Hierarchical rainfall intensity inversion method and system based on 6g backhaul network

Abstract

The application relates to the technical field of rainfall inversion, and provides a hierarchical rainfall intensity inversion method and system based on a 6G backhaul network, wherein the hierarchical rainfall intensity inversion method based on the 6G backhaul network comprises the following steps: obtaining a microwave measurement vector in the 6G backhaul network, and performing pre-processing on the microwave measurement vector to generate an original feature vector; inputting the original feature vector into a pre-trained comprehensive inversion model for processing to obtain a rainfall intensity inversion result, wherein the comprehensive inversion model comprises a regression model, a gate trigger, a peak processing module and a decision output module, and the regression model comprises a state detection branch trained by using a soft label mechanism and a hierarchical estimation branch constructed based on a probability weighted fusion strategy. The application provides the hierarchical rainfall intensity inversion method and system based on the 6G backhaul network, the precision of the inversion result is higher, and the probability of missing report of extreme rainfall can be significantly reduced.

Description

Technical Field

[0001] This invention relates to the field of rainfall intensity inversion technology, specifically a hierarchical rainfall intensity inversion method and system based on a 6G backhaul network. Background Technology

[0002] Accurate and timely rainfall estimation is crucial for urban early warning, water resource management, and numerical weather prediction. Traditional monitoring relies on rain gauges and weather radar, but the density of rain gauges is insufficient, and low-altitude radar detection is limited, especially in densely populated urban areas, making it difficult to meet the needs of high-resolution, low-cost monitoring.

[0003] With the development of 6G, IoT, and AI, integrated sensing and communication have become key technologies for next-generation networks. Microwave link monitoring technology, based on communication links, utilizes existing communication infrastructure for continuous atmospheric monitoring without requiring additional hardware. Its rainfall monitoring principle is clear: raindrops attenuate microwave signals, and the path-average rainfall intensity can be estimated by inverting the attenuation. The E-band, widely used in 5G backhaul and smart cities, has become a powerful sensing resource for urban rainfall monitoring. Although this band's signals are susceptible to attenuation by vegetation and rainfall, dense deployment in cities provides a high spatial resolution advantage without increasing hardware costs. However, this high sensitivity also makes the signal susceptible to non-stationary interference such as wet antenna attenuation and baseline drift, making effective separation of rainfall signals from noise a significant challenge.

[0004] Existing estimation methods can be divided into two categories: model-driven and data-driven. Model-driven methods mainly use power-law models, which are physically well-defined but require accurate separation of rain-induced attenuation and baseline changes, the latter of which is easily affected by various factors and difficult to calibrate. Data-driven methods utilize deep learning and other techniques to learn mappings from data, and have adopted tree models, LSTM, CNN, and attention mechanisms, but still face two fundamental challenges in practical prediction.

[0005] First, the boundary between wet and dry conditions is blurred. The attenuation from minute amounts of rainfall is often lower than the noise floor, and coupled with the wet antenna effect and quantization errors, baseline fluctuations lead to extremely high false detection rates. Existing classification models rely on hard labels, forcing binary judgments at ambiguous points, ignoring physical continuity, and introducing label noise.

[0006] Secondly, peak values ​​are underestimated. When the path length reaches several kilometers, the attenuation value is spatially averaged, smoothing out localized heavy rainfall. This results in an extreme long-tail distribution of training samples, with a very low proportion of high-intensity events. Global optimizations such as MSE cause the model to tend towards mean regression, leading to a systematically low amplitude during heavy rainfall and a delayed response, thus affecting the effectiveness of early warnings.

[0007] In summary, the main shortcomings of existing technologies are: the lack of an effective dry / wet discrimination mechanism, making it difficult to reliably distinguish between no rain and light rainfall, resulting in a high false alarm rate; the inability to overcome the systematic underestimation of high-intensity rainfall peaks caused by path smoothing effects and long-tail distribution; and the model response delay, which fails to meet the accuracy and timeliness requirements of urban rainstorm warnings, thus limiting their practical application value. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a hierarchical rainfall intensity inversion method and system based on a 6G backhaul network, which yields higher accuracy inversion results and can significantly reduce the probability of missed extreme rainfall reports.

[0009] To achieve the above objectives, the specific solution adopted by this invention is: a graded rainfall intensity inversion method based on a 6G backhaul network, comprising: Obtain microwave measurement vectors from the 6G backhaul network, and perform preprocessing on the microwave measurement vectors to generate original feature vectors; The original feature vectors are input into a pre-trained integrated inversion model for processing to obtain the rainfall intensity inversion result. The integrated inversion model includes a regression model, a gating trigger, a peak processing module, and a decision output module. The regression model includes a state detection branch trained using a soft label mechanism and a hierarchical estimation branch constructed based on a probability-weighted fusion strategy. The methods used by the integrated inversion model to process the original feature vectors include: The state detection branch of the regression model is used to determine the probabilities of multiple rainfall levels based on the original feature vector, and the classification estimation branch of the regression model is used to determine the basic rainfall estimate based on the probabilities of multiple rainfall levels. Using gated triggers, the system determines whether extreme peak rainfall has occurred based on baseline rainfall estimates. When extreme peak rainfall occurs, the peak processing module is used to process the peak portion of the basic rainfall estimate to obtain the optimized estimate process value; The decision output module generates rainfall intensity inversion results based on the optimized estimation process values.

[0010] As a further optimization of the above-mentioned graded rainfall intensity inversion method based on 6G backhaul network: the microwave measurement vector includes the attenuation value of the first frequency band, the attenuation value of the second frequency band, the carrier interference noise ratio of the first frequency band, and the carrier interference noise ratio of the second frequency band.

[0011] As a further optimization of the aforementioned graded rainfall intensity inversion method based on 6G backhaul network, the method for preprocessing microwave measurement vectors to generate original feature vectors includes: Based on the microwave measurement vectors within the first time window prior to the current time, calculate the rolling statistical characteristics within the first time window; Calculate the cross features between data of different dimensions in the microwave measurement vector; Based on the microwave measurement vector within the second time window prior to the current moment, the difference feature is obtained by subtracting the measurement values ​​of the same dimension data at the current moment from those at historical moments; Feature extraction is performed on the microwave measurement vector at the current moment to obtain the original signal features; Obtain the time-coding information of the current moment to obtain the time-coding features; The rolling statistical features, the original signal features, the cross features, the differential features, and the time-coded features are fused to generate the original feature vector.

[0012] As a further optimization of the above-mentioned graded rainfall intensity inversion method based on 6G backhaul network: the soft label mechanism is used to generate three-level soft labels based on the comparison results of real rainfall intensity data and a predetermined threshold set. The three-level soft labels include a first label value representing the dry state, a second label value representing the dry-wet transition state, and a third label value representing the wet state. During the training of the regression model, the state detection branch is trained using the three-level soft labels as the supervision target, and the predetermined threshold is dynamically calibrated through an optimization algorithm. The optimization algorithm uses the weighted sum of false alarm rate, detection probability and mean square error as the fitness function to jointly search and optimize the predetermined threshold.

[0013] As a further optimization of the above-mentioned graded rainfall intensity inversion method based on a 6G backhaul network: the graded estimation branch includes multiple graded regression units, and each graded regression unit corresponds to a rainfall level probability. The method for the graded estimation branch to determine the basic rainfall estimate based on the multiple rainfall level probabilities includes: The original feature vector is input into multiple hierarchical regression units for processing to obtain multiple rainfall intensity estimation sub-values; The basic rainfall estimate is obtained by weighting and summing the rainfall intensity estimates using the probability of rainfall level as the weight.

[0014] As a further optimization of the above-mentioned graded rainfall intensity inversion method based on 6G backhaul network: the method of using a gated trigger to determine whether extreme peak rainfall has occurred based on the baseline rainfall estimate includes: Obtain multiple historical baseline rainfall estimates for a predetermined time period prior to the current moment; The time-series variation is obtained by performing a difference operation on the multiple historical baseline rainfall estimates; The cumulative estimate is obtained by summing the multiple historical baseline rainfall estimates. The probability of high-intensity rainfall is obtained by taking the highest probability or the sum of the two highest probabilities among the multiple rainfall level probabilities. The high-intensity rainfall probability, the temporal variation, and the cumulative estimate are input into a pre-trained binary classifier to obtain the trigger flag; When the trigger flag is in the first state, it is determined that extreme peak rainfall has occurred; when the trigger flag is in the second state, it is determined that no extreme peak rainfall has occurred.

[0015] As a further optimization of the above-mentioned graded rainfall intensity inversion method based on 6G backhaul network: the peak processing module includes a time-series residual correction unit and an adaptive fusion unit; Methods for obtaining optimized estimated process values ​​by processing the peak portion of the basic rainfall estimate using a peak processing module when extreme peak rainfall occurs include: The time-series residual correction unit is used to perform state-space modeling of the historical feature sequence within a predetermined time window before the current time, extract the time-series dependency, and output the residual correction amount. The residual correction is added to the baseline rainfall estimate to obtain the corrected rainfall estimate; The adaptive fusion unit is used to perform multi-branch nonlinear mapping and adaptive fusion on the corrected rainfall estimate to obtain the optimized estimation process value. The historical feature sequence includes the original feature vectors of each historical moment within a predetermined time window prior to the current moment, as well as the basic rainfall estimate corresponding to each historical moment.

[0016] As a further optimization of the above-mentioned graded rainfall intensity inversion method based on 6G backhaul network: the time-series residual correction unit is constructed based on the Mamba state-space model; The Mamba state-space model performs causal sequence modeling on the input historical feature sequence, and dynamically adjusts the state transition according to the input content through a selective scanning mechanism to extract long-range temporal dependencies and output residual correction. The adaptive fusion unit includes a depth branch, a width branch, and a residual branch; The depth branch performs a nonlinear transformation on the input through cascaded fully connected layers, the width branch performs a linear transformation on the input through a single fully connected layer, and the residual branch is used to preserve the original information of the input features. The adaptive fusion unit performs weighted fusion of the outputs of the depth branch, the width branch, and the residual branch using learnable gating weights.

[0017] As a further optimization of the above-mentioned graded rainfall intensity inversion method based on 6G backhaul network: when the gated trigger determines that no extreme peak rainfall has occurred, the decision output module outputs the basic rainfall estimate as the rainfall intensity inversion result; when the gated trigger determines that extreme peak rainfall has occurred, the decision output module outputs the optimized estimation process value as the rainfall intensity inversion result.

[0018] A graded rainfall intensity retrieval system based on a 6G backhaul network is used to implement the aforementioned graded rainfall intensity retrieval method based on a 6G backhaul network. The system includes: The data acquisition module is used to acquire microwave measurement vectors in the 6G backhaul network; The data processing module is used to preprocess the microwave measurement vector to generate the original feature vector, and to process the original feature vector using a pre-trained integrated inversion model to obtain the rainfall intensity inversion result; The human-computer interaction module is used to display the rainfall intensity inversion results.

[0019] Beneficial effects: First, this invention replaces the traditional binary hard label by introducing a three-level soft label mechanism, representing the dry state, dry-wet transition state, and wet state as continuous label values. This enables the model to learn uncertainty in the ambiguous dry-wet boundary region rather than being forced to make deterministic decisions. Combined with the optimization algorithm, the boundary threshold is dynamically calibrated using the weighted sum of false alarm rate, detection probability, and mean square error as the fitness function. This effectively solves the technical problem of high false alarm rate caused by the difficulty in distinguishing between trace rainfall and noise in the prior art. Second, this invention constructs a time-series residual correction unit using the Mamba state-space model, utilizes its selective scanning mechanism to perform linear complexity causal sequence modeling on historical feature sequences, extracts long-range time-series dependencies and outputs residual correction amounts to compensate for the peak portion of the base rainfall estimate, and, in conjunction with a gating trigger, determines whether to trigger peak processing based on the probability of high-intensity rainfall, the amount of time-series change and the cumulative estimate, effectively solving the technical problem of systematic underestimation of high-intensity rainfall peaks caused by path smoothing effects and long-tail distribution in the prior art; Third, this invention constructs an adaptive fusion unit that includes a deep branch, a wide branch, and a residual branch. It dynamically determines the fusion ratio of each branch output with learnable gating weights, so that the model is dominated by the wide branch under stable or light rain conditions to ensure computational efficiency. Under extreme rainfall conditions, the deep branch is automatically activated to handle complex signal abrupt changes. Combined with the decision output module, the output base rainfall estimate or the optimized estimation process value is selected based on the judgment result of the gating trigger. This makes it possible to maintain lightweight computation for about 95% of non-peak periods and activate the high-complexity module only in extreme events, thus balancing estimation accuracy and computational efficiency. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the integrated inversion model processing method; Figure 3 This is a schematic diagram of the overall structure of the integrated inversion model. 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 Figures 1 to 3 As shown, the present invention first provides a graded rainfall intensity inversion method based on a 6G backhaul network, including S1 to S2.

[0023] S1. Obtain the microwave measurement vector in the 6G backhaul network, and perform preprocessing on the microwave measurement vector to generate the original feature vector. The microwave measurement vector includes data in multiple dimensions, specifically including the attenuation value of the first frequency band, the attenuation value of the second frequency band, the carrier-to-noise ratio (CNR) of the first frequency band, and the CNR of the second frequency band. In one embodiment of the present invention, the first frequency band is the 71GHz band, the second frequency band is the 81GHz band, and the microwave measurement vector can be specifically represented as follows: ,in, This refers to the signal attenuation value for the first frequency band, which is the 71GHz band. This refers to the signal attenuation value for the second frequency band, which is the 81GHz band. This refers to the carrier interference-to-noise ratio (CNR) of the first frequency band, which is the 71GHz band. This is the carrier interference noise ratio of the second frequency band, which is the 81GHz band. The superscript T indicates the vector transpose, and the subscript t indicates the current time. This vector forms the original data basis for all subsequent feature construction and model inference.

[0024] Furthermore, the method for preprocessing the microwave measurement vector to generate the original feature vector includes S11 to 16.

[0025] S11. Based on the microwave measurement vectors within the first time window prior to the current time, calculate the rolling statistical features within the first time window. In one embodiment of the present invention, the first time window includes six parts: 5 minutes, 10 minutes, 20 minutes, 30 minutes, 60 minutes, and 120 minutes. Each part corresponds to three sub-features: mean, standard deviation, and maximum value. Therefore, the rolling statistical features are 18-dimensional and can be represented as follows: ; in, This represents the length of each section of the first time window, in minutes. The mean, Standard deviation, The maximum value, This refers to the dimensional concatenation operation. Rolling statistical features can characterize the distribution characteristics of each dimension of data in the microwave measurement vector in the time domain from three perspectives: central tendency, dispersion, and extreme values.

[0026] S12. Calculate the cross-features between different dimensions of data in the microwave measurement vector. In one embodiment of the present invention, the cross-features are obtained by multiplying each dimension of data in the microwave measurement vector pairwise. The cross-features can capture the cooperative change relationship between different dimensions of data in the microwave measurement vector. For example, when rainfall causes the attenuation value to increase and the carrier interference noise ratio to decrease synchronously, the product of the two will change sharply. Based on this cooperative change relationship, single-channel noise interference can be effectively suppressed, enhancing the accuracy of the inversion results of rainfall intensity. Furthermore, the cross-features are 14-dimensional, specifically represented as follows: .

[0027] S13. Based on the microwave measurement vector within the second time window prior to the current moment, the difference between the measurement values ​​of the same dimension data at the current moment and historical moments is calculated to obtain the differential feature. The essence of the differential feature is the approximation of the first derivative of the signal in the time domain, which can effectively capture the start and end times of rainfall, as well as the rate of intensity change. When rainfall begins, the differential feature exhibits a large positive value; when rainfall ends, the differential feature exhibits a large negative value; during continuous and stable rainfall, the differential feature approaches zero. In one embodiment of the present invention, the second time window includes two parts, namely 1 minute and 5 minutes, and the corresponding differential feature also includes two parts, denoted as... and Furthermore, because the microwave measurement vector is 4-dimensional, the difference feature between the two parts is also 4-dimensional. The difference feature can be uniformly represented as: .

[0028] S14. Extract features from the microwave measurement vector at the current moment to obtain the original signal features. Specifically, the original signal features are also 4-dimensional and can be represented as: .

[0029] S15. Obtain the time-coded information of the current moment to obtain the time-coded features. Specifically, the time-coded information includes three parts: hour, minute, and day of the week. The hour and minute are encoded using either sine or cosine to form 2D features, thus the time-coded features are 5-dimensional. By mapping the hour and minute to a periodic circle, feature distance distortion caused by numerical discontinuities between 11 PM and midnight can be avoided, allowing for a more accurate determination of the distribution patterns of rainfall in diurnal and seasonal cycles, such as the characteristics of peak afternoon convective rainfall periods. The time-coded features can be represented as: .

[0030] S16. The rolling statistical features, original signal features, cross features, difference features, and time-coded features are fused to generate the original feature vector. Specifically, the original feature vector consists of 4-dimensional original signal features, 8-dimensional difference features, 18-dimensional rolling statistical features, 5-dimensional time-coded features, and 14-dimensional cross features, totaling 49 dimensions. The original feature vector can be represented as: .

[0031] S2. The original feature vector is input into a pre-trained integrated inversion model for processing to obtain the rainfall intensity inversion result. The integrated inversion model includes a regression model, a gating trigger, a peak processing module, and a decision output module. The regression model includes a state detection branch trained using a soft labeling mechanism and a hierarchical estimation branch constructed based on a probability-weighted fusion strategy. The soft labeling mechanism is used to generate three levels of soft labels based on the comparison results between the actual rainfall intensity data and a predetermined threshold set. The three levels of soft labels include a first label value representing the dry state, a second label value representing the dry-wet transition state, and a third label value representing the wet state. During the training of the regression model, the three levels of soft labels are used as the supervision target to train the state detection branch, and the predetermined threshold is dynamically calibrated through an optimization algorithm. The optimization algorithm uses the weighted sum of the false alarm rate, detection probability, and mean square error as the fitness function to jointly search and optimize the predetermined threshold.

[0032] The soft tagging mechanism can be represented as: ; Among them, transition conditions The definition of is: ; and then, For instantaneous rainfall intensity, and The instantaneous rainfall intensity threshold. and The cumulative rainfall threshold is defined as follows: The predetermined threshold set is... . The value represents the cumulative rainfall over the 60 minutes preceding the current moment. 0, 0.5, and 1.0 are the first, second, and third label values, respectively. In the soft-label mechanism, by introducing a second label value as a transition, the integrated inversion model no longer makes a rigid binary decision in the fuzzy interval between dry and wet states. Instead, the second label value preserves the uncertainty of the fuzzy interval, effectively reducing the probability of erroneous estimation.

[0033] Furthermore, to ensure the accuracy of the three-level soft labels generated by the soft labeling mechanism, this invention utilizes a fitness function and an optimization algorithm to jointly optimize the predetermined threshold set. The fitness function is expressed as: ; in, For false alarm rate, For detection probability, Mean square error, The weights are denoted as . In one embodiment of the invention, the optimization algorithm employs a modified whale optimization algorithm, with a population size of 20 and a maximum number of iterations of 15. The initial values ​​for are 2.0, 1.5, and 1.0, respectively. By minimizing this fitness function, the integrated inversion model can find the optimal balance between suppressing baseline noise and maintaining meteorological sensitivity, thereby effectively improving the accuracy of rainfall inversion results.

[0034] Furthermore, the methods for processing the original feature vectors in the integrated inversion model include S21 to S24.

[0035] S21. The state detection branch of the regression model determines multiple rainfall level probabilities based on the original feature vector, and the hierarchical estimation branch of the regression model determines the basic rainfall estimate based on these multiple rainfall level probabilities. The hierarchical estimation branch includes multiple hierarchical regression units, and each hierarchical regression unit corresponds to a rainfall level probability. The method for determining the basic rainfall estimate based on multiple rainfall level probabilities using the hierarchical estimation branch includes: inputting the original feature vector into multiple hierarchical regression units for processing to obtain multiple rainfall intensity estimates; and weighting and summing the rainfall intensity estimates using the rainfall level probabilities as weights to obtain the basic rainfall estimate. More specifically, the method for weighting and summing the rainfall intensity estimates using the rainfall level probabilities as weights to obtain the basic rainfall estimate is as follows: ; in, The output of the state detection branch is the probability of the rainfall level, and it satisfies... , The rainfall intensity estimate sub-value output by the i-th hierarchical regression unit is... This is the baseline rainfall estimate. In one embodiment of the invention, the hierarchical regression unit employs an XGBoost regressor. Each hierarchical regression unit can be specifically optimized and trained for different rainfall intensity level ranges, avoiding the problem of uneven performance of a single global hierarchical regression unit across different intensity ranges.

[0036] S22. Use a gating trigger to determine whether extreme peak rainfall has occurred based on the baseline rainfall estimate. Specifically, the method for using a gating trigger to determine whether extreme peak rainfall has occurred based on the baseline rainfall estimate includes S221 to S225.

[0037] S221. Obtain multiple historical baseline rainfall estimates for a predetermined time period prior to the current time. More specifically, there are two predetermined time periods: 30 minutes and 60 minutes.

[0038] S222. Perform a difference operation on multiple historical baseline rainfall estimates to obtain the time-series variation. The time-series variation reflects the rate of change of the baseline rainfall estimates over time, and can be denoted as... ,in The change in historical baseline rainfall estimates. The time series variation is the length of a predetermined time period. The time series variation is a key indicator for identifying extreme rainfall; when the time series variation exceeds a certain threshold, it usually indicates the occurrence of an extreme rainfall event.

[0039] S223. The cumulative estimate is obtained by summing multiple historical baseline rainfall estimates. The cumulative estimate characterizes the sustained intensity of rainfall over time. It effectively distinguishes between transient noise spikes and persistent real rainfall events, avoiding misinterpretation of occasional signal fluctuations as extreme rainfall. More specifically, the cumulative estimates for two predetermined time periods can be denoted as follows: and , S224. Obtain the highest probability of multiple rainfall levels, or the sum of the probabilities of the two highest levels, as the probability of high-intensity rainfall. In other words, it can be used alone. As a probability of high-intensity rainfall, or using As a probability of high-intensity rainfall.

[0040] S225. Input the probability of high-intensity rainfall, its temporal variation, and the cumulative estimate into a pre-trained binary classifier to obtain the trigger flag. The trigger flag can be represented as: When the trigger flag is in the first state, that is, when... When the trigger flag is in the second state, it is determined that an extreme peak rainfall has occurred. At that time, it was determined that no extreme peak rainfall had occurred. In one embodiment of the present invention, the binary classifier is implemented using an XGBoost classifier.

[0041] Furthermore, during the comprehensive inversion model processing, as each processing module is executed sequentially, the original feature vector expands from 49 dimensions to 61 dimensions. Specifically, when the state detection branch of the regression model completes processing, it outputs the probabilities of four rainfall levels. Then, a 4-dimensional hierarchical probability feature is obtained. After the tiered estimation branch outputs the baseline rainfall estimates for each time period, the cumulative estimates for 30 minutes, 60 minutes, and 120 minutes are obtained by summing them backwards from the current time. , forming cumulative characteristics Once the gated trigger and its internal submodules complete burst detection and peak detection, a flag characteristic is formed. The 4-dimensional hierarchical probability features, the 6-dimensional cumulative features, and the 2-dimensional marker features together constitute the newly added 12-dimensional features. These are then fused with the original feature vector to obtain a 61-dimensional comprehensive feature vector, which can be represented as follows: This 61-dimensional integrated feature vector, by fusing multi-dimensional information such as instantaneous values, trends, statistical distributions, nonlinear interactions, periodicity, cumulative effects, and decision feedback, provides a comprehensive signal representation for the integrated inversion model. All features rely solely on observation data from the current and historical moments, strictly satisfying real-time causal inference constraints, and is suitable for online deployment scenarios in 6G backhaul networks.

[0042] S23. When extreme peak rainfall occurs, the peak processing module is used to process the peak portion of the baseline rainfall estimate to obtain an optimized estimate. The peak processing module includes a temporal residual correction unit and an adaptive fusion unit. The temporal residual correction unit is built based on the Mamba state-space model. The Mamba state-space model performs causal sequence modeling on the input historical feature sequence and dynamically adjusts the state transitions according to the input content through a selective scanning mechanism to extract long-range temporal dependencies and output residual correction values. The adaptive fusion unit includes a depth branch, a width branch, and a residual branch. The depth branch performs a nonlinear transformation on the input through cascaded fully connected layers, the width branch performs a linear transformation on the input through a single fully connected layer, and the residual branch is used to preserve the original information of the input features. The adaptive fusion unit performs weighted fusion of the outputs of the depth branch, width branch, and residual branch using learnable gating weights.

[0043] The adaptive fusion unit uses learnable gating weights to weight and fuse the outputs of the depth branch, width branch, and residual branch as follows: ; in, , and These are the output features of the depth branch, width branch, and residual branch, respectively. For the set of learnable gating weights, Use the gate weights and satisfy the following conditions: The output can be represented as: ; in, and These represent the internal learnable weights and biases of each branch. This adaptive fusion mechanism dynamically adjusts the contribution ratio of each branch based on the current rainfall status through gated weights. Under stable or light rainfall conditions, the width branch dominates, ensuring computational efficiency and generalization performance; under extreme rainfall conditions, the depth branch is activated, handling complex signal abrupt changes through multi-layer nonlinear transformations. This dynamic routing strategy enables the integrated inversion model to maintain lightweight operation under normal conditions and automatically utilize high-capacity computing resources during extreme events.

[0044] Furthermore, the method for processing the peak portion of the basic rainfall estimate using the peak processing module to obtain the optimized estimate process value when extreme peak rainfall occurs includes S231 to S234.

[0045] S231. Using a temporal residual correction unit, perform state-space modeling on the historical feature sequence within a predetermined time window prior to the current time, extract temporal dependencies, and output the residual correction amount. The residual correction amount can be denoted as... The residual correction is essentially the temporal momentum information extracted by the Mamba model from the historical sequence, used to compensate for the systematic underestimation bias of the basic estimate during the peak phase.

[0046] S232. The residual correction is superimposed on the baseline rainfall estimate to obtain the corrected rainfall estimate. The superposition process can be expressed as: ,in This is the corrected rainfall estimate.

[0047] S233. The corrected rainfall estimate is obtained by performing multi-branch nonlinear mapping and adaptive fusion using an adaptive fusion unit.

[0048] S234. The historical feature sequence includes the original feature vectors of each historical moment within a predetermined time window prior to the current moment, and the corresponding baseline rainfall estimate for each historical moment. This historical feature sequence can be represented as: ; Where L is the length of the predetermined time window, and in one embodiment of the present invention, L is 30 minutes, and d is the feature dimension of each time point in the sequence. This historical feature sequence constitutes the input of the Mamba state-space model, enabling it to model long-range temporal dependencies and output residual corrections. The 30-minute window setting balances the ability to capture long-range dependencies with the requirements of real-time inference latency.

[0049] Furthermore, the continuous-time dynamics of the Mamba state-space model can be defined as: ; in, For the hidden state vector, , and Let be the learnable state transition matrix. After applying zero-order preserved discretization, we can obtain the discrete-time recursive form, specifically: ; in, Let be the discretized state transition matrix, and , To discretize the input matrix, and , For time step, It is the identity matrix. This discrete-time recursive form allows the integrated inversion model to directly process discrete data while maintaining its ability to model continuous-time dynamics.

[0050] S24. The decision output module generates rainfall intensity inversion results based on the optimized estimation process values. Specifically, when the gated trigger determines that no extreme peak rainfall has occurred, the decision output module outputs the basic rainfall estimate as the rainfall intensity inversion result. When the gated trigger determines that extreme peak rainfall has occurred, the decision output module outputs the optimized estimation process values ​​as the rainfall intensity inversion result.

[0051] Furthermore, the method by which the decision output module generates rainfall intensity inversion results can be expressed as follows: ,in, The optimized estimation process value is output by the adaptive fusion unit.

[0052] The gating mechanism of the decision output module enables the integrated inversion model to maintain computational lightweight and high inference speed in most off-peak periods, and only activates the highly complex peak processing module during extreme rainfall events, thereby minimizing computational overhead while maintaining accuracy.

[0053] In summary, the overall processing of the original feature vector by the integrated inversion model can be represented as follows: ; in, Indicates the state detection branch, Indicates the hierarchical estimation branch, This refers to the timing residual correction unit in the peak processing module. Represents the fusion operator, This indicates a gated trigger.

[0054] Furthermore, to ensure the accuracy of the comprehensive inversion model, the training process can be supervised and optimized using a composite loss function. This composite loss function includes a binary cross-entropy loss to control the state detection branch, a multi-class cross-entropy loss to control the hierarchical estimation branch, and a mean squared error loss to control the regression estimation. The binary cross-entropy loss, multi-class cross-entropy loss, and mean squared error loss are expressed as follows: ; ; .

[0055] Furthermore, during training, the state detection branch is first trained with three-level soft labels as the supervision target. After convergence, the parameters are fixed. Then, the hierarchical estimation branch is trained with the probabilities output by the state detection branch as weights. Finally, the temporal residual correction unit and adaptive fusion unit are trained based on the hierarchical estimation branch. This phased and progressive strategy effectively avoids gradient interference and feature leakage problems in multi-task joint optimization.

[0056] To verify the practical effect of the graded rainfall intensity inversion method based on 6G backhaul network proposed in this invention, this embodiment conducted a systematic experimental verification on a real-world deployed E-band microwave link.

[0057] Experimental data were collected from an operational-grade E-band point-to-point microwave backhaul link in Nanjing, China. The link operates in dual-band frequencies of 71 GHz and 81 GHz, has a length of 3 km, a transmit power of +7 dBm, an antenna gain of 50 dBi, QPSK modulation, and a signal level resolution of 0.1 dB. Ground-level rainfall intensity reference data was provided synchronously by the CLIMA laser raindrop spectrometer deployed at the same site. This instrument provides high temporal resolution microscale rainfall intensity measurements as ground truth for model training and evaluation. Data collection spanned four months, with data from December 2019 to February 2020 used as the training set and data from March 2020 used as the test set. The test set included five independent rainfall events with maximum rainfall rates of approximately 6 mm / h, 2.5 mm / h, 6 mm / h, 5 mm / h, and 10 mm / h, covering a range of rainfall intensities from light to extreme heavy rainfall.

[0058] To comprehensively evaluate the performance of the method of this invention, multiple comparative baselines were set up in the experiment, including: a traditional physical baseline method based on the ITU-RP.838 power-law model; an end-to-end deep time series modeling method based on a standard LSTM network; a deep learning rainfall inversion method based on the TabGRU architecture; and an ablation baseline that only includes hierarchical regression but does not include a soft label mechanism and peak processing module. The evaluation metrics used were root mean square error, mean absolute error, Pearson correlation coefficient, and bias. In addition, false alarm rate, detection probability, and key success index were evaluated for the dry and wet state classification task.

[0059] Table 1 presents a comparison of the global prediction performance of the proposed method and various baselines across the entire test set. As shown in Table 1, the proposed graded rainfall intensity retrieval method achieves a global RMSE of 0.4137 mm / h, a global MAE of 0.1119 mm / h, and a PCC of 0.8749. Compared to the standard LSTM network, the proposed method reduces the RMSE by approximately 23.6%, and also shows a significant advantage over the TabGRU architecture. Compared to the ITU-R P.838 physical baseline, the proposed method reduces the RMSE by approximately 48.4%, and the ITU-R baseline's R... 2 A negative value indicates that the pure physical model is no longer able to effectively fit the mapping relationship between the signal and rainfall when faced with non-stationary measured data.

[0060] Table 1 Comparison of Global Prediction Performance of Various Methods

[0061] Furthermore, to verify the performance of the method of the present invention under different rainfall intensities, the experiment divided the test set into multiple intervals according to the actual rainfall intensities for stratified evaluation. Table 2 shows the RMSE comparison results of each method in different rainfall intensity intervals.

[0062] Table 2. Comparison of RMSE results for different methods in different rainfall intensity ranges.

[0063] As shown in Table 2, in the high-intensity rainfall range (greater than 2.5 mm / h), the RMSE of the method of this invention is 1.9248 mm / h, significantly better than DLinear's 2.2116 mm / h, TabGRU's 2.9553 mm / h, and LSTM's 3.1063 mm / h. A one-sided Wilcoxon signed-rank test was performed on the paired absolute errors of each method and the method of this invention in the heavy rain range, with p-values ​​of 0.0981 and 2.75 × 10⁻⁶, respectively. -13 and 4.06×10 -16 The results indicate that the improvement of the method in this invention compared to TabGRU and LSTM is highly statistically significant. These results demonstrate that the peak processing module proposed in this invention can effectively overcome the systematic underestimation defect of traditional methods under heavy rainfall conditions. In the moderate rainfall range (1.5, 2.5] mm / h, TabGRU achieved a slightly lower RMSE, indicating that current correction strategies for extreme peak values ​​may have a slight overcompensation for moderate rainfall, but the method in this invention still exhibits the best overall performance across all ranges.

[0064] To address the core challenge of extreme peak rainfall scenarios, the experiment further extracted samples with rainfall intensity exceeding 5 mm / h from the test set as a peak subset for specific evaluation. Table 3 presents a performance comparison of each method on the peak subset.

[0065] Table 3. Performance comparison results of the method on the subset of rainfall peaks.

[0066] As shown in Table 3, all baseline methods exhibit significant negative biases on the peak subset, with bias values ​​ranging from -1.8210 mm / h to -2.8514 mm / h, and underestimation rates exceeding 88%. The standard LSTM network shows an underestimation rate as high as 96.5%, indicating that traditional time-series models are almost completely ineffective in extreme rainfall events. In contrast, the method of this invention corrects the bias to a positive bias of +0.5967 mm / h, reducing the underestimation rate to 43.2%, and lowering the RMSE by approximately 23.0% compared to the original hierarchical regression and approximately 37.6% compared to the standard LSTM network. This invention corrects the peak estimation bias from a systematic underestimation of -2.2413 mm / h to a conservative positive bias of +0.5967 mm / h, which has significant practical implications in urban rainstorm warning scenarios, demonstrating that conservative overestimation is more conducive to triggering timely warnings than catastrophic underestimation.

[0067] To verify the independent contribution of each core module in the method of this invention, an ablation study was designed to progressively remove the soft label mechanism, the Mamba temporal residual correction unit, and the DUET adaptive fusion unit. Table 4 presents the complete results of the ablation study.

[0068] Table 4 Ablation Experiment Results

[0069] As shown in Table 4, although the hard-labeled hierarchical regression controlled the false alarm rate to the lowest level of 0.1524, its peak RMSE was as high as 3.2996 mm / h. This indicates that while the overly conservative hard classification boundary can suppress false alarms, it severely sacrifices the ability to capture extreme events. Furthermore, the forced binarization of hard labels disrupts the physical continuity of the dry and wet boundaries, causing misclassification errors in hierarchical regression to propagate and amplify along the cascaded architecture. After adopting the soft-label mechanism, the model allows uncertainty to be retained in the boundary ambiguity region. The global MSE decreased from 0.2613 to 0.2426, and the peak RMSE decreased from 3.2996 to 3.1870, verifying that the soft-label mechanism effectively suppresses the propagation of cascaded errors by modeling boundary uncertainty. Based on this, after introducing the Mamba temporal residual correction unit, the peak RMSE further decreased from 3.1870 to 2.1074, a reduction of 33.8%, verifying that the Mamba model can effectively compensate for the systematic underestimation of hierarchical regression in the peak stage by capturing the temporal momentum of the decaying signal. Finally, by introducing the DUET adaptive fusion unit, the model further optimized the stability of peak estimation while maintaining global accuracy.

[0070] To verify the performance of the method for detecting wet and dry conditions, the experiment used a hardware noise threshold of 0.15 mm / h to binarize the continuous rainfall sequence, and the false alarm rate, detection probability, and critical success index of each method were statistically analyzed. The results show that the method of this invention achieves a detection probability of 0.9847, a false alarm rate of 0.2130, and a critical success index of 0.7774. Compared with the hard-label hierarchical regression scheme (detection probability 0.9701, false alarm rate 0.1524), the method of this invention achieves a higher detection probability and better peak estimation accuracy at the cost of a slightly higher false alarm rate. This trade-off is acceptable in urban rainstorm warning scenarios, as the cost of missing extreme rainfall events is far higher than occasional false alarms.

[0071] Regarding real-time performance, experiments were conducted in a CPU environment to test the single-inference latency of each method. The inference latency of the method described in this invention is 110.51 ms, with 154.66 K model parameters. In comparison, the inference latency of TabGRU is 306.55 ms, with 104.59 K parameters; and the inference latency of the standard LSTM network is 140.57 ms, with 152.40 K parameters. The inference speed of the method described in this invention is approximately 2.7 times faster than TabGRU, and the inference latency is approximately 21.4% lower than that of standard LSTM, while the number of parameters is roughly equivalent to that of LSTM.

[0072] Furthermore, the experiment analyzed the dynamic evolution of the three-branch gating weights in the DUET adaptive fusion unit. Under stable or light rain conditions, the gating weights of the width branch remained high at 0.70-0.80, indicating that the model tends to use simple linear mappings to ensure computational efficiency and generalization performance. When extreme rainfall events occur, the gating weights of the depth branch rapidly increased, and the model automatically switched the computational load to the multi-layer nonlinear transformation branch to handle complex signal abrupt changes. This dynamic routing mechanism verifies that the gating fusion strategy proposed in this invention can adaptively adjust the model's computational mode according to rainfall intensity, maintaining lightweight operation under normal conditions and automatically calling high-capacity computing resources during extreme events.

[0073] Based on the above experimental results, the hierarchical rainfall intensity inversion method based on 6G backhaul network proposed in this invention has achieved better experimental results than existing methods in terms of global estimation accuracy, extreme peak capture capability, dry and wet state detection reliability, and inference efficiency, verifying the effectiveness and practicality of the method in actual deployment scenarios.

[0074] The present invention further provides a graded rainfall intensity inversion system based on a 6G backhaul network, which is used to implement the above-mentioned graded rainfall intensity inversion method based on a 6G backhaul network. The system includes a data acquisition module, a data processing module, and a human-computer interaction module.

[0075] The data acquisition module is used to acquire microwave measurement vectors in the 6G backhaul network.

[0076] The data processing module is used to preprocess the microwave measurement vector to generate the original feature vector, and to process the original feature vector using a pre-trained integrated inversion model to obtain the rainfall intensity inversion result.

[0077] The human-computer interaction module is used to display the rainfall intensity inversion results.

[0078] Those skilled in the art will readily 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. Additionally, 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.

[0079] 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 graded rainfall intensity inversion method based on a 6G backhaul network, characterized in that, include: Obtain microwave measurement vectors from the 6G backhaul network, and perform preprocessing on the microwave measurement vectors to generate original feature vectors; The original feature vectors are input into a pre-trained integrated inversion model for processing to obtain the rainfall intensity inversion result. The integrated inversion model includes a regression model, a gating trigger, a peak processing module, and a decision output module. The regression model includes a state detection branch trained using a soft label mechanism and a hierarchical estimation branch constructed based on a probability-weighted fusion strategy. The methods used by the integrated inversion model to process the original feature vectors include: The state detection branch of the regression model is used to determine the probabilities of multiple rainfall levels based on the original feature vector, and the classification estimation branch of the regression model is used to determine the basic rainfall estimate based on the probabilities of multiple rainfall levels. Using gated triggers, the system determines whether extreme peak rainfall has occurred based on baseline rainfall estimates. When extreme peak rainfall occurs, the peak processing module is used to process the peak portion of the basic rainfall estimate to obtain the optimized estimate process value; The decision output module generates rainfall intensity inversion results based on the optimized estimation process values.

2. The graded rainfall intensity inversion method based on a 6G backhaul network as described in claim 1, characterized in that, The microwave measurement vector includes the attenuation value of the first frequency band, the attenuation value of the second frequency band, the carrier interference noise ratio of the first frequency band, and the carrier interference noise ratio of the second frequency band.

3. The graded rainfall intensity inversion method based on a 6G backhaul network as described in claim 2, characterized in that, Methods for preprocessing microwave measurement vectors to generate original feature vectors include: Based on the microwave measurement vectors within the first time window prior to the current time, calculate the rolling statistical characteristics within the first time window; Calculate the cross features between data of different dimensions in the microwave measurement vector; Based on the microwave measurement vector within the second time window prior to the current moment, the difference feature is obtained by subtracting the measurement values ​​of the same dimension data at the current moment and at historical moments; Feature extraction is performed on the microwave measurement vector at the current moment to obtain the original signal features; Obtain the time-coding information of the current moment to obtain the time-coding features; The rolling statistical features, the original signal features, the cross features, the differential features, and the time-coded features are fused to generate the original feature vector.

4. The graded rainfall intensity inversion method based on a 6G backhaul network as described in claim 1, characterized in that, The soft labeling mechanism is used to generate three-level soft labels based on the comparison results between real rainfall intensity data and a predetermined threshold set. The three-level soft labels include a first label value representing the dry state, a second label value representing the dry-wet transition state, and a third label value representing the wet state. During the training of the regression model, the state detection branch is trained using the three-level soft labels as the supervision target, and the predetermined threshold is dynamically calibrated through an optimization algorithm. The optimization algorithm uses the weighted sum of false alarm rate, detection probability and mean square error as the fitness function to jointly search and optimize the predetermined threshold.

5. The graded rainfall intensity inversion method based on a 6G backhaul network as described in claim 1, characterized in that, The graded estimation branch includes multiple graded regression units, and each graded regression unit corresponds to a rainfall level probability. The method for determining the baseline rainfall estimate based on the multiple rainfall level probabilities includes: The original feature vector is input into multiple hierarchical regression units for processing to obtain multiple rainfall intensity estimation sub-values; The basic rainfall estimate is obtained by weighting and summing the rainfall intensity estimates using the probability of rainfall level as the weight.

6. The graded rainfall intensity inversion method based on a 6G backhaul network as described in claim 1, characterized in that, The method for determining whether extreme peak rainfall has occurred based on a baseline rainfall estimate using a gated trigger includes: Obtain multiple historical baseline rainfall estimates for a predetermined time period prior to the current moment; The time-series variation is obtained by performing a difference operation on the multiple historical baseline rainfall estimates; The cumulative estimate is obtained by summing the multiple historical baseline rainfall estimates. The probability of high-intensity rainfall is obtained by taking the highest probability or the sum of the two highest probabilities among the multiple rainfall level probabilities. The high-intensity rainfall probability, the temporal variation, and the cumulative estimate are input into a pre-trained binary classifier to obtain the trigger flag; When the trigger flag is in the first state, it is determined that extreme peak rainfall has occurred; when the trigger flag is in the second state, it is determined that no extreme peak rainfall has occurred.

7. The graded rainfall intensity inversion method based on a 6G backhaul network as described in claim 1, characterized in that, The peak processing module includes a timing residual correction unit and an adaptive fusion unit; Methods for obtaining optimized estimated process values ​​by processing the peak portion of the basic rainfall estimate using a peak processing module when extreme peak rainfall occurs include: The time-series residual correction unit is used to perform state-space modeling of the historical feature sequence within a predetermined time window before the current time, extract the time-series dependency, and output the residual correction amount. The residual correction is added to the baseline rainfall estimate to obtain the corrected rainfall estimate; The adaptive fusion unit is used to perform multi-branch nonlinear mapping and adaptive fusion on the corrected rainfall estimate to obtain the optimized estimation process value. The historical feature sequence includes the original feature vectors of each historical moment within a predetermined time window prior to the current moment, as well as the basic rainfall estimate corresponding to each historical moment.

8. The graded rainfall intensity inversion method based on a 6G backhaul network as described in claim 7, characterized in that, The temporal residual correction unit is constructed based on the Mamba state-space model; The Mamba state-space model performs causal sequence modeling on the input historical feature sequence, and dynamically adjusts the state transition according to the input content through a selective scanning mechanism to extract long-range temporal dependencies and output residual correction. The adaptive fusion unit includes a depth branch, a width branch, and a residual branch; The depth branch performs a nonlinear transformation on the input through cascaded fully connected layers, the width branch performs a linear transformation on the input through a single fully connected layer, and the residual branch is used to preserve the original information of the input features. The adaptive fusion unit performs weighted fusion of the outputs of the depth branch, the width branch, and the residual branch using learnable gating weights.

9. The graded rainfall intensity inversion method based on a 6G backhaul network as described in claim 1, characterized in that, When the gated trigger determines that no extreme peak rainfall has occurred, the decision output module outputs the basic rainfall estimate as the rainfall intensity inversion result; when the gated trigger determines that extreme peak rainfall has occurred, the decision output module outputs the optimized estimation process value as the rainfall intensity inversion result.

10. A graded rainfall intensity retrieval system based on a 6G backhaul network, characterized in that, The system is used to implement the graded rainfall intensity inversion method based on a 6G backhaul network as described in any one of claims 1-9, the system comprising: The data acquisition module is used to acquire microwave measurement vectors in the 6G backhaul network; The data processing module is used to preprocess the microwave measurement vector to generate the original feature vector, and to process the original feature vector using a pre-trained integrated inversion model to obtain the rainfall intensity inversion result; The human-computer interaction module is used to display the rainfall intensity inversion results.

Images