Microwave link rainfall inversion method and system based on deep learning

Abstract

The invention discloses a microwave link rainfall inversion method and system based on deep learning, and the method comprises the steps: obtaining the preprocessed time sequence feature data according to the static data and dynamic data of a microwave link; according to the preprocessed time sequence feature data, through a hybrid model fusing a long short-term memory network and a multi-scale convolutional neural network, spatio-temporal feature extraction is carried out, and a multi-branch prediction value is generated; and outputting a rainfall intensity prediction result through an adaptive fusion layer according to the multi-branch prediction value. According to the method, the deep learning model and the adaptive loss function are constructed, the limitation of a traditional microwave link rainfall inversion method based on an empirical formula in the aspects of signal attenuation separation and non-rainfall interference elimination is overcome, and more accurate and real-time estimation of rainfall intensity and spatial distribution is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological monitoring and microwave link attenuation data processing, and particularly relates to a microwave link rainfall inversion method and system based on deep learning. Background Art

[0002] In the field of meteorological monitoring, a commonly used rainfall inversion method at present is based on the signal attenuation characteristics of commercial microwave links. The traditional method establishes a mathematical relationship between the attenuation caused by rainfall and the rainfall intensity through an empirical formula (such as the power-law model) to achieve rainfall estimation. Specifically, Equation (1) represents this relationship:

[0003]

[0004] Wherein, is the attenuation caused by rainfall at time step n (unit: dB), L is the link length between the microwave transmitter and the receiver, R n is the rainfall rate at time step n (unit: mm / h), and a and b are power-law (PL) coefficients, which depend on the frequency and polarization of the signal and the distribution of raindrop sizes, and are usually assumed to be known

[0005] However, the traditional method has many limitations in practical applications. The signal attenuation of the microwave link is affected not only by rainfall but also by various non-rainfall factors, such as atmospheric attenuation, free-space path loss, equipment noise, and wet antenna effect. These factors make it difficult to accurately separate the signal change caused by rainfall from the total attenuation, resulting in significant errors in rainfall prediction results. In addition, the rapid change of environmental conditions and the instability of equipment status further increase the difficulty of accurately quantifying the baseline attenuation and non-rainfall factors, seriously restricting the improvement of rainfall inversion accuracy. Summary of the Invention

[0006] To solve the above technical problems, the present invention proposes a microwave link rainfall inversion method and system based on deep learning to solve the problems existing in the above prior art.

[0007] To achieve the above object, in the first aspect, the present invention provides a microwave link rainfall inversion method based on deep learning, including:

[0008] Obtaining preprocessed time-series feature data according to the static data and dynamic data of the microwave link;

[0009] Performing spatio-temporal feature extraction through a hybrid model that fuses a long short-term memory network and a multi-scale convolutional neural network according to the preprocessed time-series feature data to generate multi-branch prediction values;

[0010] According to the multi-branch prediction values, a rainfall intensity prediction result is output through an adaptive fusion layer.

[0011] Preferably, the process of obtaining the preprocessed time series feature data includes:

[0012] Obtaining parameters of transmit signal power, link length, and polarization mode based on the static data;

[0013] Acquire a received signal strength indicator and a carrier-to-interference-plus-noise ratio timing signal based on the dynamic data;

[0014] According to the static data and the dynamic data, pre-processed time series feature data is obtained, the time series feature data is segmented by a sliding window method, and the segmented data is normalized.

[0015] Preferably, for missing values in dynamic data, data segments whose continuous missing duration exceeds a preset threshold are deleted.

[0016] Preferably, the hybrid model comprises a dual-branch parallel structure:

[0017] According to the Bi-LSTM network of the first branch, long-term dependency modeling is performed on the preprocessed time series data to generate the first prediction value;

[0018] According to the multi-scale convolutional network of the second branch, spatial features are hierarchically extracted through convolution kernels and pooling operations of different scales, and combined with LSTM to generate the second to fourth prediction values.

[0019] Preferably, the hierarchical extraction of the multi-scale convolutional network includes:

[0020] Extract macro-scale features based on 5×5 convolution kernel and maximum pooling operation;

[0021] Extract mesoscale features based on 3×3 convolution kernel and maximum pooling operation;

[0022] Local fine features are extracted based on 1×1 convolution kernel and maximum pooling operation.

[0023] Preferably, the adaptive fusion layer is implemented by a fully connected network, including:

[0024] According to the weight distribution of the multi-branch prediction values, the weighted sum of the prediction values of each branch is calculated to generate a final rainfall intensity prediction result.

[0025] Preferably, the method further comprises the step of designing a loss function:

[0026] According to the imbalance of rainfall intensity distribution, an adaptive loss function including dynamic weight parameters is defined;

[0027] Balance the loss contributions corresponding to light rainfall and heavy rainfall according to the initial value and decay rule of the dynamic weight parameter.

[0028] Preferably, the adaptive loss function satisfies the following formula:

[0029]

[0030] where L hc represents the loss component corresponding to light rainfall, and L lc represents the loss component corresponding to heavy rainfall, and λ is the loss contribution weight.

[0031] Preferably, the method further includes adding Gaussian noise to the input data to enhance the robustness to noise interference.

[0032] In a second aspect, the present invention provides a microwave link rainfall inversion system based on deep learning for implementing the method described in the first aspect, including:

[0033] A data acquisition and processing module for obtaining preprocessed time-series feature data according to the static data and dynamic data of the microwave link;

[0034] A feature extraction module for performing spatio-temporal feature extraction according to the preprocessed time-series feature data through a hybrid model that fuses a long short-term memory network and a multi-scale convolutional neural network to generate multi-branch prediction values;

[0035] A combined prediction module for outputting a rainfall intensity prediction result through an adaptive fusion layer according to the multi-branch prediction values.

[0036] Compared with the prior art, the present invention has the following advantages and technical effects:

[0037] The present invention provides a microwave link rainfall inversion method based on deep learning, including: First, obtain preprocessed time-series feature data according to the static data and dynamic data of the microwave link; Second, perform spatio-temporal feature extraction according to the preprocessed time-series feature data through a hybrid model that fuses a long short-term memory network and a multi-scale convolutional neural network to generate multi-branch prediction values; Finally, output a rainfall intensity prediction result through an adaptive fusion layer according to the multi-branch prediction values.

[0038] The present invention utilizes artificial intelligence to design a deep learning model, overcomes the limitations of traditional microwave link rainfall inversion methods based on empirical formulas in signal attenuation separation and non-rainfall interference elimination, and realizes a more accurate and real-time estimation of rainfall intensity and spatial distribution. By constructing a deep learning model and an adaptive loss function, key features are extracted from a large amount of microwave link data, improving the accuracy and stability of rainfall inversion, and providing reliable rainfall prediction data support for meteorological flood warning and environmental protection.

[0039] In addition, the present invention also aims to reduce the overall operation complexity of the system. Through an integrated complete solution that combines data preprocessing, feature extraction, and intelligent rainfall inversion, it realizes the automated processing from raw signal acquisition to rainfall estimation. This system can not only effectively filter out the interference of non-rainfall factors (such as equipment noise, wet antenna effect), but also has strong adaptive learning ability and robustness, adapting to changing environmental conditions, so it can be widely applied to fields such as intelligent meteorological forecasting in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings forming a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation to this application. In the drawings:

[0041] Figure 1 is the data processing flow chart of the embodiment of the present invention;

[0042] Figure 2 is the schematic diagram of the system model structure of the embodiment of the present invention;

[0043] Figure 3 is the schematic diagram of the PCC values of each system of the embodiment of the present invention under different noise interferences. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will refer to the drawings and combine the embodiments to detail this application.

[0045] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0046] Embodiment 1

[0047] In this embodiment, a microwave link rainfall inversion method based on deep learning is provided, including:

[0048] S1. Obtain preprocessed time-series feature data according to the static data and dynamic data of the microwave link;

[0049] Further, the process of obtaining the preprocessed time-series feature data includes:

[0050] Obtain the parameters of the transmitted signal power, link length, and polarization mode according to the static data;

[0051] Obtain the time series signals of received signal strength indication and carrier interference noise ratio according to the dynamic data;

[0052] Obtain the preprocessed time series feature data according to the static data and dynamic data, segment the time series feature data by the sliding window method, and perform normalization processing on the segmented data.

[0053] Furthermore, for the missing values in the dynamic data, delete the data segments with consecutive missing durations exceeding the preset threshold.

[0054] Specifically, first collect the data of commercial microwave links, which can be divided into two types: static data and dynamic data. Among them, the static data includes transmitted signal power, link length, and polarization mode of the link; the dynamic data includes RSSI (Received Signal Strength Indication) affected by attenuation caused by rainfall, etc.

[0055] Systematically organize the obtained dynamic data and static data in chronological order, and then perform preprocessing operations to input high-quality data for the subsequent model. The detailed steps are as Figure 1 shown. In addition, due to the large differences in data values between different features, normalization operations are performed on the feature values. To better capture the temporal relationship between data, the sliding window method is used to process the sequence when processing data. If the amount of data in the obtained dataset is large and a part of it is missing continuously for a long time, the data with missing values in the dataset also needs to be deleted.

[0056] S2. According to the preprocessed time series feature data, perform spatio-temporal feature extraction through a hybrid model that combines a long short-term memory network and a multi-scale convolutional neural network to generate multi-branch prediction values;

[0057] Furthermore, the hybrid model includes a two-branch parallel structure:

[0058] According to the Bi-LSTM network of the first branch, model the long-term dependence relationship of the preprocessed time series data to generate the first prediction value;

[0059] According to the multi-scale convolutional network of the second branch, hierarchically extract spatial features through convolutional kernels and pooling operations of different scales, and combine with LSTM to generate the second to fourth prediction values.

[0060] Furthermore, the hierarchical extraction of the multi-scale convolutional network includes:

[0061] Extract macroscopic scale features according to the 5×5 convolutional kernel and max pooling operation;

[0062] Extract mesoscopic scale features according to the 3×3 convolutional kernel and max pooling operation;

[0063] Extract local fine features according to the 1×1 convolution kernel and the max pooling operation.

[0064] Specifically, the development of deep learning technology has brought new breakthroughs to rainfall prediction. These neural network models can effectively process large-scale multi-dimensional meteorological data, extract key features from it, and learn complex spatio-temporal patterns, thus significantly improving the accuracy and timeliness of rainfall prediction. In this embodiment, a deep learning feature mining model with a hybrid architecture is constructed, combining the long-term memory ability of the long short-term memory network (LSTM) and the ability of CNN to extract data spatial features. The system model structure is as Figure 2 shown.

[0065] In this system model architecture, the model adopts a dual-branch parallel design to fully mine the spatio-temporal features of the data. The input of the model is the time series data of the preprocessed microwave attenuation signal, and the output is the rainfall intensity (mm / h) at the corresponding time step. One branch directly models the original sequence through Bi-LSTM, effectively capturing the long-term dependencies of the time series data and outputting the predicted value y b . The other branch uses a multi-scale convolutional network for feature extraction. Through convolutional kernels of different scales and layer-by-layer downsampling operations, hierarchical extraction of multi-scale features of the data is achieved. Specifically, this branch first performs macro feature extraction and downsampling through 5×5 convolution and 3×3 max pooling, and then splits to obtain the first prediction branch, and the predicted value y1 is obtained through LSTM modeling. The backbone network continues to extract meso-scale features through 3×3 convolution and 3×3 max pooling, and splits again to obtain the second prediction branch, and the predicted value y2 is output through LSTM modeling. Finally, local fine features are extracted through 1×1 convolution and 2×2 max pooling, and the predicted value y3 is obtained through LSTM modeling.

[0066] S3. Output the rainfall intensity prediction result through the adaptive fusion layer according to the multi-branch predicted values.

[0067] Furthermore, the adaptive fusion layer is implemented through a fully connected network, including:

[0068] Calculate the weighted sum of the predicted values of each branch according to the weight assignment of the multi-branch predicted values to generate the final rainfall intensity prediction result.

[0069] Specifically, the model finally adaptively fuses the outputs of the four prediction branches through a fully connected layer to obtain the final predicted value y.

[0070] Furthermore, the method also includes the design step of the loss function:

[0071] Define an adaptive loss function containing dynamic weight parameters according to the imbalance of the rainfall intensity distribution;

[0072] Balance the loss contributions corresponding to light rainfall and heavy rainfall according to the initial value and decay rule of the dynamic weight parameter.

[0073] Specifically, the loss function of the model is designed as follows:

[0074] A key problem with the existing rainfall intensity dataset is the imbalance in the rainfall intensity distribution. This imbalance may cause the training model to tend to certain rainfall distributions, thus affecting the prediction performance. Currently, a relatively simple loss function is adopted, directly calculating the mean square error (MSE) between the predicted value and the actual value. However, it is difficult for MSE itself to balance the prediction errors of different rainfall intensities, resulting in significant biases in the prediction results, especially the underestimation phenomenon commonly exists when predicting heavy rainfall.

[0075] To address this problem, this embodiment designs a loss function that can adapt to different rainfall intensities to improve the model's prediction ability for light and heavy rainfall. The core idea is to introduce an adaptive weight mechanism based on rainfall intensity. By dynamically adjusting the loss contributions corresponding to different rainfall intensities, the model can achieve balanced optimization of the prediction of light and heavy rainfall during the training process, thereby improving the overall prediction accuracy.

[0076] The specific formula of the designed loss function is as follows. L is the loss function, λ is the loss contribution weight, and the initial value is set to 1. The weight of light rainfall in L is relatively large, resulting in a larger loss. The model first trains for light rainfall, and then as λ decreases, the L where heavy rainfall is located has a larger weight loss, and the model begins to gradually increase its attention to the characteristics of heavy rainfall. Finally, when λ basically no longer changes, the two reach a balance and the training is completed. hc Among them, N is the number of samples, lc is the predicted value, y is the actual value, γ

[0077]

[0078] where N is the number of samples, is the predicted value, y is the actual value, γ lc 、γ hc 、α lc and α hc are parameters.

[0079] With the help of this adaptive loss function in this embodiment, after the deep learning model is trained, the overall system architecture can receive the attenuation signal data from the microwave link, and after preprocessing, feature extraction, and time series modeling, it outputs the rainfall intensity estimation at the corresponding time points.

[0080] Beneficial effects of this embodiment:

[0081] In this embodiment, by constructing a feature extraction model based on deep learning, rainfall information can be automatically captured from a large amount of microwave signal data, realizing the full-process intelligence from signal preprocessing to rainfall estimation. This method can not only effectively filter non-rainfall interference and improve the mapping accuracy between signal attenuation and rainfall intensity, but also has the ability to adaptively learn and process complex non-linear relationships. In addition, the combination of a variety of data preprocessing techniques and deep learning models can achieve the comprehensive mining of multi-feature responses, further improving the timeliness and accuracy of rainfall prediction, and providing more reliable data support for practical applications such as meteorological early warning and environmental monitoring.

[0082] To evaluate the performance of the proposed deep learning model in rainfall intensity prediction, some evaluation metrics and experimental designs were adopted. In addition, Python 3.9 was used as the development environment.

[0083] 1. Experimental Design

[0084] All models used the adaptive loss function designed in this embodiment as the training loss function, and the training process was set for 20 epochs to ensure that the model converged sufficiently. Under the condition of keeping the data preprocessing consistent, this embodiment selected four other common basic models to compare the performance with the model proposed in this embodiment, including LSTM, gated recurrent unit (GRU), bidirectional LSTM (Bi-LSTM), convolutional neural network (CNN), and the structure in which the CNN downsampling structure in the model architecture was replaced by a serial structure of a convolutional layer and a pooling layer (hereinafter referred to as CNN-LSTM).

[0085] 2. Experimental Data

[0086] This experiment used an E-band millimeter-wave two-way communication link built in the Nanjing area, with working frequencies of 71 GHz and 81 GHz respectively, a length of 3 km, and collected received signal strength data from December 2019 to March 2020. The receiver sampled once every 1 minute, and the quantization resolution was 0.1 dB.

[0087] The static data is shown in Table 1; the dynamic data includes the CINR (Carrier to Interference plus Noise Ratio) and RSSI (Received Signal Strength Indication) of 71 GHz microwave, the CINR and RSSI of 81 GHz microwave, and the rainfall intensity values collected by the rain gauge.

[0088] Table 1

[0089]

[0090] 3. Evaluation Metrics

[0091] To scientifically evaluate the performance of the rainfall prediction model, it is necessary to quantitatively analyze the prediction effect of the model from multiple dimensions. Considering the particularity of rainfall prediction, a single evaluation metric is difficult to comprehensively reflect the prediction ability of the model. Therefore, five complementary evaluation metrics are selected: The Pearson correlation coefficient (PCC) is used to measure the correlation between the predicted value and the actual value; the coefficient of determination (R 2 ) is used to characterize the model's ability to explain data changes; the root mean square error (RMSE) is used to evaluate the average deviation degree between the predicted value and the actual value; the mean absolute error (MAE) is used to measure the absolute magnitude of the prediction error; and the mean bias (MB) is used to evaluate the systematic bias of the predicted value. The combined use of these metrics can comprehensively evaluate the performance of the model in the rainfall prediction task.

[0092] Specifically, the calculation methods of each evaluation metric are as follows:

[0093] The closer the value of the Pearson correlation coefficient (PCC) is to 1, the stronger the correlation. The calculation formula is:

[0094]

[0095] where n is the number of samples, P i and are the actual value and the predicted value at the i-th time point, and are the average values of all actual rainfall amounts and all predicted rainfall amounts, respectively.

[0096] The closer the value of the coefficient of determination (R 2 ) is to 1, the stronger the model's explanatory ability. The calculation formula is:

[0097]

[0098] where n is the number of samples, y i and are the actual value and the predicted value at the i-th time point, is the mean value of all actual rainfall amounts.

[0099] The smaller the value of the root mean square error (RMSE), the higher the prediction accuracy. The calculation formula is:

[0100]

[0101] where n is the number of samples, y i and are the actual value and the predicted value at the i-th time point, respectively.

[0102] The smaller the value of the Mean Absolute Error (MAE), the higher the prediction accuracy. The calculation formula is:

[0103]

[0104] where n is the number of samples, y i and are the actual value and the predicted value at the i-th time point respectively.

[0105] The Mean Bias (MB) reflects the systematic bias of the predicted value relative to the actual value, and can evaluate whether there is a trend of overestimation or underestimation in the model. The calculation formula is:

[0106]

[0107] where n is the number of samples, y i and are the actual value and the predicted value at the i-th time point respectively.

[0108] 4. Experimental Results

[0109] (1) Performance

[0110] Table 2 shows the comprehensive evaluation results of the entire test data set. The model of this embodiment is superior to other models in various indicators.

[0111] Table 2

[0112]

[0113] The experimental results verify that the model designed in this embodiment can significantly improve the rainfall prediction accuracy, and also further prove the comprehensive ability of the system architecture proposed in this embodiment in the rainfall intensity prediction task.

[0114] (2) Robustness

[0115] In practical applications, microwave link measurement data will inevitably be affected by various noises, including sensor errors, environmental interference, and data transmission distortion, etc. Therefore, the robustness to noise is one of the key characteristics for the actual deployment of rainfall prediction models.

[0116] To systematically evaluate the robustness of the model, this embodiment introduces Gaussian noises with different standard deviations (from 0 to 2) into the test data set and conducts a series of experiments. By gradually increasing the noise level, this embodiment can analyze the degradation of the model's performance when facing increasingly severe noise interference. Figure 3 The PCC values for each system under different noise interferences.

[0117] As the noise level further increases, the performance of each system model shows a gradually decreasing trend, which is expected. However, even at the highest noise level (variance = 2.0), the system model designed in this embodiment can still maintain a PCC value of over 80%, indicating that it can still maintain a high anti-interference ability at a relatively high noise level, further verifying the feasibility of the system model in practical rainfall prediction applications. This strong robustness feature enables the model to still provide reliable prediction results when facing complex and noisy real meteorological data.

[0118] Embodiment 2

[0119] Based on the same inventive concept, this embodiment also provides a microwave link rainfall inversion system based on deep learning for implementing the method described in Embodiment 1, including:

[0120] A data acquisition and processing module, configured to obtain preprocessed time-series feature data according to the static data and dynamic data of the microwave link;

[0121] A feature extraction module, configured to perform spatio-temporal feature extraction according to the preprocessed time-series feature data through a hybrid model that fuses a long short-term memory network and a multi-scale convolutional neural network, and generate multi-branch prediction values;

[0122] A combined prediction module, configured to output a rainfall intensity prediction result through an adaptive fusion layer according to the multi-branch prediction values.

[0123] The microwave link rainfall inversion system based on deep learning provided in this embodiment has all the advantages of the microwave link rainfall inversion method based on deep learning provided in Embodiment 1.

[0124] The above are only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A rainfall inversion method for microwave links based on deep learning, characterized in that, It includes the following steps: Obtain preprocessed time-series feature data according to the static data and dynamic data of the microwave link; According to the preprocessed time-series feature data, perform spatio-temporal feature extraction through a hybrid model that fuses a long short-term memory network and a multi-scale convolutional neural network to generate multi-branch prediction values; According to the multi-branch prediction values, output the rainfall intensity prediction result through an adaptive fusion layer.

2. The method according to claim 1, wherein: The process of obtaining preprocessed time-series feature data includes: According to the static data, obtain the parameters of the transmitted signal power, link length, and polarization mode; According to the dynamic data, obtain the time-series signals of the received signal strength indication and carrier interference noise ratio; According to the static data and dynamic data, obtain preprocessed time-series feature data, segment the time-series feature data by the sliding window method, and perform normalization processing on the segmented data.

3. The method according to claim 1, wherein: For the missing values in the dynamic data, delete the data segments with consecutive missing durations exceeding the preset threshold.

4. The method according to claim 1, wherein: The hybrid model includes a two-branch parallel structure: According to the Bi-LSTM network of the first branch, perform long-term dependence modeling on the preprocessed time-series data to generate the first prediction value; According to the multi-scale convolutional network of the second branch, hierarchically extract spatial features through convolutional kernels and pooling operations of different scales, and combine with LSTM to generate the second to fourth prediction values.

5. The method according to claim 4, wherein: The hierarchical extraction of the multi-scale convolutional network includes: Extract macroscopic-scale features according to a 5×5 convolutional kernel and a max pooling operation; Extract mesoscopic-scale features according to a 3×3 convolutional kernel and a max pooling operation; Extract local fine features according to a 1×1 convolutional kernel and a max pooling operation.

6. The method according to claim 1, wherein: The adaptive fusion layer is implemented through a fully connected network and includes: According to the weight assignment of the multi-branch prediction values, calculate the weighted sum of each branch prediction value to generate the final rainfall intensity prediction result.

7. The method according to claim 1, wherein: It further includes the design step of the loss function: According to the imbalance of the rainfall intensity distribution, define an adaptive loss function including dynamic weight parameters; According to the initial value and decay rule of the dynamic weight parameters, balance the loss contributions corresponding to small rainfall amounts and large rainfall amounts.

8. The method according to claim 7, wherein: The adaptive loss function satisfies the following formula: Among them, L hc represents the loss component corresponding to light rainfall, and L lc represents the loss component corresponding to heavy rainfall, and λ is the loss contribution weight.

9. The method according to claim 1, wherein: It further includes adding Gaussian noise to the input data to enhance the robustness to noise interference.

10. A microwave link rainfall inversion system based on deep learning, characterized in that, For implementing the method according to any one of claims 1-9, it includes: A data acquisition and processing module for obtaining preprocessed time-series feature data according to the static data and dynamic data of the microwave link; A feature extraction module, which is used to perform spatio-temporal feature extraction on the preprocessed time-series feature data through a hybrid model that combines a long short-term memory network and a multi-scale convolutional neural network to generate multi-branch prediction values; A combined prediction module, which is used to output a rainfall intensity prediction result through an adaptive fusion layer according to the multi-branch prediction values.

Images