Currently, research on channel state recognition focuses on indoor positioning, mainly by extracting parameters such as received signal strength indication and channel measurement value statistical characteristics, combining threshold setting or simple classification algorithms to distinguish between two occlusion states: line-of-sight occlusion and non-line-of-sight occlusion. In the signal processing link, some technologies process the real and imaginary parts of the signal by converting the modulus and phase angle, and use a model design approach that constructs two-dimensional data from multiple antenna collections of impulse signals at different times and extracts features with multiple convolution layers, while indirectly reflecting the occlusion situation based on traditional shadow parameters.

[0004] However, the existing technology has defects. First, the application scenario is limited, and it is difficult to adapt to the complex occlusion environment of marine communication, and the direct diameter occlusion state is not considered, only the classification of two occlusion states: line-of-sight occlusion and non-line-of-sight occlusion can be achieved, which cannot meet the demand for joint recognition of three types of channel states in ship-shore communication. Second, there are deficiencies in feature extraction and model design. The processing method of the real and imaginary parts of some technologies will lose the phase relationship information of the original signal, and the multi-convolution layer design will cause excessive processing of low-dimensional data composed of real and imaginary parts in marine channel signals, and the traditional k-factor needs to define a critical value for different scenarios to determine the occlusion intensity. The fitting effect can only reflect the state at a single time point, and cannot capture the complete occlusion process. SUMMARY

[0005] Therefore, the present application provides a method and device for determining the intensity of channel occlusion at sea to realize the automatic recognition of channel state in marine communication.

[0006] Specifically, the present application is realized by the following technical solutions:

[0007] The first aspect of the present application provides a method for determining the intensity of channel occlusion at sea, comprising:

[0008] Obtaining marine frequency modulation channel data at the current sampling time to form a channel array including a plurality of time delay domain data points;

[0009] extracting features of the channel array based on a single-layer CNN, obtaining a phase relationship between each time delay domain data point in the channel array, and obtaining first features corresponding to the channel array;

[0010] classifying and analyzing the first features based on an LSTM, obtaining a blocking degree category and a blocking degree value corresponding to the channel array;

[0011] In the process of analyzing the first features, the LSTM fits a distribution function corresponding to the channel array based on the first features, wherein only the blocking degree value is a fitting variable in the distribution function.

[0012] The second aspect of the present application provides a sea channel blocking strength determination device, which comprises an acquisition module, an extraction module and an analysis module; wherein,

[0013] The acquisition module is configured to acquire sea frequency modulation channel data at a current sampling time, and construct a channel array comprising a plurality of time delay domain data points.

[0014] The extraction module is configured to extract features of the channel array based on a single-layer CNN, obtain a phase relationship between each time delay domain data point in the channel array, and obtain first features corresponding to the channel array.

[0015] The analysis module is configured to classify and analyze the first features based on an LSTM, and obtain a blocking degree category and a blocking degree value corresponding to the channel array.

[0016] In the process of analyzing the first features, the LSTM fits a distribution function corresponding to the channel array based on the first features, wherein only the blocking degree value is a fitting variable in the distribution function.

[0017] The sea channel blocking strength determination method and device provided by the present application can construct a channel array comprising a plurality of time delay domain data points by acquiring sea frequency modulation channel data at a current sampling time, extract a phase relationship between each time delay domain data point in the channel array by using a single-layer CNN to obtain first features, and classify and analyze the first features by using an LSTM. Not only can the sea channel blocking strength determination method and device output a blocking degree category corresponding to the channel array, but also can obtain a quantitative blocking degree value by fitting a distribution function with only the blocking degree value as a fitting variable. In this way, the sea channel blocking strength determination method and device can be adapted to the sea communication scene, avoid the loss of phase relationship information and the problem of excessive processing by multiple convolution layers in the traditional technology, realize the integrated output of channel state from category recognition to strength quantification, focus on the key variable through the LSTM fitting process, improve the accuracy and efficiency of blocking strength determination, effectively meet the demand for automatic recognition of channel state in sea communication, and provide reliable technical support for subsequent adaptive adjustment of communication parameters and improvement of system robustness. Attached Figure Description

[0018] Figure 1 A flowchart of Embodiment 1 of the method for determining the intensity of maritime channel obstruction provided in this application;

[0019] Figure 2 This is a schematic diagram of the structure of Embodiment 1 of the marine channel obstruction intensity determination device provided in this application. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0021] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0022] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0023] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0024] Figure 1 This is a flowchart of an embodiment of the method for determining the intensity of maritime channel obstruction provided in this application. Please refer to... Figure 1 The method provided in this embodiment may include:

[0025] S101. Obtain the maritime frequency modulation channel data at the current sampling time and construct a channel array including multiple time delay domain data points.

[0026] Specifically, only the offshore FM channel data of a single current sampling time is collected, without mixing data of other times; the channel array is based on a pulse signal unit, a single channel array contains multiple pulse signals, and each pulse signal needs to contain multiple time delay domain data points, and each time delay domain data point needs to retain the real part and imaginary part information of the signal at the same time, and finally form a structured array with dimensions adapted to the subsequent single-layer CNN input.

[0027] Specifically, the offshore FM channel data is instantaneous data of the current sampling time, and is based on a pulse signal to form a unit, contains multiple pulse signals, each pulse signal corresponds to multiple time delay domain data points, and each time delay domain data point retains the real part data and imaginary part data of the signal at the same time, the real part data is the basic strength feature reflecting the amplitude size of the offshore FM channel signal in the time delay domain, and the imaginary part data is the phase information of the signal; the time delay domain data point refers to the discrete data unit reflecting the signal variation characteristics with propagation time in the time delay dimension through sampling of the offshore FM channel signal, specifically, when a single pulse signal is analyzed in the time delay domain, the signal can be sampled at a fixed time interval, each sampling point is a time delay domain data point, and a single pulse signal corresponds to multiple such data points.

[0028] Further, through a high-precision offshore channel measurement device, the original signal of the FM channel at the current sampling time is captured in real time, ensuring that the data only corresponds to the current time, avoiding cross-time signal interference, the captured original signal is split according to the pulse signal to obtain multiple independent pulse signals, and then each pulse signal is analyzed in the time delay domain to extract multiple time delay domain data points, the real part and imaginary part of each time delay domain data point are taken as the dimension components of the array respectively, and the channel array is combined according to the logic of pulse signal-time delay domain data point-real part or imaginary part.

[0029] Optionally, the channel array only includes FM channel data of a sampling time point, and one channel array includes M pulse signals at a sampling time point, each pulse signal contains N time delay domain data points, M is a positive integer greater than 64, and N is a positive integer greater than 1024.

[0030] Specifically, at a single sampling moment, the channel array is composed of impulse signals as a unit, containing M impulse signals. M is defined as a positive integer greater than 64, because too few impulse signals will result in insufficient representation of channel characteristics (such as the inability to cover signal reflection and scattering characteristics at different angles), and a number greater than 64 can ensure that the amount of data is sufficient to reflect the basic characteristics of the channel at that moment, providing sufficient samples for feature extraction; each impulse signal is further decomposed into N time delay domain data points, N being a positive integer greater than 1024, and the time delay domain data points are discrete sampling results of the signal at different propagation times (each point contains a real part and an imaginary part, reflecting the amplitude and phase, respectively), and a number greater than 1024 can completely cover the time delay dimension information of the signal (such as the full range from near direct to far reflection), avoiding the loss of key features such as phase relationship and intensity change due to insufficient sampling points. By strictly limiting the time dimension, the number of impulse signals and the number of time delay domain data points, both the accurate description of the instantaneous state of the channel array and the reliability of subsequent feature extraction and occlusion analysis are ensured by sufficient data, which meets the needs of channel state recognition in complex marine communication environments.

[0031] In this way, only the current sampling moment is focused on, ensuring that the data completely matches the current channel occlusion state, laying a foundation for accurate analysis in the time dimension; at the same time, by limiting the parameter range of M>64, N>1024, both the insufficient number of impulse signals leading to poor feature representation are avoided, and the complete channel time delay information is covered by sufficient time delay domain data points, and the real part and the imaginary part are preserved to avoid the loss of key signal information, providing a data basis for subsequent extraction of phase relationships.

[0032] S102, performing feature extraction on the channel array based on a single-layer CNN, obtaining the phase relationship between each time delay domain data point in the channel array, and obtaining a first feature corresponding to the channel array.

[0033] Specifically, the channel array is normalized for pretreatment; local spatial features are extracted through a single-layer CNN architecture and converted into two-dimensional spatial features; one-dimensional time sequence features are output through a flattening layer as the first features. The first features represent the phase relationship between each time delay domain data point in the channel array, including the phase difference between data points at different time delay positions (such as the phase difference between direct signals at close distances and reflected signals at long distances), the local trend of the phase change with the pulse signal (such as the phase mutation pattern caused by occlusion), and the phase correlation law formed by the real part and the imaginary part data (without modulus length and phase angle conversion, the original phase information is completely retained). The first features not only retain the core phase features of the sea channel (which is different from the defect of losing phase information in traditional technologies), but also adapt to the time sequence analysis capability of LSTM through the one-dimensional time sequence format, providing a direct basis for subsequent LSTM to realize occlusion degree category classification and quantitative value fitting, and are the core feature data for realizing accurate identification of the occlusion strength of the sea channel.

[0034] By replacing the multi-convolution layer with a single-layer CNN, the over-processing of the low-dimensional sea channel data of the real part data and the imaginary part data by the multi-convolution layer is avoided, and the phase relationship information is prevented from being destroyed; by normalizing the pretreatment, the interference of signal intensity differences is eliminated, and it is ensured that the single-layer CNN can focus on phase relationship extraction, and the first features output finally not only retain the core phase correlation information of the channel, but also adapt to the time sequence processing capability of LSTM through format conversion, providing high-quality feature input for subsequent occlusion state analysis.

[0035] Further, the single-layer CNN structure includes a single-layer convolution layer, a pooling layer, and an activation layer, which traverses all time delay domain data points in the channel array at the current sampling time, and filters out the maximum value and the minimum value of all time delay domain data points; the normalization calculation is performed on each time delay domain data point to eliminate the signal intensity differences caused by different distances and avoid the intensity interference with the phase relationship extraction.

[0036] Further, the implementation steps of the normalization calculation on each time delay domain data point include:

[0037] (1) Traversing all time delay domain data points at a sampling time point, filtering the maximum value and the minimum value of the channel array;

[0038] Specifically, through the traversal algorithm, the real part value and the imaginary part value of each time delay domain data point are read in sequence, and the maximum value and the minimum value among all the values are recorded. For example, if a data point has a real part of 2.5 and an imaginary part of -1.8, and another data point has a real part of 1.9 and an imaginary part of -2.3, the minimum value of the channel array after traversal is -2.3, and the maximum value is 2.5. The maximum value and the minimum value define the upper and lower boundaries of the signal intensity at the collection time, providing a reference for subsequent elimination of signal intensity differences caused by different distances, and avoiding the masking of key features such as phase relationship due to intensity differences.

[0039] (2) subtract the value of each time delay domain data point in the channel array from the minimum value, and divide by the difference between the maximum value and the minimum value to obtain a normalized channel array.

[0040] Specifically, in combination with the above description, for example, the real part of a data point has an original value of 1.5, the minimum value of the channel array determined is -2.3, and the maximum value is 2.5, then the normalized value of the real part is ; similarly, the normalized value of the imaginary part of the data point is calculated, and finally a channel array in which all data points are standardized is formed.

[0041] In marine communication, signal strength will differ significantly due to different propagation distances (such as the amplitude of a direct signal at close range being much higher than that of a reflected signal at a long distance), and after normalization, all data point values are mapped to the [0, 1] interval, which can eliminate the intensity interference caused by distance, so that the subsequent single-layer CNN can focus on the phase relationship between data points; on the other hand, the normalized data has a uniform scale, which can avoid the model training bias caused by too large amplitude differences in the original data, ensuring that the subsequent single-layer CNN and LSTM module can stably and accurately process features, improving the reliability of occlusion strength identification.

[0042] Further, the normalized channel array is input into a single convolution layer, the convolution kernel size and step size are set to adapt to the low-dimensional data of the real and imaginary parts, the local correlation between the time delay domain data points is captured through convolution operation, and the local spatial features (mainly containing phase relationship information) are output; the local spatial features are input into a pooling layer, the key information (such as significant phase difference points) in the local spatial features is retained, while the data dimension is reduced and the redundancy is reduced, avoiding excessive computational complexity in the subsequent calculation; the pooled local spatial features are input into an activation layer, a nonlinear transformation is introduced by using an activation function such as ReLU, effective features (such as phase mutation features related to occlusion) are strengthened, and invalid noise is suppressed, and two-dimensional spatial features are output; the two-dimensional spatial features are input into a flattening layer, and the multi-dimensional spatial features are converted into one-dimensional time series features through flattening processing, the one-dimensional time series features are the first features, the dimension is adapted to the input requirements of the subsequent LSTM, and the phase relationship between the time delay domain data points is completely retained.

[0043] Further, the implementation steps of feature extraction on the channel array based on the single-layer CNN include:

[0044] (1) input the channel array into a single convolution layer, and the single convolution layer outputs local spatial features;

[0045] Specifically, the normalized channel array is first adjusted to a format acceptable to a single convolutional layer, and a suitable convolution kernel parameter is selected in view of the low-dimensional characteristics of the real part combined with the imaginary part of the offshore channel data; the convolution kernel integrates the amplitude variation and phase correlation of the data points in the local region into a characteristic value through cross-correlation operation with the local region in the channel array, outputs the local spatial feature, and the local spatial feature has preliminarily included the phase relationship between the data points in each time delay domain.

[0046] Further, it can be understood that the dimension of the channel array is Mx2xN, and after adjustment, the dimension forms a 4D tensor of 1xMx2xN, wherein 1 represents a single sample, the convolution kernel size is set to 3x3, the step is 1, and the number of output channels (such as 32) is set. By traversing the channel array with a sliding convolution kernel, a local spatial feature with a dimension of 1xM'x2'xC is outputted (M', 2' are the dimensions of the pulse signal and the real and imaginary parts after convolution, and C is the number of output channels).

[0047] (2) The local spatial feature is inputted into a pooling layer, and the pooling layer performs pooling processing on the local spatial feature to output a pooled local spatial feature;

[0048] Specifically, the maximum pooling is adopted for the phase mutation point in the offshore channel, the local spatial feature outputted by the single convolutional layer is divided into a plurality of non-overlapping local regions according to the pooling kernel size, the maximum value in each region is taken as the representative feature of the region, the dimension of the pooled local spatial feature is compressed, the redundant information is removed, and the key features are reserved.

[0049] For example, in an embodiment, the pooling kernel size is set to 2x2, and the step is 2; the four feature values in a certain 2x2 region are [0.3, 0.8, 0.5, 0.6], and the output after pooling is 0.8, and the pooled local spatial feature with a dimension of 1xM''x2''xC'' is obtained, wherein M''=M' / 2 and 2''=2' / 2. By reducing the dimension, the amount of subsequent calculation is reduced, and model overfitting is avoided; on the other hand, by retaining the key features in the local region, the phase difference information related to occlusion is strengthened, and the effectiveness of the features is improved.

[0050] (3) The pooled local spatial feature is inputted into an activation layer, and the activation layer processes the local spatial feature by using an activation function to output a two-dimensional spatial feature;

[0051] Specifically, the ReLU function is adopted to adapt to the nonlinear expression requirement of the sea channel characteristics; all negative values in the pooled local spatial features are set to 0, and the positive values remain unchanged; the output dimension is consistent with the two-dimensional spatial features after the local spatial features are pooled, and the features in the two-dimensional spatial features have been strengthened through nonlinear transformation of effective features (positive values corresponding to phase correlation information) and suppressed noise (negative values corresponding to interference information). For example, in an embodiment, when a certain feature value is -0.2, it is changed to 0 after ReLU function processing; when the feature value is 0.9, it remains 0.9 unchanged.

[0052] The nonlinear transformation is introduced to solve the problem that the linear convolution cannot depict the complex phase relationship of the sea channel, and at the same time, the negative features are set to zero to filter environmental noise, so that the two-dimensional spatial features are more focused on effective phase correlation information.

[0053] (4) The two-dimensional spatial features are input into a flattening layer, the flattening layer performs flattening processing on the two-dimensional spatial features, and outputs one-dimensional time sequence features as the first features.

[0054] Specifically, the two-dimensional spatial features output by the activation layer are dimensionally expanded in the order of channel-pulse signal-real and imaginary parts, and the multi-dimensional spatial features are converted into one-dimensional vectors; that is, one-dimensional time sequence features as first features. For example, in an embodiment, the two-dimensional spatial feature dimension 1x32x1x32 (M''=32, 2''=1, C=32) is flattened into a one-dimensional time sequence feature with a length of 32x1x32=1024. It is ensured that the subsequent LSTM can mine the time sequence trend of the phase relationship based on the first features, and provide high-quality input for the occlusion degree category classification and quantization value fitting.

[0055] Further, after the single-layer CNN extracts the features of the channel array, it further includes:

[0056] (1) The first features output by the single-layer CNN are acquired and input into a subsequent CNN layer to extract high-order spatial features;

[0057] Specifically, the subsequent CNN layer is composed of 1 to 2 convolution layers, each layer is configured with an adaptive convolution kernel, a step length and an output channel number, and the first features are processed in depth through multi-layer convolution operation.

[0058] Further, the first feature (one-dimensional time sequence feature) output by the single-layer CNN is re-adjusted into a spatial feature format (such as restored into a two-dimensional matrix form) as the input of the subsequent CNN layer; for the implementation process of the subsequent CNN layer for feature extraction of the first feature, please refer to the description in the related art, which will not be repeated here. Through the progressive feature extraction of the subsequent CNN layer, the basic phase relationship is upgraded to a high-order feature with more semantic value, solving the problem of limited feature expression capability of the single-layer CNN, and providing more abundant feature basis for subsequent occlusion strength analysis.

[0059] (2) The high-order spatial feature is input into a connection layer, and the connection layer performs pooling and flattening on the high-order spatial feature to output a one-dimensional time sequence feature;

[0060] Specifically, the connection layer includes a pooling layer and a flattening layer. The adaptive pooling (such as global average pooling) is used to reduce the dimension of the high-order spatial feature, eliminate the difference in spatial dimension, and at the same time retain the overall distribution trend of the key feature. Then the flattening layer is used to convert the high-order spatial feature after the pooling into a one-dimensional vector, so as to ensure that the length of the output one-dimensional time sequence feature matches the input dimension of the subsequent module (such as LSTM). Through the combination of the pooling and flattening operations, the format conversion of the high-order spatial feature to the time sequence feature is realized, and a bridge between the subsequent CNN layer and the subsequent analysis module is built. The abstract correlation information of the high-order spatial feature is retained, and the feature is standardized in format to ensure that it can be effectively processed by the subsequent module, avoiding information loss or processing failure caused by mismatched feature dimensions.

[0061] (3) The one-dimensional time sequence feature is used to replace the first feature as the corresponding feature information of the channel array.

[0062] Specifically, the one-dimensional time sequence feature (fusing the basic phase relationship and the high-order abstract feature) output by the connection layer is used to replace the first feature output by the original single-layer CNN as the final feature information representing the channel array, which is used for subsequent LSTM classification and quantization of the occlusion degree category. In this way, the feature information of the channel array can be upgraded from a single basic feature to a higher-order composite feature. The high-order composite feature can more accurately depict the essential differences of different occlusion states, and the high-order composite feature also contains multi-level information, which can reduce the influence of noise on a single feature, making the subsequent occlusion strength analysis more stable and reliable. The process realizes the deep mining and format adaptation of the feature information by adding the subsequent CNN layer and the connection layer, and the finally output feature can more comprehensively and accurately reflect the occlusion-related characteristics of the sea channel, laying a foundation for improving the accuracy of occlusion strength determination.

[0063] S103, classifying and analyzing the first feature based on the LSTM to obtain the occlusion degree category and the occlusion degree value corresponding to the channel array;

[0064] The LSTM is used for fitting a distribution function corresponding to the channel array based on the first feature when analyzing the first feature, wherein only the occlusion degree value in the distribution function is a variable to be fitted.

[0065] Specifically, the LSTM is used for establishing an association between the first feature and a preset classification label, and a category is output by probability calculation to realize classification; and the LSTM is used for fitting a distribution function to output a quantized occlusion degree value, so as to realize dual output of an occlusion degree category and an occlusion degree value.

[0066] Further, the first feature can be input into the LSTM, the LSTM is used for capturing a time sequence association (such as a trend of a phase relationship changing with a time delay) implied in the first feature by using a gating mechanism (an input gate, a forgetting gate, and an output gate) of the LSTM, and a target feature reflecting an association degree between the first feature and each preset classification label (LOS, OLOS, and NLOS) is output; the target feature is input into a fully connected layer, and then is input into a Softmax layer after adjusting a feature dimension by the fully connected layer, so as to calculate a probability distribution of the first feature belonging to each preset classification label; and a preset classification label corresponding to a maximum probability value in the probability distribution is selected as an occlusion degree category of the channel array.

[0067] Compared with a traditional binary classification model (only distinguishing a line of sight or a non-line of sight), the LSTM can mine subtle differences (such as a local phase disorder but a stable overall trend mode specific to a direct diameter being occluded state) of a phase relationship in the first feature by using a time sequence modeling capability, so as to realize accurate classification of three types of occlusion states, fill a technical gap of recognizing a partial occlusion state in maritime communication, and provide a possibility for fine classification of channel states.

[0068] Further, the distribution function is a mathematical model pre-constructed based on historical data and associated with the first feature and the occlusion degree, and a formula form of the distribution function is an occlusion distribution * weight + a direct distribution * (1-weight), wherein the occlusion distribution represents a feature rule when a signal is occluded, the direct distribution represents a feature rule when a signal is directly transmitted, and the weight is positively correlated with the occlusion degree (the stronger the occlusion is, the greater the weight is). When the LSTM is analyzed, all parameters (such as basic morphological parameters of the occlusion distribution or the direct distribution) in the distribution function except the occlusion degree value are fixed based on a historical fitting result, and only the occlusion degree value is a variable to be fitted. Specifically, the current first feature is substituted into the distribution function, the LSTM adjusts the occlusion degree value by using an optimization algorithm (such as a least square method) to make an error between a theoretical feature output by the distribution function and the current first feature minimum (for example, an occlusion degree value of 0.3 is found by iterative calculation to make the error less than 0.01).

[0069] By only retaining the occlusion degree value as a variable to be fitted, the fitting complexity is greatly reduced, and the quantization accuracy is improved; at the same time, fixing other parameters ensures the consistency of the fitting standard at different sampling times, so that the occlusion degree value can objectively reflect the occlusion intensity at the same time, and provides a directly reusable quantization basis for dynamic adjustment of marine communication parameters. On the one hand, through the time series processing capability of LSTM, the dynamic trend of the phase relationship in the first feature can be mined, and the precise classification of the three types of occlusion states of line-of-sight, non-line-of-sight and direct diameter occlusion can be realized by combining the Softmax probability calculation, to meet the needs of marine communication for intermediate occlusion state recognition; on the other hand, by outputting the quantized occlusion degree value, the occlusion state is upgraded from a qualitative category to a quantitative index, providing a more detailed decision basis for subsequent adaptive adjustment of communication parameters (such as adjusting the transmission power according to the occlusion degree value), and improving the practical value of the technology.

[0070] Further, the distribution function needs to be constructed based on the association between the first feature and the occlusion degree, and the construction process needs to combine the weighted fusion of the occlusion distribution and the direct distribution; all parameters except the occlusion degree value need to be fixed based on historical fitting results before fitting; during fitting, only the occlusion degree value is adjusted to make the distribution function output match the first feature.

[0071] Further, the first feature is classified based on LSTM, including:

[0072] (1) determining a target feature of the association between the first feature and each preset classification label based on LSTM;

[0073] Specifically, the first feature is time series modeled through the gating mechanism of LSTM, and a target feature quantifying the association strength between the first feature and each preset classification label is outputted, wherein the dimension of the target feature needs to be consistent with the number of preset classification labels, and each dimension value directly reflects the association degree between the first feature and the corresponding label.

[0074] Further, LSTM can capture long-term dependencies in time series data through the synergistic effect of the input gate, the forget gate and the output gate; the target feature refers to a high-dimensional feature vector outputted by LSTM after processing the first feature, each element corresponds to the association strength between the first feature and a preset classification label, and the preset classification label is a pre-set marine channel occlusion state category, including line-of-sight occlusion, partial occlusion and non-line-of-sight occlusion.

[0075] Further, the first feature is input into the LSTM sequentially by time steps, each time step corresponds to a segment of the first feature, and the LSTM processes the first feature through its gating mechanism, the forget gate filters out the historical information in the first feature irrelevant to the occlusion state (such as phase fluctuations caused by random noise), retains the key time trend, the input gate controls the feature segment of the current time step to enter the LSTM cell state, updates the cell state to include new phase relationship information, and the output gate outputs the hidden state of the current time step based on the updated cell state, reflecting the time sequence feature summary up to this step. The hidden state of the last time step of the LSTM is mapped to a dimension matching the number of preset classification labels through a fully connected layer to obtain the target feature. For example, the target feature [0.8, 0.1, 0.1] indicates that the first feature is most strongly associated with the line-of-sight occlusion and less strongly associated with the partial occlusion and non-line-of-sight occlusion.

[0076] (2) calculating a probability distribution of the first feature belonging to each preset classification label based on the target feature;

[0077] Specifically, the target feature is input into a Softmax activation function, and the probability value of the first feature belonging to each preset classification label is output, the sum of the probability values of all preset classification labels is 1, and each probability value is in the interval [0, 1], directly quantifying the possibility of the first feature belonging to the corresponding category.

[0078] Further, the probability distribution is a vector composed of multiple probability values matching the preset classification labels, each element corresponding to the probability of the first feature belonging to a certain preset classification label, for example, the probability distribution is [0.85, 0.1, 0.05], indicating that the first feature has an 85% probability of belonging to the line-of-sight occlusion, a 10% probability of belonging to the partial occlusion, and a 5% probability of belonging to the non-line-of-sight occlusion. For the specific implementation process of the Softmax activation function converting the target feature into a probability value, please refer to the description in the related technology, which will not be repeated here.

[0079] (3) selecting the preset classification label corresponding to the maximum probability value as the classification result of the first feature.

[0080] Specifically, in combination with the above description, after obtaining the probability distribution corresponding to the first feature, the maximum probability value in the probability distribution is taken as the maximum probability value, and the preset classification label corresponding to the maximum probability value is taken as the classification result of the first feature.

[0081] Further, the probability distribution vector is traversed to find the maximum value and the corresponding preset classification label. For example, the maximum value of the probability distribution [0.85, 0.1, 0.05] is 0.85, and the corresponding preset classification label is the line-of-sight occlusion, so it is determined that the classification result of the first feature is the line-of-sight occlusion.

[0082] Further, in the probability distribution vector, if there are multiple same maximum values, the preset classification label with higher matching degree is selected as the classification result in combination with the matching degree of the occlusion degree value and the category interval.

[0083] Further, the first feature is analyzed based on the LSTM, including:

[0084] (1) A distribution function is established according to the first feature, and the distribution function is a relationship between the first feature and the occlusion degree;

[0085] Specifically, based on the correlation between the time sequence feature of the first feature and the occlusion degree, a distribution function is constructed with the occlusion degree as the core variable, wherein the input of the distribution function is the key feature value (such as the phase relationship parameter, the amplitude distribution statistic) of the first feature, and the output is the theoretical feature distribution matched with the first feature, and the function form needs to clearly reflect the regulation of the occlusion degree on the feature distribution (for example, the higher the occlusion degree, the closer the function output to the occlusion distribution feature). The distribution function is a mathematical expression for describing the mapping relationship between the first feature (including phase relationship, amplitude change, etc.) and the occlusion degree, and the first feature is a one-dimensional time sequence feature extracted by CNN, which includes the phase relationship and time sequence trend of each time delay domain data point in the channel array, and is the core feature representing the state of the sea channel.

[0086] Further, key statistics (such as phase mutation frequency, amplitude peak position, feature value variance, etc.) are extracted from the first feature as input dimensions of the distribution function, and a shielding distribution (reflecting the feature rule when completely shielding) and a direct distribution (reflecting the feature rule when there is no shielding) are constructed. The two distributions are weighted and fused by a weight coefficient (i.e. the occlusion degree value) to form the distribution function, with time delay position as the horizontal axis and feature value probability density as the vertical axis.

[0087] Further, the global parameter set is determined ; the first feature is taken as an input variable x, wherein in the global parameter set, as a variable related to the occlusion degree, other parameters can be set or estimated according to historical data or prior knowledge.

[0088] The distribution function can be constructed by the following formula:

[0089] ;

[0090] wherein, is a shared global distribution parameter;

[0091] is the weight parameter of the i th group of data.

[0092] (2) fixing the parameters in the distribution function other than the degree of occlusion according to the historical fitting results;

[0093] Specifically, based on a large amount of fitting data of historical first features and corresponding degrees of occlusion, all parameters in the distribution function other than the degree of occlusion (such as the mean and variance of the occlusion distribution, the peak position and decay coefficient of the direct light distribution, etc.) are determined through statistical analysis or machine learning methods, and these parameters are fixed as constants, ensuring the consistency of the fitting standards at different sampling times.

[0094] Further, a sufficient number of historical first feature samples and their corresponding manually labeled or high-precision device measured degrees of occlusion are collected. For each historical sample, the best parameter combination is obtained through multi-parameter fitting, and the optimal fixed value of each parameter is determined through statistical methods (such as mean, median) or regression models (such as linear regression). The optimized parameters are written into the distribution function to form a simplified model containing only one variable of the degree of occlusion.

[0095] Further, a large amount of historical first feature data and corresponding known degrees of occlusion are used to fit the established distribution function. In the fitting process, the global simultaneous fitting target function is used.

[0096] wherein, is the global parameter set;

[0097] is the observation value of the jth sampling point of the ith data set;

[0098] is the jth sampling point of the ith data set;

[0099] M is the number of sampling points of each data set;

[0100] N is the number of data sets.

[0101] Through this target function, the parameters in the global parameter set are adjusted so that the difference between the observation value and the function value is minimized. After the fitting is completed, the other parameters other than the degree of occlusion value are fixed, obtaining a distribution function with only the degree of occlusion value as a variable.

[0102] Further, for the ith data set, the distribution function is defined as:

[0103] ;

[0104] wherein, is the weight coefficient of the ith data set;

[0105] is a probability density function of a lognormal distribution;

[0106] is a probability density function of a Rician distribution;

[0107] is a shared global distribution parameter.

[0108] Further, the probability density function of a lognormal distribution is:

[0109] ;

[0110] wherein, is a log scale parameter;

[0111] is a log shape parameter, greater than 0;

[0112] is greater than 0.

[0113] Further, the probability density function of a Rician distribution is:

[0114] ;

[0115] wherein, is a non-centrality parameter of the Rician distribution, ;

[0116] is a scale parameter of the Rician distribution, ;

[0117] is a first kind zeroth order modified Bessel function;

[0118] .

[0119] (3) fitting the distribution function according to the first feature, to obtain the occlusion degree value.

[0120] Specifically, the current first feature data is substituted into the distribution function after the fixed parameters. At this time, the distribution function has only one unknown variable, the occlusion degree value, and the value of the occlusion degree is adjusted through an optimization algorithm (such as gradient descent method, etc.) to make the output of the distribution function most consistent with the actual situation of the current first feature (i.e. minimize the objective function). When the convergence condition is reached, the value of the occlusion degree at this time is the occlusion degree value to be solved.

[0121] Further, before fitting the distribution function corresponding to the channel array based on the first feature, the implementation steps include:

[0122] (1) selecting multiple continuous pulse signals as a set of fitting data;

[0123] Specifically, the multiple continuous pulse signals refer to selecting M continuous pulse signals from the channel array at the current sampling time, and the M pulse signals are continuous in time sequence, which means arranging in the original collection order without skipping or disturbing the order. Continuous pulse signals can preserve the time sequence correlation of signal characteristics and avoid feature breakage caused by non-continuous selection. By selecting continuous pulse signals, it is ensured that the fitting data can fully reflect the dynamic changes of channel characteristics in a certain period of time.

[0124] (2) determining the occlusion distribution and the direct distribution according to the fitting data, respectively;

[0125] Specifically, the occlusion distribution refers to the probability distribution model of the change of channel characteristics with time delay position when the signal propagation is blocked by obstacles. The segments that obviously reflect the occlusion characteristics (such as pulse signals with frequent phase mutations and drastic real amplitude fluctuations) can be selected from the fitting data, and the time delay domain data points (including real and imaginary parts) of these segments are statistically modeled to obtain the probability distribution (such as mean, variance, peak position, etc. parameterized distribution) reflecting the phase relationship and amplitude change rule when the signal is blocked, as the occlusion distribution. In the time delay dimension, the probability density of the occlusion distribution presents the characteristics of dispersion and multiple peaks, that is, the signal energy is not concentrated in the near time delay position (the typical region corresponding to the direct diameter), but is dispersed to multiple time delay positions due to the reflection and scattering effect caused by the occlusion, and the characteristic values (real and imaginary parts) of different time delay positions fluctuate greatly without stable rules.

[0126] Further, the direct distribution refers to the probability distribution model of the change of channel characteristics with time delay position when the signal propagation is not blocked by obstacles and is mainly in the direct diameter. The segments that reflect the direct characteristics (such as pulse signals with stable phase and stable real amplitude) can be selected from the fitting data, and the probability distribution reflecting the phase relationship and amplitude change rule when the signal is directly transmitted is obtained by statistical modeling, as the direct distribution. In the time delay dimension, the probability density of the direct distribution presents the characteristics of centralization and single peak, that is, the signal energy is highly concentrated in the near time delay position (the region corresponding to the time of direct diameter propagation), the probability density of this position is much higher than that of other time delay positions, and the characteristic values (real and imaginary parts) fluctuate small and present a stable linear or periodic change trend.

[0127] The occlusion distribution and the direct distribution both take the time delay position as the horizontal axis and take the eigenvalue probability density as the vertical axis, and quantify the appearance probability of the signal characteristics at different time delay positions (for example, the direct distribution has a higher probability density at the near distance time delay position, and the occlusion distribution has a more dispersed probability density at the far distance time delay position). By separating and extracting the occlusion distribution and the direct distribution, the channel characteristic rules in the two extreme states (complete occlusion and complete direct) are accurately captured, which provides a basis for subsequent construction of a mixed distribution (adaptation to a partially occluded state), and solves the problem that the traditional model is difficult to depict the mixed occlusion characteristics.

[0128] (3) determining a weight coefficient of the occlusion distribution and the direct distribution according to the occlusion intensity;

[0129] Specifically, the weight coefficient of the occlusion distribution is a, and the weight coefficient of the direct distribution is 1-a. The weight coefficient of the occlusion distribution is positively correlated with the occlusion intensity. The higher the occlusion intensity, the closer the weight coefficient of the occlusion distribution to 1 (occlusion distribution dominant). The lower the occlusion intensity, the closer the weight coefficient of the occlusion distribution to 0 (direct distribution dominant). In a partially occluded state, the weight coefficient of the occlusion distribution takes an intermediate value between 0 and 1, for example, when the occlusion intensity is 0.5, a=0.5, and the two distributions each account for half of the weight.

[0130] Based on the association rules of historical occlusion intensity data and corresponding characteristics, a mapping relationship between the occlusion intensity and the weight coefficient of the occlusion distribution is established through linear mapping or machine learning fitting (such as least squares fitting), to ensure that the weight coefficient can accurately reflect the influence weight of the occlusion degree on the two distributions.

[0131] (4) performing weighted summation on the occlusion distribution and the direct distribution based on the weight coefficient to establish a distribution function.

[0132] Specifically, the distribution function can be established by the following formula:

[0133]

[0134] wherein, is the weight coefficient of the occlusion distribution;

[0135] is the occlusion distribution;

[0136] is the direct distribution.

[0137] The distribution function constructed by the weighted summation can accurately depict the actual characteristics of the occlusion and direct mixing of the sea channel (for example, in a partially occluded state, the near distance time delay position is greatly affected by the direct, and the far distance is greatly affected by the occlusion), and the function form is simple and the parameters are clear, which provides a feasible mathematical model for subsequent LSTM to fit the current first characteristics only by optimizing the occlusion degree value, and greatly improves the efficiency and accuracy of the occlusion degree quantization.​

[0138] In summary, by constructing the distribution function in steps, both the extreme characteristics of occlusion and direct light are preserved, and the mixed state is continuously described through the weight coefficient. The final distribution function can accurately match the complex occlusion characteristics of the sea channel, laying a core foundation for the LSTM single variable fitting of the occlusion degree value.

[0139] Further, after obtaining the occlusion degree category and the occlusion degree value corresponding to the channel array, the method comprises:

[0140] (1) setting an occlusion degree value interval corresponding to the occlusion degree category;

[0141] Specifically, based on the statistical association of the occlusion degree category and the occlusion degree value in the historical data (such as the occlusion degree values of a large number of line-of-sight occlusion samples being concentrated in 0-0.3), combined with the physical characteristics of different occlusion states in the marine communication scene (such as the high proportion of direct light in the line-of-sight occlusion state, the occlusion degree value must be low), it is ensured that the interval boundary can accurately distinguish different categories. For the preset occlusion degree category (line-of-sight occlusion, partial occlusion, and non-line-of-sight occlusion), the corresponding occlusion degree value (quantized value in the 0-1 interval) range is determined. For example, in an embodiment, the occlusion degree value interval of line-of-sight occlusion is set to (0, 0.3), the occlusion degree value interval of partial occlusion is set to (0.3, 0.7), and the occlusion degree value interval of non-line-of-sight occlusion is set to (0.7, 1.0).

[0142] (2) calculating the matching degree of the occlusion degree value falling into the interval corresponding to the interval;

[0143] Specifically, the matching degree is a measure of the degree of fit between the current occlusion degree value and the category interval. The higher the value of the matching degree, the better the fit. If the occlusion degree value falls within the interval of a certain category, the distance between the value and the center of the interval is calculated (the smaller the distance, the higher the matching degree). For example, the center of the partial occlusion interval (0.3, 0.7) is 0.5, and if the current value is 0.5, the matching degree is 1.0; if the value is 0.3 (the edge of the interval), the matching degree is 0.5. If the occlusion degree value falls outside the interval (such as the classification result is line-of-sight occlusion but the occlusion degree value is 0.4), the matching degree is attenuated according to the distance from the interval boundary (the farther the distance, the lower the matching degree, the lowest being 0). When the matching degree is high, it means that the occlusion degree category and the occlusion degree value are consistent and do not need to be adjusted significantly; when the matching degree is low, there is a contradiction and the occlusion degree category needs to be corrected through weight adjustment.

[0144] (3) assigning weights to the preset multiple classification labels in the LSTM based on the matching degree;

[0145] Specifically, the preset classification labels include line-of-sight occlusion, partial occlusion, and non-line-of-sight occlusion, the weights of the preset classification labels are positively correlated with the matching degrees of the corresponding categories, if the current classification result is partial occlusion and the matching degree is 0.9, the weight of the partial occlusion is set to 0.9, and the weights of the line-of-sight occlusion and the non-line-of-sight occlusion are allocated according to their correlation degrees (for example, 0.05 and 0.05, respectively). If the current classification result is line-of-sight occlusion but the matching degree is only 0.2, the weight of the line-of-sight occlusion is reduced, the weight of the partial occlusion is increased, and the weight of the non-line-of-sight occlusion is still low. By dynamically allocating the weights according to the matching degrees, the importance of the categories is adaptively adjusted, the category with a high fitting degree with the quantization value is given a higher weight, and the influence of the category in the final result is strengthened; the weight of the category with a low fitting degree is reduced, and the influence of the category is weakened, which provides a mechanism for correcting contradictory results.

[0146] (4) Adjusting the probability distribution values corresponding to each preset classification label of the first feature according to the weights, and adjusting the classification result output by the LSTM.

[0147] Specifically, the probability distribution values output by the LSTM are fused with the adjusted weights to obtain an adjusted probability distribution. The category with the highest probability after adjustment is selected as the final classification result.

[0148] By fusing the weights, the information of the quantization value and the original feature is solved, and the deviation that may exist in single feature classification (for example, the case that the features are similar but the quantization values are significantly different) is solved. For example, when the first feature is disturbed by noise and the original classification is biased towards line-of-sight occlusion, but the occlusion degree value is definitely in the partial occlusion interval, the result can be corrected to partial occlusion through weight adjustment, so that the final classification result meets both the feature rule and the quantization index, and the robustness and accuracy of the offshore channel occlusion state recognition are improved.

[0149] The offshore channel shielding strength determination method provided in the embodiment comprises the following steps: acquiring offshore frequency modulation channel data at a current sampling moment to construct a channel array comprising a plurality of pulse signals (M>64) and a plurality of time delay domain data points (N>1024 and the real part and the imaginary part are reserved) in each pulse, which not only ensures accurate correspondence between the data and the instantaneous shielding state, but also provides sufficient samples for feature extraction; after eliminating the signal strength difference caused by the distance through normalization preprocessing, a single-layer CNN is used to extract the phase relationship between each time delay domain data point to obtain first features, which avoids excessive processing of low-dimensional data and loss of phase information by multiple convolution layers, and further extraction of high-order spatial features in the subsequent CNN layer can improve the richness and abstraction degree of feature expression; then, the first features are classified and analyzed by relying on an LSTM, which can not only mine the time sequence trend of the phase relationship to realize accurate classification of three types of shielding states, namely, the line-of-sight shielding, the partial shielding and the non-line-of-sight shielding, and fill the gap in the identification of the intermediate shielding state in offshore communication, but also output the shielding degree value based on the distribution function taking only the shielding degree value as the variable to be fitted, thereby solving the problems of large fitting error and the need for a scenario-based critical value of k-factor in the traditional multi-parameter fitting; finally, the classification weight is adjusted by matching the shielding degree value with the category interval to correct the single-dimensional deviation. The offshore channel shielding strength determination method provided in the embodiment not only adapts to the complex shielding environment at sea, but also improves the accuracy and robustness of the shielding strength determination, thereby providing reliable technical support for the adaptive adjustment of communication parameters, the optimization scheduling of system resources and the improvement of link reliability, and effectively meeting the core needs of automatic identification of channel states in offshore communication.

[0150] Corresponding to the foregoing embodiment of the offshore channel shielding strength determination method, the present application further provides an embodiment of an offshore channel shielding strength determination device.

[0151] Figure 2 FIG. 1 is a structural schematic diagram of the embodiment one of the offshore channel shielding strength determination device provided in the present application. As shown in FIG. 1, the device comprises an acquisition module 210, an extraction module 220 and an analysis module 230. Figure 2 The device provided in the embodiment comprises the acquisition module 210, the extraction module 220 and the analysis module 230, wherein,

[0152] The acquisition module 210 is configured to acquire offshore frequency modulation channel data at a current sampling moment to construct a channel array comprising a plurality of time delay domain data points.

[0153] The extraction module 220 is configured to perform feature extraction on the channel array based on a single-layer CNN to acquire the phase relationship between each time delay domain data point in the channel array, thereby obtaining first features corresponding to the channel array.

[0154] The analysis module 230 is configured to perform classification and analysis on the first features based on an LSTM to obtain a shielding degree category and a shielding degree value corresponding to the channel array.

[0155] The LSTM fits a distribution function corresponding to the channel array based on the first feature when analyzing the first feature, wherein only the occlusion degree value in the distribution function is a variable to be fitted.

[0156] The device of the embodiment can be used to execute Figure 1 The steps of the method embodiment are similar to the specific implementation principles and implementation processes, and thus will not be described here.

[0157] The implementation processes of the functions and roles of the units in the device are specifically described in the implementation processes of the corresponding steps in the above method, and thus will not be described here.

[0158] For the device embodiment, the related parts can be understood by referring to the parts of the method embodiment. The device embodiments described above are only illustrative, and the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, that is, they can be located in one place or distributed on multiple network units. According to actual needs, part or all of the modules can be selected to achieve the purposes of the present application. Those skilled in the art can understand and implement without creative labor.

[0159] The above only describes the preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method of determining a sea channel obscuration strength, characterized by, The method comprises: Obtaining the marine frequency modulation channel data at the current sampling time to form a channel array comprising a plurality of time delay domain data points; Performing feature extraction on the channel array based on a single-layer CNN to obtain the phase relationship between each time delay domain data point in the channel array, and to obtain first features corresponding to the channel array; Wherein, the feature extraction on the channel array based on the single-layer CNN comprises: Inputting the channel array into a single convolution layer, and the single convolution layer outputs local spatial features; The local spatial features are inputted into a pooling layer, and the pooling layer performs pooling processing on the local spatial features to output pooled local spatial features; The pooled local spatial features are inputted into an activation layer, and the activation layer processes the local spatial features using an activation function to output two-dimensional spatial features; The two-dimensional spatial features are inputted into a flattening layer, and the flattening layer performs flattening processing on the two-dimensional spatial features to output one-dimensional time sequence features as the first features; After the feature extraction on the channel array based on the single-layer CNN, the method further comprises: Inputting the first features outputted by the single-layer CNN into a subsequent CNN layer to extract high-order spatial features; The high-order spatial features are inputted into a connection layer, and the connection layer performs pooling and flattening on the high-order spatial features to output one-dimensional time sequence features; The one-dimensional time sequence features replace the first features as the corresponding feature information of the channel array; Performing classification and analysis on the first features based on an LSTM to obtain the occlusion degree category and the occlusion degree value corresponding to the channel array; Wherein, when analyzing the first features, the LSTM fits a distribution function corresponding to the channel array based on the first features, wherein only the occlusion degree value is a fitting variable in the distribution function.

2. The method of claim 1, wherein, The channel array only comprises frequency modulation channel data at one sampling time point, one channel array comprises M pulse signals at one sampling time point, each pulse signal contains N time delay domain data points, M is a positive integer greater than 64, and N is a positive integer greater than 1024.

3. The method of claim 1, wherein, The classification on the first features based on the LSTM comprises: Determining target features of the first features and each preset classification label based on the LSTM; Calculating the probability distribution of the first features belonging to each preset classification label based on the target features; Selecting the preset classification label corresponding to the maximum probability value as the classification result of the first features.

4. The method of claim 1, wherein, The analysis on the first features based on the LSTM comprises: Establishing a distribution function according to the first features, the distribution function is the relationship between the first features and the occlusion degree; Fixing the parameters in the distribution function except the occlusion degree according to historical fitting results; Fitting the distribution function according to the first features to obtain the occlusion degree value.

5. The method of claim 1, wherein, Before the feature extraction on the channel array based on the single-layer CNN, it comprises: Traversing all time delay domain data points at one sampling time point to screen the maximum value and the minimum value of the channel array; Subtract the value of each time delay domain data point in the channel array from the minimum value, and divide by the difference between the maximum value and the minimum value to obtain a normalized channel array.

6. The method of claim 1, wherein, Before fitting the distribution function corresponding to the channel array based on the first feature, the method comprises: selecting a plurality of continuous pulse signals as a set of fitting data; determining the occlusion distribution and the direct distribution respectively according to the fitting data; determining the weight coefficients of the occlusion distribution and the direct distribution according to the occlusion intensity; weighting and summing the occlusion distribution and the direct distribution based on the weight coefficients to establish the distribution function.

7. The method of claim 1, wherein, After obtaining the occlusion degree category and the occlusion degree value corresponding to the channel array, the method comprises: setting an occlusion degree value interval corresponding to the occlusion degree category; calculating the matching degree of the occlusion degree value falling into the interval corresponding to the interval; allocating weights of a plurality of preset classification labels in the LSTM based on the matching degree; adjusting the probability distribution value corresponding to each preset classification label of the first feature according to the weights, and adjusting the classification result output by the LSTM.

8. A device for determining the obscuring strength of a maritime channel, characterized in that The device comprises an acquisition module, an extraction module and an analysis module; wherein, the acquisition module is configured to acquire marine frequency modulation channel data at a current sampling time to form a channel array comprising a plurality of time delay domain data points; the extraction module is configured to perform feature extraction on the channel array based on a single-layer CNN to obtain the phase relationship between each time delay domain data point in the channel array, and obtain a first feature corresponding to the channel array; wherein the feature extraction on the channel array based on the single-layer CNN comprises: inputting the channel array into a single convolution layer, and the single convolution layer outputs local spatial features; inputting the local spatial features into a pooling layer, and the pooling layer performs pooling processing on the local spatial features to output pooled local spatial features; inputting the pooled local spatial features into an activation layer, and the activation layer processes the local spatial features using an activation function to output two-dimensional spatial features; inputting the two-dimensional spatial features into a flattening layer, and the flattening layer performs flattening processing on the two-dimensional spatial features to output one-dimensional time sequence features as the first feature; after the feature extraction on the channel array based on the single-layer CNN, the method further comprises: inputting the first feature output by the single-layer CNN into a subsequent CNN layer to extract high-order spatial features; inputting the high-order spatial features into a connection layer, and the connection layer performs pooling and flattening on the high-order spatial features to output one-dimensional time sequence features; replacing the first feature with the one-dimensional time sequence features as the feature information corresponding to the channel array; the analysis module is configured to classify and analyze the first feature based on an LSTM to obtain an occlusion degree category and an occlusion degree value corresponding to the channel array; wherein the LSTM fits a distribution function corresponding to the channel array based on the first feature when analyzing the first feature, and only the occlusion degree value is a to-be-fitted variable in the distribution function.

Images