A rainfall estimation method, system, computer device and storage medium based on a microwave communication link

Abstract

A rainfall estimation method based on a microwave communication link comprises the following steps: constructing multiple microwave communication links, and obtaining received signal levels of the microwave communication links during rainfall; preprocessing the obtained received signal levels to generate an original signal sequence; inputting the original signal sequence into a pre-trained rainfall estimation model for processing to obtain rainfall intensity; the rainfall estimation model is constructed based on a Transformer model and comprises a position encoding module for injecting position information into the original signal sequence, a multi-layer encoder for extracting multi-scale time sequence features and long-term rainfall trends from the original signal sequence, a bidirectional gated recurrent unit for dynamically gating and adjusting the multi-scale time sequence features to obtain short-term rainfall fluctuations, and a data output module for outputting the rainfall intensity. The present application does not rely on the characteristics of regional rainfall and can be applied to rainfall monitoring in different regions, different link lengths and different geographical condition scenarios.

Description

Technical Field

[0001] This invention relates to the field of rainfall monitoring technology, specifically a rainfall estimation method, system, computer equipment, and storage medium based on a microwave communication link. Background Technology

[0002] With the increasing frequency of extreme weather events, strengthening cities' ability to prevent natural disasters caused by extreme weather has become crucial. Traditional methods for rainfall detection have many limitations. For example, while rain gauges are widely used, they are limited by high equipment costs, limited spatiotemporal resolution, and uneven regional coverage. Weather radar and satellite remote sensing can provide wide-area rainfall observations, but their accuracy in monitoring near-surface rainfall is limited.

[0003] With the rapid development of communication technology, the technology of retrieving rainfall intensity based on commercial microwave communication links has gradually solved the shortcomings of the aforementioned traditional methods. The core principle of this technology is to deduce the corresponding rainfall intensity by analyzing the attenuation changes of microwave signals caused by rainfall during atmospheric propagation.

[0004] Currently, the power-law model used by the International Telecommunication Union (ITU) to describe the relationship between signal attenuation and rainfall intensity is widely adopted due to its high computational efficiency. However, the core parameters of this model are highly sensitive to regional precipitation characteristics, making it difficult to achieve universality across different link lengths and complex geographical environments.

[0005] Chinese patent document CN118042309A discloses a water vapor monitoring method and a rainfall trend prediction method based on a microwave communication link. It also uses a model provided by the International Telecommunication Union (ITU), and the inversion algorithm is as follows: Where Aw is the first attenuation data or the second attenuation data, f is the communication frequency of the intra-area communication link or the inter-area communication link, p is the air pressure in the rain area, T is the temperature in the rain area, and e is the water vapor density.

[0006] The above methods also have the problem of being difficult to apply universally across different link lengths and complex geographical environments. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a rainfall estimation method, system, computer equipment, and storage medium based on microwave communication links. This method is independent of regional rainfall characteristics and is applicable to rainfall monitoring in different regions, with different link lengths, and under different geographical conditions.

[0008] To achieve the above objectives, the specific solution adopted by the present invention is as follows: A rainfall estimation method based on a microwave communication link includes the following steps: Multiple microwave communication links were constructed, and the received signal levels of the microwave communication links were obtained during rainfall. The received signal level is preprocessed to generate the original signal sequence; The original signal sequence is input into a pre-trained rainfall estimation model for processing to obtain the rainfall intensity. The rainfall estimation model is based on the Transformer model and includes a location encoding module for injecting location information into the original signal sequence, a multi-layer encoder for extracting multi-scale temporal features and long-term rainfall trends from the original signal sequence, a bidirectional gated loop unit for dynamically gating the multi-scale temporal features to obtain short-term rainfall fluctuations, and a data output module for outputting rainfall intensity.

[0009] As a further optimization of the above-mentioned rainfall estimation method based on microwave communication links, the method by which the location coding module injects location information into the original signal sequence includes: ; Where X is the time step in the original signal sequence. The time step after the location information is injected. It contains location information, and has ,in The total length of the original signal sequence. It is the feature dimension of the time step.

[0010] As a further optimization of the above-mentioned rainfall estimation method based on microwave communication links: the multilayer encoder includes multiple linearly arranged encoders, each encoder including a multi-head self-attention block and a feedforward network arranged sequentially, the feedforward network including two linear transformation layers and an activation function.

[0011] As a further optimization of the above-mentioned rainfall estimation method based on microwave communication links: the bidirectional gated loop unit includes a forward gated loop unit and a reverse gated loop unit. The forward gated loop unit processes the first input sequence in a first order, and the reverse gated loop unit processes the first input sequence in a second order. The first input sequence is output by the multilayer encoder, with the first order being the reverse of the second order.

[0012] As a further optimization of the above-mentioned rainfall estimation method based on microwave communication links: both the forward-gated loop unit and the reverse-gated loop unit include an update gate and a reset gate, and the processing procedures of the forward-gated loop unit and the reverse-gated loop unit are as follows: ; ; ; ; in, To update the gate output, To reset the gate's output, This is the weight matrix. For bias vectors, This represents the candidate hidden layer state. This is the output at state t-1.

[0013] As a further optimization of the above-mentioned rainfall estimation method based on microwave communication link: the data output module includes an attention pooling layer and a fully connected layer arranged in sequence. The attention pooling layer is used to transform the second input sequence, which is the output of the bidirectional gated recurrent unit.

[0014] As a further optimization of the above-mentioned rainfall estimation method based on microwave communication links, the method also includes the following steps: The rainfall estimation model is trained by backpropagation based on the rainfall intensity to optimize the location encoding module.

[0015] A rainfall estimation system based on a microwave communication link, used to implement the above-mentioned rainfall estimation method based on a microwave communication link, the system comprising: The microwave communication subsystem is used to construct multiple microwave communication links and acquire the received signal levels of the microwave communication links during rainfall. The data acquisition module is used to preprocess the received signal levels to generate the original signal sequence; The model execution module is used to run a pre-trained rainfall estimation model to process the original signal sequence and obtain the rainfall intensity.

[0016] Computer equipment, including: Memory, used to store computer programs; A processor is used to read and execute the computer program to implement the above-described rainfall estimation method based on a microwave communication link.

[0017] A storage medium for storing a computer program that, when executed, implements the aforementioned rainfall estimation method based on a microwave communication link.

[0018] Beneficial Effects: This invention optimizes the method for retrieving rainfall intensity from microwave communication links using deep learning technology. It is independent of regional rainfall characteristics and applicable to rainfall monitoring in different regions, with varying link lengths and geographical conditions. Furthermore, this invention introduces a bidirectional gated recurrent unit (GRU) on top of the Transformer model. The Transformer learns long-distance relationships within the sequence, while the GRU captures short-distance dynamics. The complementary advantages of both significantly improve the model's accuracy in predicting small to medium-sized datasets. Additionally, this invention introduces a custom, learnable position encoding module into the Transformer model. This module dynamically adjusts the position encoding representation of the input sequence through backpropagation, more accurately characterizing the relevant features of positions within the sequence, thus improving the model's flexibility and generalization ability. Finally, this invention introduces an attention pooling layer after the GRU, which adaptively assigns weights to each position in the input sequence, automatically learns key time step information, and assigns it higher weights, thereby improving the model's accuracy. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the rainfall estimation model in this invention; Figure 2 This is a schematic diagram of a two-way gating unit. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figure 1 As shown, a rainfall estimation method based on a microwave communication link includes S1 to S3.

[0022] S1. Construct multiple microwave communication links and obtain the received signal level of the microwave communication links during rainfall.

[0023] S2. Preprocess the acquired received signal level to generate the original signal sequence.

[0024] S3. The original signal sequence is input into a pre-trained rainfall estimation model for processing to obtain the rainfall intensity. The rainfall estimation model is based on the Transformer model and includes a location encoding module for injecting location information into the original signal sequence, a multi-layer encoder for extracting multi-scale temporal features and long-term rainfall trends from the original signal sequence, a bidirectional gated loop unit for dynamically gating the multi-scale temporal features to obtain short-term rainfall fluctuations, and a data output module for outputting the rainfall intensity.

[0025] Furthermore, the Transformer model itself is an unordered structure, unlike neural networks such as RNNs which can process data sequentially according to the input sequence. To overcome this problem, the rainfall estimation model of this invention introduces a learnable location encoding module to generate learnable location codes, thereby injecting location information into each time step of the input sequence. Through the learnable location encoding module, a location vector is added to the input at each time step, enabling the inherently unordered Transformer model to not only recognize the temporal structure in the data but also maintain parallel computing capabilities. The method for injecting location information into the original signal sequence by the location encoding module includes: ; Where X is the time step in the original signal sequence. The time step after the location information is injected. It contains location information, and has ,in The total length of the original signal sequence. It is the feature dimension of the time step.

[0026] In traditional Transformer models, positional encoding is typically calculated using predefined sine and cosine functions. This traditional positioning encoding assumes a fixed periodicity in the positional relationships at each time step during data training and cannot be adjusted based on the input data. Therefore, it is more suitable for time-series data with obvious periodic patterns. However, for rainfall data, the temporal patterns are complex and non-periodic, and traditional positioning encoding cannot fully capture the temporal characteristics of the data. This invention employs a learnable positioning encoding module that, through backpropagation training, allows the model to dynamically adjust the positional encoding representation based on the characteristics of the actual data during training. This flexibility enables the model to automatically learn short-term sudden changes and long-term trend changes in rainfall data, greatly improving the model's accuracy and generalization ability.

[0027] Furthermore, the multilayer encoder includes multiple linearly arranged encoders, each encoder comprising a multi-head self-attention block and a feedforward network arranged sequentially, the feedforward network comprising two linear transform layers and an activation function.

[0028] A multi-head self-attention block consists of multiple single-head self-attention mechanisms. The single-head self-attention mechanism primarily uses scaled dot product attention, and the calculation process is as follows: ; In the formula, Q, K, and V represent the query, key, and value, respectively, and are matrices composed of vectors obtained from the input data through different linear transformations. The activation function is used to normalize the attention scores so that the sum of each row is 1. The dimension is K; This is a scaling factor used to prevent the dot product result from becoming too large, avoiding gradient vanishing or exploding. Multi-head self-attention mechanisms not only allow the model to focus on certain important information of the input features, but also allow each attention mechanism to optimize different parts of the input features, balancing the potential biases that a single attention mechanism might produce. More specifically, multi-head self-attention mechanisms are an extension of scaled dot product attention, and can be computed in parallel by multiple different attention heads. The computation method of each attention head is the same as that of the self-attention mechanisms described above. Multi-head self-attention mechanisms map input features to multiple different subspaces, compute attention mechanisms in parallel in each subspace, and then fuse these results together to obtain richer expressive power. The entire computation process can be represented as: in, Let h be the output of the i-th attention head; h is the number of heads. To output the mapping matrix; , , , In rainfall prediction tasks, multi-head self-attention mechanisms focus on the correlations between different positions in the input sequence and learn the relationship between the received signal strength of different links and rainfall intensity. This multi-head design enables the model to capture features at different time scales, such as minute-level light rain fluctuations or hourly rainfall trends. While extracting multi-view features and learning the correlations between different features, it can also focus on features from all historical moments in the window at each time step, thus preserving key historical information in the rainfall process.

[0029] Furthermore, since the self-attention mechanism itself is a linear structure, although it can aggregate information, it lacks sufficient nonlinear transformation capabilities. Therefore, the encoder in this invention also includes a feedforward network. The feedforward network can introduce nonlinearity through the ReLU activation function to learn the complex relationships between input features, enhancing the model's ability to model nonlinear relationships. The feedforward network layer typically consists of two linear transformation layers and one activation function. The implementation is as follows: ; in, The input vector; , These are the weight matrices of the two linear layers; It is the bias of the two linear layers before and after; This refers to the intermediate layer dimension. The feedforward network layer can perform independent nonlinear transformations on the features at each location, complementing the global interaction of the attention layer to capture more complex patterns, thereby enhancing the model's expressive power. In this invention, a Dropout layer is also added to the feedforward network to prevent overfitting, making the model more stable in scenarios with small samples or noise, and further improving the model's generalization ability.

[0030] Furthermore, the bidirectional gated loop unit includes a forward gated loop unit and a reverse gated loop unit. The forward gated loop unit processes the first input sequence in a first order, and the reverse gated loop unit processes the first input sequence in a second order. The first input sequence is output by a multilayer encoder, and the first order and the second order are reversed.

[0031] More specifically, both the forward-gated recurrent unit and the reverse-gated recurrent unit include update gates and reset gates. The update gate determines the extent to which the state information from the previous moment is retained in the current moment, controlling the flow of information. The reset gate determines the degree of influence of the state information from the previous moment on the current candidate state, adjusting the influence on the candidate state. The forward-gated recurrent unit processes the input sequence in chronological order, while the reverse-gated recurrent unit processes the sequence in reverse chronological order. This bidirectional modeling approach connects information from before and after each input moment. By simultaneously utilizing past and future information within each window of the input sequence, the bidirectional gated recurrent unit helps the model consider changes in the input data at each moment, improving its utilization rate. On the other hand, the bidirectional gated recurrent unit, through its gating mechanism and bidirectional structure, works in conjunction with the multi-head self-attention blocks of the multi-layer encoder to compensate for the model's global bias. Furthermore, the bidirectional structure in the model of this invention utilizes past and future information within the historical window, avoiding the problem of leaking future information.

[0032] like Figure 2 As shown, the processing steps of the forward-gated loop unit and the reverse-gated loop unit are as follows: ; ; ; ; in, To update the gate output, To reset the gate's output, This is the weight matrix. For bias vectors, This represents the candidate hidden layer state. This is the output at state t-1. The value of takes between 0 and 1. If it's close to 1, it means the previous hidden state is more important, and more information from the hidden state needs to be retained. If it's close to 0, it means the previous hidden state is not important, and the information from the previous time step will be forgotten by the model. Similarly, The value of is between 0 and 1. If it is close to 1, it means that the hidden state has not been updated. and Almost identical; when When the value is close to 0, The value is determined by Decide. and Represents a hidden state. For the input layer, This is the output layer.

[0033] Furthermore, the data output module includes an attention pooling layer and a fully connected layer arranged sequentially. The attention pooling layer transforms the second input sequence, which is the output of the bidirectional gated recurrent unit (BRU). Since the output of the BRU is a time-step sequence with the shape (B, L, H), where B is the batch size, L is the sequence length, and H is the hidden state dimension at each time step, a uniform, fixed-length vector is needed in the actual output process to match the input requirements of the fully connected layer. Therefore, an attention pooling layer is first introduced in the data output module to aggregate input information by adaptively assigning weights to each position of the input sequence. The attention pooling layer first calculates the attention score at each time step, and then calculates the attention weights through a simple linear layer, as shown in the following formula: ; in, It is the attention weight at position t; This is the score for the feature at position t. After obtaining the attention weights for all positions, the final pooling result is obtained through a weighted summation method, as shown below: ; Where Z is the pooled vector, which is the weighted average of the features at each time step of the input sequence.

[0034] Using the methods described above, attention pooling layers enable the model to automatically learn which time steps are more important for rainfall prediction and assign them higher weights. Compared to methods like average pooling and max pooling, which simply take the output of the last time step, attention pooling does not discard crucial time segments in the sequence. This approach allows the model to maintain the integrity of temporal information while automatically learning and emphasizing the time steps most useful for rainfall prediction, thereby improving the model's prediction accuracy.

[0035] The present invention further provides a rainfall estimation system based on a microwave communication link, which is used to implement the above-mentioned rainfall estimation method based on a microwave communication link. The system includes a microwave communication subsystem, a data acquisition module, and a model running module.

[0036] The microwave communication subsystem is used to construct multiple microwave communication links and to acquire the received signal levels of the microwave communication links during rainfall.

[0037] The data acquisition module is used to preprocess the received signal levels to generate the original signal sequence.

[0038] The model execution module is used to run a pre-trained rainfall estimation model to process the original signal sequence and obtain the rainfall intensity.

[0039] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division; in actual implementation, there may be other division methods. Furthermore, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or modules may be electrical, mechanical, or other forms.

[0040] The present invention also provides a computer device, including a memory and a processor.

[0041] Memory is used to store computer programs.

[0042] A processor is used to read and execute computer programs to implement the above-described rainfall estimation method based on a microwave communication link.

[0043] Finally, the present invention provides a storage medium for storing a computer program that, when executed, implements the above-described rainfall estimation method based on a microwave communication link.

[0044] The memory, as a carrier of resources, can be a read-only memory, random access memory, disk, or optical disk, etc. The resources stored on it can include the operating system, computer programs, etc., and the storage method can be temporary or permanent storage. The operating system is used to manage and control the various hardware devices and computer programs on the electronic device, and can be Windows Server, Netware, Unix, Linux, etc. In addition to including computer programs capable of performing the adaptive emotion regulation method based on personalized reconfigurable music disclosed in any of the foregoing embodiments, the computer programs may further include computer programs capable of performing other specific tasks. The processor can be a general-purpose processor product based on architectures such as x86, IA64, RISC, MIPS, and ARM.

[0045] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in this invention may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A rainfall estimation method based on a microwave communication link, characterized in that, Includes the following steps: Multiple microwave communication links were constructed, and the received signal levels of the microwave communication links were obtained during rainfall. The received signal level is preprocessed to generate the original signal sequence; The original signal sequence is input into a pre-trained rainfall estimation model for processing to obtain the rainfall intensity. The rainfall estimation model is based on the Transformer model and includes a location encoding module for injecting location information into the original signal sequence, a multi-layer encoder for extracting multi-scale temporal features and long-term rainfall trends from the original signal sequence, a bidirectional gated loop unit for dynamically gating the multi-scale temporal features to obtain short-term rainfall fluctuations, and a data output module for outputting rainfall intensity.

2. The rainfall estimation method based on a microwave communication link as described in claim 1, characterized in that, The method by which the position encoding module injects position information into the original signal sequence includes: ； Where X is the time step in the original signal sequence. The time step after the location information is injected. It contains location information, and has ,in The total length of the original signal sequence. It is the feature dimension of the time step.

3. The rainfall estimation method based on a microwave communication link as described in claim 1, characterized in that, The multilayer encoder includes multiple linearly arranged encoders, each encoder comprising a multi-head self-attention block and a feedforward network arranged sequentially. The feedforward network includes two linear transform layers and an activation function.

4. The rainfall estimation method based on a microwave communication link as described in claim 1, characterized in that, The bidirectional gated loop unit includes a forward gated loop unit and a reverse gated loop unit. The forward gated loop unit processes the first input sequence in a first order, and the reverse gated loop unit processes the first input sequence in a second order. The first input sequence is output by the multilayer encoder, with the first order being the reverse of the second order.

5. The rainfall estimation method based on a microwave communication link as described in claim 4, characterized in that, Both the forward-gated loop unit and the reverse-gated loop unit include an update gate and a reset gate. The processing steps of the forward-gated loop unit and the reverse-gated loop unit are as follows: ； ； ； ； in, To update the gate output, To reset the gate's output, This is the weight matrix. For bias vectors, This represents the candidate hidden layer state. This is the output at state t-1.

6. The rainfall estimation method based on a microwave communication link as described in claim 1, characterized in that, The data output module includes an attention pooling layer and a fully connected layer arranged sequentially. The attention pooling layer is used to transform the second input sequence, which is the output of the bidirectional gated recurrent unit.

7. The rainfall estimation method based on a microwave communication link as described in claim 1, characterized in that, The method further includes the following steps: The rainfall estimation model is trained by backpropagation based on the rainfall intensity to optimize the location encoding module.

8. A rainfall estimation system based on a microwave communication link, characterized in that, The system is used to implement a rainfall estimation method based on a microwave communication link as described in any one of claims 1-7, the system comprising: The microwave communication subsystem is used to construct multiple microwave communication links and acquire the received signal levels of the microwave communication links during rainfall. The data acquisition module is used to preprocess the received signal levels to generate the original signal sequence; The model execution module is used to run a pre-trained rainfall estimation model to process the original signal sequence and obtain the rainfall intensity.

9. A computer device, characterized in that, include: Memory, used to store computer programs; A processor for reading and executing the computer program to implement a rainfall estimation method based on a microwave communication link as described in any one of claims 1-7.

10. A storage medium, characterized in that, Used to store a computer program, which, when executed, implements a rainfall estimation method based on a microwave communication link as described in any one of claims 1-7.

Images