Rainfall estimation method and system based on communication perception integration

Abstract

The application relates to the technical field of rainfall estimation, and provides a rainfall estimation method and system based on communication and perception integration, which comprises the following steps: obtaining rainfall monitoring data of a rainfall perception device pre-deployed at a fixed position and generating reference weather features; obtaining received signal strength data based on a preset time resolution in a commercial microwave link and generating original perception signals; performing feature extraction on the original perception signals by using a plurality of time windows with different lengths to obtain a plurality of original feature sequences with different time scales; inputting the plurality of original feature sequences into a pre-trained multi-scale prediction model for processing to generate branch rainfall prediction results; weighting and fusing the branch rainfall prediction results based on a fusion weight to obtain integrated prediction values; and performing physical constraint filtering on the integrated prediction values based on a signal fluctuation degree to obtain final rainfall estimation results. Compared with the prior art, the application has higher estimation accuracy.

Description

Technical Field

[0001] This invention relates to the field of rainfall estimation technology, specifically a rainfall estimation method and system based on integrated communication and sensing. Background Technology

[0002] Integrated Communication and Sensing (ISAC) is a key technology direction for next-generation mobile communication systems. Its core idea is to simultaneously achieve information transmission and environmental sensing functions on the same wireless infrastructure, thereby significantly improving spectrum efficiency, reducing deployment costs, and expanding network intelligence capabilities. With the large-scale deployment of 5G networks and the in-depth advancement of 6G research, environmental sensing using existing commercial microwave links (CML) has gradually become a research hotspot. CML has unique advantages as a rainfall estimation tool: commercial microwave links are widely distributed globally, with well-developed infrastructure; the wavelength of microwave signals in the operating frequency band (6GHz to 40GHz) is comparable to the size of raindrops; rainfall significantly attenuates and scatters microwave signals, and there is a definite physical correlation between the degree of attenuation and rainfall intensity. This forms the physical basis for using CML to estimate rainfall. Compared with traditional rainfall monitoring methods (such as rain gauges and weather radar), CML has significant advantages such as wide coverage, well-developed infrastructure, and no need for additional hardware deployment. It can effectively fill the spatial coverage blind spots of existing rainfall observation networks, providing high spatiotemporal resolution rainfall data for urban flood warning, hydrological monitoring, and agricultural meteorology.

[0003] However, existing CML-based rainfall estimation methods still face key bottlenecks. On the one hand, the received signal strength (RSS) of microwave links is susceptible to interference from non-rainfall factors such as temperature drift, antenna alignment deviation, and multipath effects. Existing models lack effective physical discrimination mechanisms, making it difficult to distinguish such signal fluctuations from actual rainfall attenuation, resulting in a high false alarm rate. On the other hand, in multi-branch prediction architectures, the fusion weights of features from each branch largely rely on manual experience and cannot be adaptively adjusted based on link status, regional characteristics, or seasonal changes, severely limiting the model's generalization ability in different scenarios. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a rainfall estimation method and system based on integrated communication and sensing, which has higher estimation accuracy.

[0005] To achieve the above objectives, the specific solution adopted by this invention is: a rainfall estimation method based on integrated communication and sensing, comprising: Acquire rainfall monitoring data from rainfall sensing devices pre-deployed at fixed locations, and generate reference weather features based on the rainfall monitoring data; In commercial microwave links, received signal strength data is acquired based on a preset time resolution, and the received signal strength data is standardized to generate raw sensing signals. Multiple time windows of different lengths are used to extract features from the original sensing signal to obtain multiple original feature sequences with different time scales; Multiple original feature sequences are input into a pre-trained multi-scale prediction model. The multi-scale prediction model includes multiple prediction branches built based on the Bi-LSTM model. The prediction branches process the original feature sequences one-to-one based on reference weather features to generate branch rainfall prediction results. The fusion weights for each time scale are determined using the particle swarm optimization algorithm, and the branch rainfall prediction results are weighted and fused based on the fusion weights to obtain the integrated prediction value. The final rainfall estimate is obtained by applying physical constraint filtering to the integrated prediction values ​​based on signal fluctuation metrics.

[0006] As a further optimization of the above-mentioned rainfall estimation method based on integrated communication and sensing: the method for acquiring rainfall monitoring data from rainfall sensing devices pre-deployed at fixed locations and generating reference weather features based on the rainfall monitoring data includes: Rainfall monitoring data within a preset time period is obtained from rainfall sensing devices and combined into historical rainfall statistics: ,in, Here is the rainfall monitoring data at time t. The length of the time period. This represents the maximum value of rainfall monitoring data within the specified time period. Reference weather characteristics are determined based on preset rainfall thresholds and historical rainfall statistics: ; in, This is the rainfall threshold.

[0007] As a further optimization of the above-mentioned rainfall estimation method based on integrated communication and sensing: the original feature sequence is ,in, This represents the original sensory signal at time t. This represents the length of the time window.

[0008] As a further optimization of the above-mentioned rainfall estimation method based on integrated communication and sensing: the method for generating branch rainfall prediction results by processing the original feature sequence based on a one-to-one correspondence of reference weather features includes: The hidden layer of the prediction branch processes the original feature sequence to generate hidden layer features, and maps them to the initial result through a fully connected layer; The modified linear unit activation function (MTU) is used to process the initial results to generate branched rainfall prediction results. The processing method of the modified linear unit activation function is as follows: ; in, For branch rainfall prediction results, Predict the state of the hidden layer in the branch at time t. This is the weight matrix of the fully connected layer. This is a bias term.

[0009] As a further optimization of the above-mentioned rainfall estimation method based on integrated communication and sensing: the multi-scale prediction model includes an expectation loss function, which is: ; in, To represent the actual rainfall intensity, The asymmetric weighting parameters correspond to the time scale of the original feature sequence.

[0010] As a further optimization of the above-mentioned rainfall estimation method based on integrated communication and sensing: the method for obtaining the integrated prediction value by weighting and fusing the branch rainfall prediction results based on fusion weights includes: ; in, , and This is the branch rainfall estimation result when the multi-scale prediction model includes three prediction branches. The fusion weights are determined using the particle swarm optimization algorithm.

[0011] As a further optimization of the above-mentioned rainfall estimation method based on integrated communication and sensing: the method for obtaining the final rainfall estimation result by performing physical constraint filtering on the integrated prediction value based on signal fluctuation measurement is as follows: ; in, Let t be the measure of fluctuation at time t. This is the physical filtering threshold.

[0012] As a further optimization of the above-mentioned rainfall estimation method based on integrated communication and sensing: in the prediction branch, the forget gate and input gate are optimized based on the fluctuation metric, and the optimized gating formula is: ; in, Let t be the measure of fluctuation at time t. The coupling coefficients are learnable.

[0013] As a further optimization of the above-mentioned rainfall estimation method based on integrated communication and sensing, the calculation method of the fluctuation metric is as follows: ; in, The number of data streams for receiving signal strength data. The preset fluctuation statistics window length, Let be the received signal strength of the m-th data stream at time t. Let be the mean of the received signal strength of all data streams at time t, and have .

[0014] A rainfall estimation system based on integrated communication and sensing is used to implement the above-mentioned rainfall estimation method based on integrated communication and sensing. The system includes: The data acquisition module is used to acquire rainfall monitoring data and received signal strength data; The rainfall estimation module is used to estimate the amount of rainfall based on rainfall monitoring data and received signal strength data.

[0015] Beneficial effects: This invention utilizes multiple time windows of varying lengths to extract multi-scale features from the original sensing signal and employs multiple Bi-LSTM prediction branches to process feature sequences at different time scales. This allows the model to simultaneously capture the dynamic changes of rainfall processes at macro, meso, and micro time scales, solving the problem that a single time window cannot adequately capture both long-term trends and short-term abrupt changes, and improving the model's adaptability to different types of rainfall events.

[0016] This invention introduces a particle swarm optimization (PSO) algorithm to adaptively search for the fusion weights of each timescale branch, avoiding the subjectivity and limitations of manually setting weights based on experience. This allows the fusion weights to be automatically optimized according to different regions, seasons, and link conditions, improving the model's generalization ability and rainfall estimation accuracy. In experimental verification, the PSO-optimized fusion weights were 0.350 for the macroscale, 0.453 for the mesoscale, and 0.197 for the microscale. The highest weight at the mesoscale indicates that the signal variation characteristics at medium timescales contribute the most to rainfall estimation.

[0017] This invention employs physical constraint filtering based on signal fluctuation metrics to effectively identify signal anomalies caused by non-rainfall factors such as temperature drift, multipath effects, and electromagnetic interference. This significantly reduces the false alarm rate of rainfall events and improves the reliability of rainfall estimation results. The physical filtering threshold obtained through PSO optimization is 0.1379, and experimental data demonstrates the ability to accurately distinguish between rainfall signal fluctuations and non-rainfall signal fluctuations.

[0018] This invention introduces a signal fluctuation metric as a coupling term into the forget gate and input gate of Bi-LSTM, enabling the model to adaptively adjust the ratio of information retention and updating according to the degree of signal fluctuation. When the signal fluctuation is severe, it automatically suppresses the contamination of the hidden state by noise information, and fully retains historical information when the signal is stable, thereby improving the robustness of the model to complex signal environments.

[0019] This invention introduces reference weather features from rainfall sensing devices as auxiliary information, providing the model with prior knowledge of the actual rainfall conditions. This helps the model establish an accurate mapping relationship between signal attenuation and rainfall intensity during training, accelerating model convergence and improving prediction accuracy. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the structure of a multi-scale prediction model; Figure 3 This is a graph showing the test results of the data from June 11, 2017, in the verification experiment; Figure 4 This is a graph showing the test results of data from June 12, 2017, in the verification experiment. Detailed Implementation

[0021] 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.

[0022] like Figure 1 and Figure 2 As shown, the present invention first provides a rainfall estimation method based on integrated communication and sensing, including S1 to S6.

[0023] S1. Acquire rainfall monitoring data from rainfall sensing devices pre-deployed at fixed locations, and generate reference weather features based on the rainfall monitoring data. The rainfall sensing devices are conventional technologies in the field; for example, a Davis 6322 calibrated weather station equipped with a tipping bucket rain gauge can be used. The temporal resolution of the rainfall sensing devices is determined according to actual needs; for example, a 1-minute temporal resolution can be used. More specifically, the method for acquiring rainfall monitoring data from rainfall sensing devices pre-deployed at fixed locations and generating reference weather features based on the rainfall monitoring data includes S11 to S12.

[0024] S11. Obtain rainfall monitoring data within a preset time period from the rainfall sensing device and combine it into historical rainfall statistics: ,in, Here is the rainfall monitoring data at time t. The length of the time period. This represents the maximum value of rainfall monitoring data within the specified time period.

[0025] S12. Determine reference weather characteristics based on preset rainfall thresholds and historical rainfall statistics: ; in, This is the rainfall threshold. Specifically, when... When the reference weather characteristic is 0, the corresponding real weather is sunny, meaning there is no rain; when When the value is 1, the actual weather corresponding to the reference weather feature is light rainfall; when... When the value is 2, the actual weather corresponding to the reference weather feature is rainfall.

[0026] In some regions, there may be numerous trace rainfall events, meaning frequent rainfall events with an intensity not exceeding 0.1 mm / min. These trace rainfall events have a small impact on microwave signals, but their contribution to cumulative rainfall is not negligible. Therefore, this invention categorizes reference weather characteristics into three types: sunny, trace rainfall, and heavy rainfall. This enables the multi-scale prediction model to more accurately distinguish different rainfall intensity ranges in subsequent processes, improving the estimation capability for weak rainfall events.

[0027] Based on S1, this invention introduces reference weather features based on real rainfall observations, providing a clear weather state prior for multi-scale prediction models, better covering the continuous characteristics of rainfall intensity, reducing the uncertainty of multi-scale prediction models inferring rainfall state solely from RSS signals, and improving the accuracy of rainfall event detection and the precision of rainfall intensity estimation.

[0028] S2. In a commercial microwave link, received signal strength data is acquired based on a preset time resolution, and the received signal strength data is standardized to generate the original sensing signal. The received signal strength is recorded in dBm. In one embodiment of the invention, the time resolution is set to 1 minute to facilitate time alignment with rainfall monitoring data and avoid time misalignment that could reduce the accuracy of the final estimation result.

[0029] S3. Multiple time windows of varying lengths are used to extract features from the original sensory signal, resulting in multiple original feature sequences at different time scales. The original feature sequences are... ,in, This represents the original sensory signal at time t. This represents the length of the time window.

[0030] In one embodiment of the present invention, three time windows of different lengths are used to extract features from the original sensed signal, resulting in three original feature sequences. The lengths of the three time windows are set to 25, 15, and 7, respectively, meaning that the three time windows can extract 25, 15, and 7 data points from the original sensed signal, respectively. The three time windows correspond to macroscopic, mesoscopic, and microscopic time scales, respectively. With a time resolution of 1 minute, the 25 sampling points at the macroscopic timescale correspond to approximately 25 minutes of observation history. The macroscopic timescale branch mainly captures the long-term trend of rainfall processes, such as the continuous attenuation characteristics of frontal rain and the overall evolution trend of rainfall processes. The longer window allows subsequent multi-scale prediction models to observe the complete process of rainfall events from their inception to development, providing sufficient contextual information for rainfall trend prediction. The 15 sampling points at the mesoscopic timescale correspond to approximately 15 minutes of observation history. The mesoscopic timescale branch captures changes in rainfall intensity at intermediate timescales, such as the movement of convective rain clusters and the fluctuations in moderate-intensity rainfall. The mesoscopic timescale can achieve a balance between long-term trends and short-term rapid fluctuations. The 7 sampling points at the microscopic timescale correspond to approximately 7 minutes of observation history. The microscopic timescale branch can capture short-term rapid fluctuations, such as the sudden onset and end of showers and instantaneous changes in rainfall intensity. The shorter window allows subsequent multi-scale prediction models to have high sensitivity to rapid changes in rainfall.

[0031] Based on S3, three time windows of different lengths cover the dynamic characteristics of rainfall at different time scales, achieving full coverage of the dynamic characteristics of the rainfall process across all time scales. This solves the technical problem that a single window length cannot simultaneously capture long-term trends and short-term abrupt changes, laying a data foundation for high-precision predictions by subsequent multi-scale prediction models. It should also be noted that the length of the time window is not fixed and can be adjusted according to the temporal characteristics of rainfall processes in different regions and the sampling rate of the microwave link.

[0032] S4. Multiple original feature sequences are input into a pre-trained multi-scale prediction model. This model includes multiple prediction branches built on a Bi-LSTM model. Each branch processes the original feature sequences based on a reference weather feature to generate its own rainfall prediction result. Each prediction branch employs a two-layer bidirectional LSTM (Bi-LSTM) network architecture. The Bi-LSTM multi-scale prediction model contains a forward LSTM layer and a backward LSTM layer. The forward layer processes the feature sequences in ascending time order, and the backward layer processes them in descending time order. The final hidden states of both layers are concatenated along the feature dimension to form the hidden layer features. Each LSTM layer has 128 neurons, meaning the hidden layer dimension of a unidirectional LSTM is 128. The bi-layer Bi-LSTM design enables the multi-scale prediction model to learn deeper temporal dependencies and capture complex nonlinear patterns in RSS sequences. Additionally, a dropout layer with a dropout rate of 0.2 is added after the Bi-LSTM layers. Dropout is a regularization technique that randomly sets the output of 20% of neurons to zero during training, effectively preventing overfitting in multi-scale prediction models and improving their generalization ability on unseen data. Following the dropout layer is a fully connected layer, which maps the 256-dimensional hidden layer features *h* to 1-dimensional rainfall intensity predictions. During training, each prediction branch is jointly trained, sharing training data but learning parameters suitable for its own time scale.

[0033] Furthermore, the method for generating branched rainfall prediction results by processing the original feature sequence based on a one-to-one correspondence of reference weather features includes S41 to S42.

[0034] S41. The hidden layer of the prediction branch processes the original feature sequence to generate hidden layer features, and maps them to the initial result through a fully connected layer.

[0035] S42. The initial results are processed using the modified linear unit activation function to generate branched rainfall prediction results. The processing method of the modified linear unit activation function is as follows: ; in, For branch rainfall prediction results, Predict the state of the hidden layer in the branch at time t. This is the weight matrix of the fully connected layer. This is the bias term. The ReLU activation function truncates negative predicted values ​​to zero because rainfall intensity cannot physically be negative. Simultaneously, the nonlinear nature of ReLU enhances the ability of multi-scale prediction models to fit complex mapping relationships. The branch rainfall estimates for the multi-scale prediction model, including three prediction branches, are denoted as follows: , and .

[0036] Based on S4, this invention fully captures the contextual dependencies and deep temporal patterns in RSS sequences by utilizing the bidirectional temporal modeling capability of a two-layer Bi-LSTM. The 128-neuron hidden layer provides sufficient capacity for the multi-scale prediction model, and the dropout rate of 0.2 effectively prevents overfitting. The design of fully connected layers with ReLU activation ensures both the physical rationality (non-negativity) of the output and provides nonlinear mapping capability. The architecture design, in which each branch independently processes features at different scales, enables the multi-scale prediction model to learn specialized prediction patterns for the characteristics of each time scale.

[0037] Furthermore, to ensure the accuracy of the multi-scale prediction model, the multi-scale prediction model includes an expectation loss function, which is: ; in, To represent the actual rainfall intensity, The asymmetric weight parameters correspond to the time scale of the original feature sequence. This expected loss function is an asymmetric weighted loss function, its core characteristic being the different penalties for underestimating and overestimating the predicted values. By adjusting... The value of controls the degree of preference of the forecast branch for underestimation and overestimation. When When the value is greater than 0.5, the penalty for underestimation is greater than the penalty for overestimation, and the forecast branch tends to overestimate rainfall intensity to reduce the risk of underreporting.

[0038] In this embodiment, the asymmetric weight parameters of the prediction branches corresponding to the three time scales are... They are respectively set as: the prediction branches corresponding to the macroscopic time scale Equals 0.52, mesoscale branch Equal to 0.80, micro-timescale branch It equals 0.68. Slightly higher than 0.5, so that the forecast branch corresponding to the macro time scale is slightly biased towards conservative forecasts while maintaining basic symmetry, in order to reduce the risk of missed reports in flood warnings; A higher overestimation penalty weight makes the prediction branch corresponding to the meso-time scale pay more attention to avoiding overestimation; By moderately biasing towards underestimating penalties, the prediction branches corresponding to micro-timescales maintain high prediction accuracy in real-time scheduling scenarios. This invention, through setting an asymmetric expectation loss function, enables the multi-scale prediction model to selectively optimize prediction accuracy based on the characteristics of each timescale, with differentiated asymmetric weight parameters. The configuration allows the three prediction branches to achieve optimal prediction performance on their respective time scales.

[0039] S5. The particle swarm optimization algorithm is used to determine the fusion weights for each time scale, and the branch rainfall prediction results are weighted and fused based on these weights to obtain the integrated prediction value. Specifically, the method for weighting and fusing the branch rainfall prediction results based on the fusion weights to obtain the integrated prediction value includes: ; in, , and This is the branch rainfall estimation result when the multi-scale prediction model includes three prediction branches. The fusion weights are determined using the particle swarm optimization algorithm.

[0040] In one embodiment of the present invention, the Particle Swarm Optimization (PSO) algorithm uses 28 particles and 45 iterations. The position of each particle represents a set of candidate fusion weights. The fitness function is defined as the root mean square error (RMSE) between the weighted fusion predictions of that set of weights on the validation set and the actual rainfall intensity. PSO finds the optimal weight combination that minimizes the RMSE by iteratively updating the particle velocity and position. In the optimal weight combination... The values ​​are 0.35, 0.453, and 0.197, respectively. The sum of the three fusion weights is 1, satisfying the normalization constraint.

[0041] Furthermore, the PSO optimization process is performed offline once after the multi-scale prediction model training is completed. In actual implementation, the optimized fusion weights are directly used for weighted fusion, without the need for additional optimization calculations, thus ensuring the inference efficiency of the multi-scale prediction model.

[0042] Based on S5, this invention employs a particle swarm optimization algorithm to efficiently search for near-globally optimal fusion weight combinations in the weight space, avoiding the subjectivity and limitations of manually setting weights based on experience. The adaptive weight fusion mechanism enables the system to automatically adjust the contribution ratio of each scale branch according to the rainfall characteristics of different regions and seasons. Experimental results show that the optimized fusion weight allocation is reasonable and effective, significantly improving the rainfall estimation accuracy of the multi-scale prediction model.

[0043] S6. The final rainfall estimate is obtained by applying physical constraint filtering to the integrated prediction values ​​based on signal fluctuation metrics. Microwave signal attenuation caused by actual rainfall typically exhibits slow, continuous changes, with relatively small fluctuations within a short time window. However, non-rainfall factors such as temperature drift, antenna wind vibration, electromagnetic interference, and equipment failure often cause drastic, irregular signal jumps, resulting in significantly larger fluctuation metrics within a short time window. To eliminate the influence of signal attenuation caused by other factors on the final estimation result, this invention uses signal fluctuation metrics for physical constraint filtering. In one embodiment of this invention, the physical filtering threshold... The fusion weights are also determined synchronously by the particle swarm optimization algorithm during the PSO ensemble optimization phase. In other words, the PSO search space simultaneously includes the fusion weights. and physical filter threshold , The search range is set to 0 to 0.15. In this embodiment, the optimal physical filter threshold θ obtained by PSO optimization is equal to 0.1379. The physical meaning of this threshold is: when the standardized fluctuation metric of the signal within a short time window is lower than 0.1379, the signal fluctuation is considered to be mainly caused by rainfall, and the predicted value is retained; when the fluctuation metric reaches or exceeds 0.1379, the signal fluctuation is considered to be possibly caused by non-rainfall factors, and the predicted value is set to zero to avoid false alarms.

[0044] Furthermore, the method for obtaining the final rainfall estimate by performing physical constraint filtering on the integrated forecast values ​​based on signal fluctuation metrics is as follows: ; in, Let t be the measure of fluctuation at time t. This is the physical filtering threshold. The fluctuation metric is calculated as follows: ; in, The number of data streams for receiving signal strength data. The preset fluctuation statistics window length, Let be the received signal strength of the m-th data stream at time t. Let be the mean of the received signal strength of all data streams at time t, and have In one embodiment of the present invention, The value equals 6, meaning that the two microwave links generate a total of 6 concurrent data streams, with 2 from the 38GHz SISO link and the remaining 4 from the 32GHz MIMO link. All data streams use vertical polarization. It can be set to 10 to 60 sampling points, corresponding to an observation duration of approximately 10 to 60 minutes. The signal fluctuation metric is calculated using the redundancy information from the six data streams, making... It is not sensitive to single-channel anomalies and has good statistical stability.

[0045] In S6, physical constraint filtering can be executed independently at each time step, effectively distinguishing between fluctuations in rainfall signals and fluctuations in non-rainfall signals, minimizing false predictions caused by non-rainfall factors. Based on S6, this invention establishes a security defense line based on the physical characteristics of the signal at the output of the multi-scale prediction model through physical constraint filtering. The physical filtering threshold is automatically optimized and determined by PSO, requiring no manual setting, and can effectively identify and filter false rainfall predictions caused by non-rainfall factors, significantly reducing the false alarm rate of the system. Furthermore, physical constraint filtering does not rely on the discrimination capability within the multi-scale prediction model, but rather makes independent judgments based on the statistical characteristics of the signal, complementing the multi-scale prediction model and improving overall reliability.

[0046] Furthermore, in the prediction branch, gating optimization is performed on the forget gate and input gate based on the volatility metric. The optimized gating formula is as follows: ; in, Let t be the measure of fluctuation at time t. The coupling coefficients are learnable.

[0047] The forget gate in a standard LSTM controls the proportion of information retained from the previous hidden state, while the input gate controls the proportion of new information written into the current input. This invention introduces an additional signal fluctuation metric into the computation of these two gates. As a coupling term. Coupling coefficient. These are parameters that are automatically learned through backpropagation during the training process. The learning value is usually positive, which means that when the signal fluctuation increases, the output of the forget gate increases, and the multi-scale prediction model tends to retain more historical information to resist the impact of instantaneous noise. Typically, the learned value is negative, meaning that as signal fluctuations increase, the input gate output decreases, and the multi-scale prediction model tends to reduce the update of the cell state by the current noisy input. This adaptive gating mechanism allows Bi-LSTM to dynamically adjust the information flow according to signal quality: fully utilizing input information to update the state during signal stability periods, and automatically protecting the cell state from noise contamination during signal fluctuation periods. The sigmoid function compresses the two gate outputs to between 0 and 1, representing the proportion of information retained or passed. This gating optimization can be applied to all timescale prediction branches, or only to the microscale branches most sensitive to signal fluctuations.

[0048] By employing the above technical solution, this invention achieves an adaptive response of the multi-scale prediction model to signal quality by directly embedding signal fluctuation measurement into the LSTM gating mechanism. When signal fluctuations are severe, it automatically reduces the confidence level in the current input and enhances the retention of historical information, effectively preventing noise contamination from hiding the state and continuing to propagate in subsequent moments, thus improving the prediction stability and robustness of the multi-scale prediction model in complex electromagnetic environments.

[0049] The present invention further provides a rainfall estimation system based on integrated communication and sensing, for implementing the above-mentioned rainfall estimation method based on integrated communication and sensing, the system including a data acquisition module and a rainfall estimation module.

[0050] The data acquisition module is used to acquire rainfall monitoring data and received signal strength data.

[0051] The rainfall estimation module is used to estimate the amount of rainfall based on rainfall monitoring data and received signal strength data.

[0052] To verify the effectiveness of the method of this invention, a systematic experimental evaluation was conducted on the microwave link dataset from Gothenburg, Sweden. The experimental dataset was collected in June 2017, with data from June 11th and 12th, 2017, used as the test set, and the remaining data used for training and validation of the multi-scale prediction model. The data was partitioned by date rather than randomly to ensure no temporal data leakage between the training and test sets. Standardized parameters were only fitted on the training set and then applied to all data, ensuring the fairness of the evaluation.

[0053] from Figure 3 and Figure 4As can be seen, the final estimation results of this invention have a high degree of fit with the actual monitored rainfall data, meaning that most rainfall events can be successfully identified by this invention, and the predicted values ​​maintain physical reasonableness. Only a few residual differences occur during peak periods with strong abrupt changes and high intensity—at which point the predicted sequence is smoother than the observed rainfall data. Specifically, in the test data from June 11, 2017, the root mean square error (RMSE) of the method of this invention was 0.1840 mm / min, the mean absolute error (MAE) was 0.0507 mm / min, and the Pearson correlation coefficient (PCC) was 0.9019. In the test data from June 12, 2017, the RMSE was 0.1265 mm / min, the MAE was 0.0327 mm / min, and the PCC was 0.9264. The PCC for both days exceeded 0.90, indicating a high correlation between the rainfall intensity predicted by the multi-scale prediction model and the actual rainfall intensity. The MAE was below 0.06 mm / min, indicating that the absolute error of the minute-by-minute rainfall intensity estimation was very small. Regarding cumulative precipitation, the cumulative precipitation PCC on June 11, 2017 was 0.90, and the cumulative precipitation MAE was 1.29 mm; the cumulative precipitation PCC on June 12, 2017 was 0.93, and the MAE was 1.84 mm. High-precision estimation of cumulative precipitation is of significant practical value for flood early warning and hydrological monitoring. The method of this invention can achieve high-precision cumulative precipitation estimation while ensuring minute-by-minute estimation accuracy.

[0054] In terms of physical verification, experiments observed a significant decrease in received signal strength (dBm) during rainfall events, and the timing of this signal decrease perfectly matched the peak rainfall (mm / min) recorded by the weather station. This physical phenomenon verifies the physical basis and feasibility of using commercial microwave links for rainfall estimation. The above experimental results fully demonstrate the effectiveness and reliability of the method of this invention in practical rainfall estimation tasks, and its predictive performance meets the accuracy requirements of practical engineering applications.

[0055] 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 herein 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 integrated communication and sensing, characterized in that, include: Acquire rainfall monitoring data from rainfall sensing devices pre-deployed at fixed locations, and generate reference weather features based on the rainfall monitoring data; In commercial microwave links, received signal strength data is acquired based on a preset time resolution, and the received signal strength data is standardized to generate raw sensing signals. Multiple time windows of different lengths are used to extract features from the original sensing signal to obtain multiple original feature sequences with different time scales; Multiple original feature sequences are input into a pre-trained multi-scale prediction model. The multi-scale prediction model includes multiple prediction branches built based on the Bi-LSTM model. The prediction branches process the original feature sequences one-to-one based on reference weather features to generate branch rainfall prediction results. The fusion weights for each time scale are determined using the particle swarm optimization algorithm, and the branch rainfall prediction results are weighted and fused based on the fusion weights to obtain the integrated prediction value. The final rainfall estimate is obtained by applying physical constraint filtering to the integrated prediction values ​​based on signal fluctuation metrics.

2. The rainfall estimation method based on integrated communication and sensing as described in claim 1, characterized in that, The method for acquiring rainfall monitoring data from rainfall sensing devices pre-deployed at fixed locations and generating reference weather features based on the rainfall monitoring data includes: Rainfall monitoring data within a preset time period is obtained from rainfall sensing devices and combined into historical rainfall statistics: ,in, Here is the rainfall monitoring data at time t. The length of the time period. This represents the maximum value of rainfall monitoring data within the specified time period. Reference weather characteristics are determined based on preset rainfall thresholds and historical rainfall statistics: ； in, This is the rainfall threshold.

3. The rainfall estimation method based on integrated communication and sensing as described in claim 1, characterized in that, The original feature sequence is ,in, This represents the original sensory signal at time t. This represents the length of the time window.

4. The rainfall estimation method based on integrated communication and sensing as described in claim 1, characterized in that, The method for generating branched rainfall prediction results by processing the original feature sequence based on a one-to-one correspondence with reference weather features includes: The hidden layer of the prediction branch processes the original feature sequence to generate hidden layer features, and maps them to the initial result through a fully connected layer; The modified linear unit activation function (MTU) is used to process the initial results to generate branched rainfall prediction results. The processing method of the modified linear unit activation function is as follows: ； in, For branch rainfall prediction results, Predict the state of the hidden layer in the branch at time t. This is the weight matrix of the fully connected layer. This is a bias term.

5. The rainfall estimation method based on integrated communication and sensing as described in claim 4, characterized in that, The multi-scale prediction model includes an expected loss function, which is: ； in, To represent the actual rainfall intensity, The asymmetric weighting parameters correspond to the time scale of the original feature sequence.

6. The rainfall estimation method based on integrated communication and sensing as described in claim 4, characterized in that, The method for obtaining an integrated prediction value by weighted fusion of branch rainfall prediction results based on fusion weights includes: ； in, , and This is the branch rainfall estimation result when the multi-scale prediction model includes three prediction branches. The fusion weights are determined using the particle swarm optimization algorithm.

7. The rainfall estimation method based on integrated communication and sensing as described in claim 6, characterized in that, The method for obtaining the final rainfall estimate by performing physical constraint filtering on the integrated prediction values ​​based on signal fluctuation metrics is as follows: ； in, Let t be the measure of fluctuation at time t. This is the physical filtering threshold.

8. The rainfall estimation method based on integrated communication and sensing as described in claim 1, characterized in that, In the prediction branch, gating optimization is performed on the forget gate and input gate based on the volatility metric. The optimized gating formula is as follows: ； in, Let t be the measure of fluctuation at time t. The coupling coefficients are learnable.

9. The rainfall estimation method based on integrated communication and sensing as described in claim 8, characterized in that, The method for calculating the volatility metric is as follows: ； in, The number of data streams for receiving signal strength data. The preset fluctuation statistics window length, Let be the received signal strength of the m-th data stream at time t. Let be the mean of the received signal strength of all data streams at time t, and have .

10. A rainfall estimation system based on integrated communication and sensing, characterized in that, The system is used to implement the rainfall estimation method based on communication and sensing integration as described in any one of claims 1-9, the system comprising: The data acquisition module is used to acquire rainfall monitoring data and received signal strength data; The rainfall estimation module is used to estimate the amount of rainfall based on rainfall monitoring data and received signal strength data.

Images