An electromagnetic wave path loss time sequence prediction method in an evaporation waveguide environment

Abstract

The present application relates to the technical field of evaporation waveguide path loss prediction, and discloses a kind of electromagnetic wave path loss time sequence prediction method under evaporation waveguide environment, comprising: obtaining the time series data of electromagnetic wave historical path loss, and inputting it into path loss prediction model;By bidirectional perception layer, the historical dependence relationship and future context information of time series data are captured using bidirectional long short-term memory network;By semantic alignment layer, non-causal information of future context information is filtered out using gated recurrent unit, and then causal alignment is carried out;Through attention reconstruction layer, the influence degree of each feature vector in the aligned feature sequence on electromagnetic wave signal fading is used to weight feature vector using self-attention mechanism;Through time sequence decoding layer, the time sequence logic of each feature vector in the weighted feature sequence is obtained using post-gated recurrent unit, to predict future path loss according to the weighted feature sequence and the time sequence logic of each feature vector therein.

Description

Technical Field

[0001] This invention relates to the field of evaporation waveguide path loss prediction technology, and in particular to a method, system, device and medium for timing prediction of electromagnetic wave path loss in an evaporation waveguide environment. Background Technology

[0002] Evaporation waveguides are a common atmospheric refraction structure occurring over oceans and marginal seas. Due to water vapor evaporation from the sea surface, the atmospheric humidity near the sea surface decreases sharply with increasing altitude, leading to a negative gradient in the atmospheric refractive index and forming a waveguide layer capable of trapping electromagnetic waves. When electromagnetic waves propagate in an evaporation waveguide environment, their path loss is significantly lower than in a standard atmospheric environment, enabling beyond-line-of-sight propagation. However, dynamic changes in the waveguide environment can also cause severe signal fading and dead zones. Therefore, accurately predicting the electromagnetic wave path loss in an evaporation waveguide environment is crucial for assessing the detection range of maritime radar and optimizing the quality of maritime wireless communication links.

[0003] Currently, path loss prediction mainly relies on deterministic physical models (such as parabolic equations and ray tracing). These methods typically require high-precision real-time sea surface meteorological parameters (such as air temperature, air pressure, relative humidity, and wind speed) or evaporation waveguide height as input. However, in practical engineering applications, these methods have significant limitations. First, marine meteorological sensors are susceptible to salt spray corrosion, wave swaying, and ship turbulence, resulting in high-precision meteorological data that often contains noise or is incomplete. This directly leads to a large deviation between the calculation results of the physical model and the measured values. In addition, the inversion calculation process of the physical model is very complex and difficult to meet real-time requirements.

[0004] In recent years, deep learning methods, represented by recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), have been introduced into path loss prediction, predicting future states by mining the temporal patterns of historical data. However, when processing non-stationary marine electromagnetic environment data, RNNs and LSTMs face technical bottlenecks, failing to simultaneously capture both "global evolution trend" and "focus on local abrupt change features (deep fading)," thus hindering accurate temporal prediction of electromagnetic wave path loss in evaporating waveguide environments. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, device and medium for predicting electromagnetic wave path loss in an evaporating waveguide environment, which can solve the problem that existing technologies cannot achieve accurate timing prediction.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a method for predicting electromagnetic wave path loss timing in an evaporating waveguide environment, comprising the following steps: Obtain time-series data of historical path loss of electromagnetic waves in an evaporating waveguide environment; Time series data is input into the path loss prediction model, which includes a bidirectional perception layer, a semantic alignment layer, an attention reconstruction layer, and a temporal decoding layer. By using the bidirectional perception layer to capture the historical dependencies and future context information of time series data through the two independent chains of the bidirectional long short-term memory network (forward and backward), a bidirectional hidden state sequence is obtained. The semantic alignment layer uses the reset and update gates of the gated recurrent unit to filter out non-causal information of future context information in the bidirectional hidden state sequence, and performs causal alignment on the filtered bidirectional hidden state sequence to obtain the aligned feature sequence. The attention reconstruction layer uses a self-attention mechanism to weight the feature vectors based on the degree of influence of each feature vector in the aligned feature sequence on the fading of the electromagnetic wave signal, thus obtaining a weighted feature sequence. The timing logic of each feature vector in the weighted feature sequence is obtained by using a post-gated cyclic unit through a timing decoding layer. Based on the weighted feature sequence and the timing logic of each feature vector, the future path loss of electromagnetic waves can be predicted.

[0007] Furthermore, the path loss prediction model is obtained through iterative training in the following manner: In each iteration, based on the Bayesian hyperparameter optimizer, the path loss prediction model is searched for hyperparameters in the current iteration by using the expectation-lift acquisition function to search for the posterior distribution between the objective function fitted with a Gaussian process as a surrogate model and the hyperparameters of the path loss prediction model. The path loss prediction model is then trained using the searched hyperparameters. The objective function is constructed with the goal of minimizing the root mean square error between the predicted and actual values ​​of the path loss prediction model.

[0008] Furthermore, the attention reconstruction layer utilizes a self-attention mechanism to weight the feature vectors based on the degree of influence of each feature vector in the aligned feature sequence on the fading of the electromagnetic wave signal, obtaining a weighted feature sequence, including: The query vector of the aligned feature sequence is calculated using a self-attention mechanism through an attention reconstruction layer. Key vector Sum value vector ; Calculate the attention weight matrix using a scaled dot product model : In the formula, For feature dimensions; Obtain the weighted feature sequence .

[0009] Furthermore, the step of obtaining the temporal logic of each feature vector in the weighted feature sequence through a post-gated loop unit via a temporal decoding layer, and predicting the future path loss of electromagnetic waves based on the weighted feature sequence and the temporal logic of each feature vector therein, includes: The timing decoding layer uses a post-gated loop unit to read data sequentially over time. The read features are then compressed into the hidden state of the last time step. ; Through the fully connected layer Mapped to the final predicted output vector To predict the output vector Predicting the future path loss of electromagnetic waves; where, and For the weights and biases of the fully connected layer, This is the activation function.

[0010] Furthermore, after obtaining the corresponding output results, the bidirectional perception layer, the semantic alignment layer, and the temporal decoding layer first perform dropout regularization on the output results before inputting them into the next layer; After obtaining the corresponding output result, the attention reconstruction layer first performs layer normalization on the output result, and then inputs it into the next layer.

[0011] Embodiments of the present invention also provide an electromagnetic wave path loss timing prediction system in an evaporating waveguide environment, comprising: The data acquisition module is used to acquire time-series data of historical path loss of electromagnetic waves in an evaporating waveguide environment; The model prediction module is used to input time series data into the path loss prediction model, which includes a bidirectional perception layer, a semantic alignment layer, an attention reconstruction layer, and a temporal decoding layer. By using the bidirectional perception layer to capture the historical dependencies and future context information of time series data through the two independent chains of the bidirectional long short-term memory network (forward and backward), a bidirectional hidden state sequence is obtained. The semantic alignment layer uses the reset and update gates of the gated recurrent unit to filter out non-causal information of future context information in the bidirectional hidden state sequence, and performs causal alignment on the filtered bidirectional hidden state sequence to obtain the aligned feature sequence. The attention reconstruction layer uses a self-attention mechanism to weight the feature vectors based on the degree of influence of each feature vector in the aligned feature sequence on the fading of the electromagnetic wave signal, thus obtaining a weighted feature sequence. The timing logic of each feature vector in the weighted feature sequence is obtained by using a post-gated cyclic unit through a timing decoding layer. Based on the weighted feature sequence and the timing logic of each feature vector, the future path loss of electromagnetic waves can be predicted.

[0012] Embodiments of the present invention also provide a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described electromagnetic wave path loss timing prediction method in an evaporating waveguide environment.

[0013] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for predicting electromagnetic wave path loss timing in an evaporating waveguide environment.

[0014] The electromagnetic wave path loss timing prediction method in an evaporating waveguide environment provided by this invention has at least the following beneficial effects: This invention designs a path loss prediction model with a deep serial collaborative architecture of "panoramic perception-semantic alignment-saliency reconstruction-temporal decoding". Panoramic perception uses a bidirectional long short-term memory network to capture the historical dependencies and future context information (i.e., global temporal dependencies, capturing global evolution trends) of historical path loss time series data of electromagnetic waves. Semantic alignment uses a gated recurrent unit to filter out non-causal information of future context information, such as interference of future information on the current moment, and performs causal alignment to make it conform to the temporal causal order. Saliency reconstruction uses a self-attention mechanism to weight the key points that affect path loss, i.e. the feature vectors that affect the fading of electromagnetic wave signals (focusing on the local mutation features of instantaneous and severe fading). Finally, temporal decoding uses a post-gated recurrent unit to ensure the temporal logic of the features.

[0015] As can be seen, this invention simultaneously considers both "capturing the global evolution trend of electromagnetic wave historical path loss time series data" and "focusing on local mutation features." Furthermore, it takes into account that during local mutation feature focusing, the self-attention mechanism may disrupt the causal temporal correlation between the local mutation feature and its occurrence time, making it difficult for the model to accurately predict the recovery trend of the electromagnetic wave signal after deep fading. Therefore, after capturing the global evolution trend and before focusing on local mutation features, non-causal information is filtered and causal temporal logic is aligned. After focusing on local mutation features, its temporal logic is also referenced, effectively improving the accuracy of subsequent electromagnetic wave path loss time series prediction. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This invention provides a flowchart illustrating a method for predicting electromagnetic wave path loss timing in an evaporating waveguide environment. Figure 2 This invention provides a schematic diagram comparing the actual value and the predicted value of a model prediction. Figure 3 This is a schematic diagram of the error distribution of a model prediction result provided by the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0018] Currently, path loss prediction methods based on standard Long Short-Term Memory (LSTM) networks or single gated recurrent units (GRUs) are commonly used. This existing approach first performs simple normalization on the collected historical path loss data or meteorological data; then, it constructs a neural network model containing an input layer, a unidirectional LSTM (or GRU) hidden layer, and a fully connected output layer; next, it sets the model's hyperparameters based on human experience; finally, iteratively trains the model using training data, and uses the trained model to perform regression prediction of path loss at future times.

[0019] However, existing technical solutions have significant drawbacks in practical applications. Due to the complex and variable marine evaporation waveguide environment, existing solutions typically use fixed or grid-search-based methods to set hyperparameters. This makes it difficult for manually fixed parameters to adapt to the differences in data distribution across different sea areas and seasons, easily leading to a lack of environmental adaptability in the model and problems such as local optima or non-convergence. Furthermore, standard LSTM or GRU networks only have unidirectional temporal mobility, utilizing only historical information prior to the current moment and failing to perceive the "backward" contextual information of the input sequence, resulting in an incomplete modeling of waveguide change trends. In addition, existing solutions treat all data points within the input time series window equally, lacking a key information weighting mechanism, and cannot distinguish which data points are "key features" (such as waveguide abrupt changes) from "invalid noise." Consequently, when faced with long input sequences, prediction accuracy often decreases with increasing sequence length, making it difficult to accurately capture the deep fading characteristics of the signal.

[0020] To address the common technical bottlenecks in existing evaporation waveguide path loss prediction techniques, such as insufficient sensitivity to capturing long-sequence deep fading features, easy loss of temporal logic during dimensionality reduction, and difficulty in adapting model hyperparameters to complex sea conditions, this invention aims to provide a prediction method based on Bayesian optimization and a four-level hybrid deep collaborative architecture.

[0021] The core objective of this invention is to overcome the limitations of traditional single-model or simple parallel models in processing non-stationary marine electromagnetic environment data, and to construct a deep serial collaborative architecture encompassing "panoramic perception, semantic alignment, saliency reconstruction, and temporal decoding." This method aims to accurately infer future signal change trends based solely on historical loss data without the need for meteorological parameters. In particular, by introducing a deep coupling between a pre-gating mechanism and a self-attention mechanism, this invention addresses the challenge of smoothing abrupt signal changes in long sequences. Furthermore, it utilizes a post-temporal decoding mechanism to replace traditional static pooling operations, thereby compressing feature dimensions while fully preserving the causal logic of path loss evolution, ultimately achieving high-precision multi-step prediction of over-the-horizon propagation loss at sea.

[0022] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] One embodiment of the present invention relates to a method for predicting the timing of electromagnetic wave path loss in an evaporating waveguide environment. The specific process of the method for predicting the timing of electromagnetic wave path loss in an evaporating waveguide environment in this embodiment can be as follows: Figure 1 As shown, it includes: Step 101: Obtain time series data of electromagnetic wave historical path loss under evaporation waveguide environment.

[0024] Step 102: Input the time series data into the path loss prediction model. The path loss prediction model includes a bidirectional perception layer, a semantic alignment layer, an attention reconstruction layer, and a temporal decoding layer.

[0025] Step 103: By using the bidirectional perception layer, the historical dependencies and future context information of the time series data are captured by the two independent chains of the bidirectional long short-term memory network (forward and backward) to obtain the bidirectional hidden state sequence.

[0026] Step 104: The non-causal information of future context information in the bidirectional hidden state sequence is filtered out by the reset gate and update gate of the gated recurrent unit through the semantic alignment layer, and the filtered bidirectional hidden state sequence is causally aligned to obtain the aligned feature sequence.

[0027] Step 105: Using the attention reconstruction layer and the self-attention mechanism, the feature vectors are weighted according to the degree of influence of each feature vector in the aligned feature sequence on the fading of the electromagnetic wave signal, and the weighted feature sequence is obtained.

[0028] Step 106: The timing logic of each feature vector in the weighted feature sequence is obtained by using the post-gated loop unit through the timing decoding layer, so as to predict the future path loss of electromagnetic waves based on the weighted feature sequence and the timing logic of each feature vector.

[0029] The following is a detailed description of the implementation details of the electromagnetic wave path loss timing prediction method in the evaporating waveguide environment of this embodiment. The following content is only for the convenience of understanding and is not necessary for implementing this solution.

[0030] This embodiment designs a serial network structure of "bidirectional perception-semantic alignment-attention reconstruction-temporal decoding". By training this network structure, a four-level serial hybrid deep collaborative prediction model (i.e. path loss prediction model) is obtained.

[0031] To achieve time-series prediction of electromagnetic wave path loss in an evaporating waveguide environment, measured time-series data of electromagnetic wave path loss are used as the input sequence. The input is a bidirectional perceptron layer, which uses a bidirectional long short-term memory network (BiLSTM) to capture historical dependencies and future context information of time series data through two independent forward and backward chains, thereby obtaining the long-range dependencies and periodic features of the sequence, and outputting a bidirectional hidden state sequence. .

[0032] To filter out non-causal noise in bidirectional features (bidirectional hidden state sequences), The input semantic alignment layer utilizes the update gate of the Intermediate GRU (Gated Recurrent Unit) layer based on sequence patterns. and reset door Perform causal alignment and cleaning on the features, and output the aligned feature sequence. .

[0033] Introducing an attention reconstruction layer, utilizing the self-attention mechanism to... Perform saliency reconstruction: Calculate the query vector Key vector Sum value vector (All by) (Obtained through linear transformation), the attention weight matrix is ​​calculated using a scaled dot product model. : In the formula, For feature dimensions; obtain the weighted feature sequence. .

[0034] This layer automatically identifies key mutation points in the sequence and assigns them high weights, outputting a weighted feature sequence. .

[0035] The timing decoding layer utilizes a Last GRU (Last GRU) layer with "Last" output mode as the timing decoder, replacing the traditional global average pooling layer. This layer reads data sequentially over time. The temporal logic of a variable-length sequence is compressed to the hidden state of the last time step. Through a fully connected layer Mapped to the final predicted output vector : To predict the output vector Time-series prediction of electromagnetic wave path loss; where, and For the weights and biases of the fully connected layer, This is the activation function.

[0036] In the specific implementation, after obtaining the corresponding output results, the bidirectional perception layer, semantic alignment layer, and temporal decoding layer first perform dropout regularization on the output results before inputting them into the next layer; after obtaining the corresponding output results, the attention reconstruction layer first performs layer normalization on the output results before inputting them into the next layer.

[0037] The specific four-level serial collaborative architecture described in this embodiment overcomes the inherent defects of conventional deep learning methods in processing evaporating waveguide nonlinear data, and achieves a substantial improvement in model performance, specifically in the following aspects: 1. This invention overcomes the technical contradiction in long sequence prediction where it is difficult to simultaneously ensure the salience of abrupt changes and the logical consistency of temporal evolution. In existing path loss prediction research, simply introducing an attention mechanism can capture deep fading points, but in the subsequent feature reduction process, traditional global average pooling (GAP) operations often lead to the collapse of temporal information, meaning the model loses the specific temporal location information of the abrupt changes. This invention solves this problem through a specific collaborative structure of "Attention (highlighting key points) + post-GRU (logical concatenation)". The Attention layer first "amplifies" the weak precursor signals buried in the long sequence, and the subsequent post-GRU layer uses a gating mechanism to strictly read and memorize these amplified key features in chronological order. This "highlight first, then decode sequentially" processing mechanism allows the model to retain the intensity of deep fading features while fully maintaining the physical causal chain of waveguide changes, significantly improving the model's ability to predict future signal recovery or continued fading trends.

[0038] 2. This invention addresses the challenges of information flow redundancy and parameter oscillation in hybrid deep networks. Given the high noise and strong non-stationarity of evaporative waveguide data, directly connecting BiLSTM and Attention often leads to attention dispersion and difficulty in focusing on core features. This invention embeds a pre-GRU layer between the two as a "semantic cleaner," constructing a funnel-shaped feature processing pipeline of "wide (BiLSTM) - narrow (GRU) - strong (Attention)." The pre-GRU layer effectively filters out non-causal background noise generated by BiLSTM, providing a high-quality query vector for the attention layer, thereby significantly improving the accuracy of attention weight allocation. This deep pipeline design ensures that the features input to the decision layer have both a global perspective and strict causal consistency, effectively improving the convergence speed and stability of the Bayesian optimizer during hyperparameter search.

[0039] 3. Adaptive deep adaptation of complex heterogeneous network architecture to marine environmental characteristics is achieved. The parameter space of the proposed BiLSTM-GRU-Attention-GRU four-level heterogeneous network is extremely complex. The impact of node number matching and regularization strength between layers on prediction performance is highly nonlinear, making it extremely difficult to find the optimal solution through manual trial and error. This invention constructs a Bayesian joint search space for this specific heterogeneous network and uses Gaussian process regression to automatically perceive the optimal balance between "feature extraction capability" (determined by BiLSTM / pre-GRU) and "temporal decoding capability" (determined by post-GRU). This enables the model to automatically adjust the capacity configuration of each layer of the network according to the waveguide data characteristics of different sea areas and seasons, thereby achieving adaptive high-precision modeling of path loss variations under various complex sea conditions without manual intervention.

[0040] In some embodiments, the path loss prediction model is iteratively trained as follows: In each iteration, based on a Bayesian hyperparameter optimizer, the hyperparameters of the path loss prediction model are searched using an expectation-boosting acquisition function by means of the posterior distribution between the objective function fitted with a Gaussian process as a surrogate model and the hyperparameters of the path loss prediction model. The path loss prediction model is then trained using the searched hyperparameters. The objective function is constructed with the goal of minimizing the root mean square error between the predicted and actual values ​​of the path loss prediction model, as shown below: ; In the formula, For hyperparameters The objective function value under the configuration, i.e., the root mean square error of the model on the validation set; To verify the total number of samples in the set; In hyperparameters Under this configuration, the path loss prediction model predicts the i-th sample in the validation set. To verify the true value of the measured path loss corresponding to the i-th sample in the set; This represents the combination of hyperparameters searched by the Bayesian hyperparameter optimizer in the current iteration, including the initial learning rate, the number of hidden layer neurons, and the regularization coefficient.

[0041] In one embodiment, this invention proposes a multi-step temporal prediction method for evaporating waveguide path loss based on Bayesian optimization and a hybrid attention mechanism. First, the acquired long-term measured path loss data is hierarchically partitioned and reconstructed using a sliding window to build a supervised learning sample set. A Bayesian optimizer is constructed using Gaussian process regression to search for the optimal network configuration in the joint hyperparameter space. A four-level serial deep collaborative model, including a bidirectional long short-term memory network, gated recurrent units, and a self-attention mechanism, is constructed and retrained with the full dataset to obtain the final prediction model. The real-time observed path loss sequence is input into this model, which outputs predicted path loss values ​​for the next several steps.

[0042] In a specific embodiment, taking measured data from a microwave over-the-horizon propagation link in a certain sea area as an example, the specific implementation steps of the electromagnetic wave path loss timing prediction method in the evaporating waveguide environment of the present invention are illustrated as follows: Step 1: Data preprocessing and construction of hierarchical nested sample sets: Continuous measured path loss time series data (for two consecutive months) was obtained by constructing a microwave propagation link. ,in, This represents the total sample length. Assuming a sampling interval of 1, the total sample length is... There are 1253 points. To eliminate the influence of dimensions and accelerate model convergence, the max-min normalization method is used to map the original data to... Interval: ; In the formula, For the normalized first The path loss value at each time point. and These are the minimum and maximum values ​​in the training set, respectively.

[0043] Subsequently, a hierarchical nested strategy was used to partition the dataset: first, the end of the sequence was truncated. Each data point is used as an independent test set. Used for final model performance evaluation; remaining data The dataset is further divided into an optimized training set and an optimized validation set. A supervised learning sample matrix is ​​constructed using the sliding window technique, with the input window length set to [value missing]. The predicted step size is For time Construct the input feature vector and the corresponding label vector : ; ; Construct the final sample set ,in, The total number of samples.

[0044] Set input window length (That is, using data from the past 12 hours) predict the step size (That is, predicting the loss in the next 2 hours). For time... Construct the input feature sequence and the corresponding tag value The dataset was divided into a training set (first 80%) and a test set (last 20%) in chronological order.

[0045] Step 2: Construct a Bayesian hyperparameter optimizer based on Gaussian process regression: To address the challenge of manually fitting hyperparameters in hybrid deep networks, a Bayesian optimization framework is established. A joint optimization variable space is defined. ,in, The initial learning rate (the recommended search range is...) ), The number of neurons in the hidden layer (the recommended search range is...) ), Here is the L2 regularization coefficient. A Gaussian process (GP) is used as a surrogate model to fit the objective function. The posterior distribution between the hyperparameters and the target function is the root mean square error (RMSE) on the validation set.

[0046] The search process is guided by the Expected Improvement (EI) function, which selects the hyperparameter combination that maximizes the expected improvement in each iteration. The evaluation continues until the preset number of iterations is reached, at which point the globally optimal combination of hyperparameters is output. Assuming that after 25 iterations, this embodiment obtains the globally optimal combination of hyperparameters. For: Initial learning rate Number of neurons in the hidden layer L2 regularization coefficient .

[0047] Step 3: Construct a four-level serial hybrid deep collaborative prediction model: Using the optimal hyperparameters obtained in step 2 A path loss prediction model is constructed. This model, as described in the above embodiment, employs a deep sequential architecture of "panoramic perception - semantic alignment - saliency reconstruction - temporal decoding," with the specific structural configuration as follows: Bidirectional Perception Layer: Employs a Bidirectional Long Short-Term Memory (BiLSTM) network with 64 hidden units and a sequence output mode. This layer captures the input sequence using two independent chains, one forward and one backward. The system considers historical dependencies and future contexts, outputting a bidirectional hidden state sequence. A Dropout layer with a dropout rate of 0.2 is added after each layer to prevent overfitting.

[0048] Semantic alignment layer: A pre-gated gated recurrent unit (GRU) is used, with 64 hidden units and sequential output. Update and reset gates are used to filter non-causal noise in bidirectional features, achieving causal alignment of features. A Dropout layer with a dropout rate of 0.2 is also connected after this layer.

[0049] Attention Reconstruction Layer: Employs a multi-head self-attention mechanism with two heads. Calculate the query vector Q, key vector K, and value vector V of the aligned feature sequence, and use a scaled dot product model to compute the attention weight matrix (see the above embodiment for the formula). The feature vectors are weighted according to the weights, highlighting key moments that affect deep signal fading. A post-layer connection normalization layer (LayerNorm) stabilizes the data distribution.

[0050] The temporal decoding layer employs a Last GRU (Last GRU) with 64 hidden units and an output mode set to Last time step. This layer reads the weighted feature sequence in chronological order, compresses and maps it to the hidden state of the last time step. Preserve the evolution logic. Finally, a fully connected layer will be used to... The linear mapping is used to predict the output vector (see the above example for the formula).

[0051] Step 4: Full Retraining and Region Multi-Step Prediction: Using the globally optimal hyperparameters obtained in step 2 Initialize the hybrid network model constructed in step 3. Merge the optimized training set and optimized validation set to construct a full training set for final iterative training of the model, using the Adam optimizer to update the network weights. After training, input the latest observation sequence into the model to obtain the normalized prediction result, and then use the inverse normalization formula to restore it to the true path loss value: ; In the formula, This is the normalized predicted value output by the path loss prediction model.

[0052] This enables accurate multi-step prediction of the path loss trend of the evaporation waveguide at future moments.

[0053] The predictive performance of the model (BiLSTM-GRU-Attention(Opt)) described in this invention is evaluated below using test set data: Mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), mean relative percentage error (MAPE), and coefficient of determination (R²) are used as evaluation indicators. The statistical results of the prediction error in this embodiment are shown in Table 1. Table 1. Statistics of Model Prediction Errors

[0054] As can be seen from Table 1, the model of this invention achieved a low error level on the test set, with a determination coefficient (R²) of 78.99%, indicating that the model can explain most of the variance of path loss variation and has high prediction accuracy.

[0055] To visually demonstrate the prediction results, Figure 2 The curves comparing the actual and predicted values ​​of some test set samples are presented. It can be seen that the prediction curve (solid line) generated by the proposed prediction model highly overlaps with the actual observed curve (dashed line). Especially in the intervals where path loss fluctuates drastically, thanks to the attention mechanism's weighting of abrupt changes and the post-GRU's preservation of temporal logic, the model can accurately track the instantaneous trend of path loss changes without significant lag or smoothing.

[0056] also, Figure 3 The error distribution of the prediction results is shown. It can be seen that the prediction error amplitude for the vast majority of test samples is small, and the error distribution is uniform. This further verifies that the four-level collaborative architecture has good robustness and generalization ability when processing non-stationary marine environmental data, and can meet the practical engineering needs of electromagnetic wave path loss time series prediction in evaporative waveguide environments.

[0057] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the protection scope of this invention. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, without changing the core design of the algorithm and process, are also within the protection scope of this invention.

[0058] Another embodiment of the present invention relates to an electromagnetic wave path loss timing prediction system in an evaporating waveguide environment. The implementation details of this embodiment's electromagnetic wave path loss timing prediction system in an evaporating waveguide environment are described below. The following details are provided for ease of understanding and are not essential for implementing this solution. The electromagnetic wave path loss timing prediction system in this embodiment in an evaporating waveguide environment includes: The data acquisition module is used to acquire time-series data of historical path loss of electromagnetic waves in an evaporating waveguide environment; The model prediction module is used to input time series data into the path loss prediction model, which includes a bidirectional perception layer, a semantic alignment layer, an attention reconstruction layer, and a temporal decoding layer. By using the bidirectional perception layer to capture the historical dependencies and future context information of time series data through the two independent chains of the bidirectional long short-term memory network (forward and backward), a bidirectional hidden state sequence is obtained. The semantic alignment layer uses the reset and update gates of the gated recurrent unit to filter out non-causal information of future context information in the bidirectional hidden state sequence, and performs causal alignment on the filtered bidirectional hidden state sequence to obtain the aligned feature sequence. The attention reconstruction layer uses a self-attention mechanism to weight the feature vectors based on the degree of influence of each feature vector in the aligned feature sequence on the fading of the electromagnetic wave signal, thus obtaining a weighted feature sequence. The timing logic of each feature vector in the weighted feature sequence is obtained by using a post-gated cyclic unit through a timing decoding layer. Based on the weighted feature sequence and the timing logic of each feature vector, the future path loss of electromagnetic waves can be predicted.

[0059] It is not difficult to see that this embodiment is a system embodiment corresponding to the above method embodiments, and this embodiment can be implemented in conjunction with the above method embodiments. The relevant technical details and technical effects mentioned in the above embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiments.

[0060] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this invention, this embodiment does not introduce units that are not closely related to solving the technical problem proposed by this invention; however, this does not mean that other units are absent from this embodiment.

[0061] Another embodiment of the present invention relates to a computer device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the electromagnetic wave path loss timing prediction method in the evaporating waveguide environment described in the above embodiments.

[0062] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0063] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0064] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0065] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0066] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing the present invention, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A method for predicting electromagnetic wave path loss timing in an evaporating waveguide environment, characterized in that, The method includes: Obtain time-series data of historical path loss of electromagnetic waves in an evaporating waveguide environment; Time series data is input into the path loss prediction model, which includes a bidirectional perception layer, a semantic alignment layer, an attention reconstruction layer, and a temporal decoding layer. By using the bidirectional perception layer to capture the historical dependencies and future context information of time series data through the two independent chains of the bidirectional long short-term memory network (forward and backward), a bidirectional hidden state sequence is obtained. The semantic alignment layer uses the reset and update gates of the gated recurrent unit to filter out non-causal information of future context information in the bidirectional hidden state sequence, and performs causal alignment on the filtered bidirectional hidden state sequence to obtain the aligned feature sequence. The attention reconstruction layer uses a self-attention mechanism to weight the feature vectors based on the degree of influence of each feature vector in the aligned feature sequence on the fading of the electromagnetic wave signal, thus obtaining a weighted feature sequence. The timing logic of each feature vector in the weighted feature sequence is obtained by using a post-gated cyclic unit through a timing decoding layer. Based on the weighted feature sequence and the timing logic of each feature vector, the future path loss of electromagnetic waves can be predicted.

2. The method for predicting electromagnetic wave path loss timing in an evaporating waveguide environment according to claim 1, characterized in that, The path loss prediction model is obtained through iterative training in the following manner: In each iteration, based on the Bayesian hyperparameter optimizer, the path loss prediction model is searched for hyperparameters in the current iteration by using the expectation-lift acquisition function to search for the posterior distribution between the objective function fitted with a Gaussian process as a surrogate model and the hyperparameters of the path loss prediction model. The path loss prediction model is then trained using the searched hyperparameters. The objective function is constructed with the goal of minimizing the root mean square error between the predicted and actual values ​​of the path loss prediction model.

3. The method for predicting electromagnetic wave path loss timing in an evaporating waveguide environment according to claim 1, characterized in that, The attention reconstruction layer utilizes a self-attention mechanism to weight the feature vectors based on the degree of influence of each feature vector in the aligned feature sequence on the fading of the electromagnetic wave signal, resulting in a weighted feature sequence, including: The query vector of the aligned feature sequence is calculated using a self-attention mechanism through an attention reconstruction layer. Key vector Sum value vector ; Calculate the attention weight matrix using a scaled dot product model : In the formula, For feature dimensions; Obtain the weighted feature sequence .

4. The method for predicting electromagnetic wave path loss timing in an evaporating waveguide environment according to claim 3, characterized in that, The step of obtaining the temporal logic of each feature vector in the weighted feature sequence through a post-gated loop unit via a temporal decoding layer, and predicting the future path loss of electromagnetic waves based on the weighted feature sequence and the temporal logic of each feature vector therein, includes: The timing decoding layer uses a post-gated loop unit to read data sequentially over time. The read features are then compressed into the hidden state of the last time step. ; Through the fully connected layer Mapped to the final predicted output vector To predict the output vector Predicting the future path loss of electromagnetic waves; where, and For the weights and biases of the fully connected layer, This is the activation function.

5. The method for predicting electromagnetic wave path loss timing in an evaporating waveguide environment according to claim 1, characterized in that, After obtaining the corresponding output results, the bidirectional perception layer, the semantic alignment layer, and the temporal decoding layer first perform discard regularization on the output results before inputting them into the next layer. After obtaining the corresponding output result, the attention reconstruction layer first performs layer normalization on the output result, and then inputs it into the next layer.

6. A timing prediction system for electromagnetic wave path loss in an evaporating waveguide environment, characterized in that, The system includes: The data acquisition module is used to acquire time-series data of historical path loss of electromagnetic waves in an evaporating waveguide environment; The model prediction module is used to input time series data into the path loss prediction model, which includes a bidirectional perception layer, a semantic alignment layer, an attention reconstruction layer, and a temporal decoding layer. By using the bidirectional perception layer to capture the historical dependencies and future context information of time series data through the two independent chains of the bidirectional long short-term memory network (forward and backward), a bidirectional hidden state sequence is obtained. The semantic alignment layer uses the reset and update gates of the gated recurrent unit to filter out non-causal information of future context information in the bidirectional hidden state sequence, and performs causal alignment on the filtered bidirectional hidden state sequence to obtain the aligned feature sequence. The attention reconstruction layer uses a self-attention mechanism to weight the feature vectors based on the degree of influence of each feature vector in the aligned feature sequence on the fading of the electromagnetic wave signal, thus obtaining a weighted feature sequence. The timing logic of each feature vector in the weighted feature sequence is obtained by using a post-gated cyclic unit through a timing decoding layer. Based on the weighted feature sequence and the timing logic of each feature vector, the future path loss of electromagnetic waves can be predicted.

7. A computer device, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the electromagnetic wave path loss timing prediction method in an evaporating waveguide environment as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the electromagnetic wave path loss timing prediction method in the evaporating waveguide environment as described in any one of claims 1 to 5.

Images