Wireless sensor network multi-modal feature fusion system and method for flood disasters

Abstract

The invention relates to the technical field of data processing, in particular to a flood disaster-oriented wireless sensor network multi-modal feature fusion system and method, and provides the following scheme: acquiring multi-dimensional sensing data through a wireless sensor network, constructing a multi-modal feature matrix, and inputting a preset intelligent AI model for disaster assessment. The rainfall intensity change trend is dynamically captured through a self-adaptively adjusted time window and a sliding step length in combination with an LSTM prediction model, and the disaster prediction precision is optimized. The problems of multi-source data fusion, poor real-time performance and low precision in a traditional flood disaster early warning system are solved, and efficient and accurate flood disaster prediction and risk assessment are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a multi-modal feature fusion system and method for a wireless sensor network for flood disasters. Background Art

[0002] With the intensification of global climate change, the frequency and intensity of flood disasters have been increasing continuously, causing serious impacts on human society, economy and ecological environment. Traditional flood disaster monitoring methods rely on a single data source (such as satellite remote sensing, ground meteorological stations or hydrological monitoring), but these methods have certain limitations in terms of spatio-temporal resolution, real-time performance and accuracy. For example, although satellite remote sensing can cover a wide area, the frequency of its data acquisition is low, and it is difficult to provide ground detail information; although ground sensors can collect data in real time, their spatial coverage is limited, and they cannot comprehensively reflect the changes in the development process of regional disasters.

[0003] For example, the Chinese patent with the authorization announcement number CN112949414B discloses a method for intelligent mapping of surface water bodies in wide-field high-resolution No. 6 satellite images. On the basis of selecting high-quality wide-field high-resolution No. 6 remote sensing image data, a geometric rectification and image fusion method based on grid division is used to obtain wide-field 2m resolution 8-band high-resolution No. 6 fused image data. It is proposed to carry out multi-scale rapid segmentation of images by combining principal component analysis and object-oriented methods. On the scale of segmented objects, a feature classification vector is constructed, and water body / non-water body training samples with high spatio-temporal representativeness are selected based on spatial grids to construct an object-oriented deep neural network intelligent mapping model for surface water bodies, so as to realize the automatic mapping of surface water bodies in wide-field high-resolution No. 6 remote sensing images. It mainly improves the efficiency and accuracy of geometric rectification, multi-scale segmentation and intelligent extraction of surface water bodies in wide-field images, and has good application potential in flood monitoring, river and lake supervision and water ecological investigation.

[0004] The above patents all have the problems proposed in this background art: there are problems of insufficient accuracy and poor real-time performance in multi-modal data fusion, spatio-temporal alignment and feature extraction, and it is difficult to effectively support dynamic and refined disaster risk assessment and early warning. To solve the above problems, the present application designs a multi-modal feature fusion system and method for a wireless sensor network for flood disasters. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a multi-modal feature fusion system and method for a wireless sensor network for flood disasters in view of the deficiencies of the prior art. Multidimensional sensing data is obtained through a wireless sensor network, a multi-modal feature matrix is constructed, and a preset intelligent AI model is input for disaster assessment. By adaptively adjusting the time window and sliding step size, combined with the LSTM prediction model, the changing trend of precipitation intensity is dynamically captured, and the accuracy of disaster prediction is optimized. The system effectively improves the timeliness and accuracy of data processing through multi-physical field coupling compensation, cross-modal data fusion, and dynamic feature distillation. The second processing network further improves the accuracy of the disaster assessment result through a physically driven non-linear mapping, feature reconstruction, and two-way collaborative optimization mechanism. The invention solves the problems of multi-source data fusion, poor real-time performance, and low accuracy in traditional flood disaster warning systems, and realizes efficient and accurate flood disaster prediction and risk assessment.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A multi-modal feature fusion method for a wireless sensor network for flood disasters, the multi-modal feature fusion method for the wireless sensor network includes:

[0008] Obtain multi-dimensional sensing data of the basin through a time window, and construct a multi-modal feature matrix, wherein the length of the time window is adaptively adjusted according to the precipitation intensity change rate, and the window sliding step size has a non-linear mapping relationship with the basin runoff propagation speed;

[0009] Input the multi-modal feature matrix into a preset intelligent AI model to output a disaster assessment result, wherein the intelligent AI model includes a first processing network and a second processing network, the first processing network extracts features from the multi-modal feature matrix, and the second processing network compensates for the output of the first processing network through a compensation factor in the multi-modal feature matrix for environmental factors;

[0010] Feed back the disaster assessment result to the monitoring platform to generate a dynamic risk map of the basin and send a graded warning signal.

[0011] The length and window sliding step size are calculated based on the LSTM prediction model, specifically including:

[0012] Obtain historical precipitation intensity data, and predict the changing trend of precipitation intensity in the future for a period of time according to the historical precipitation intensity data;

[0013] Calculate the maximum change amplitude of the precipitation intensity according to the precipitation intensity change trend;

[0014] Set the length of the time window according to the maximum change amplitude;

[0015] Input the precipitation intensity change trend into a preset modeling model to output a window sliding step size, where the modeling model is established based on the terrain, precipitation pattern, and runoff propagation characteristics of the basin.

[0016] The constructing of the multi-modal feature matrix includes:

[0017] Map the multi-dimensional sensing data to the three-dimensional geographic coordinate system of the basin to generate spatio-temporally unified grid sensing data, where the multi-dimensional sensing data includes heterogeneous sensing data distributed at water body, soil, and meteorological monitoring nodes;

[0018] Perform multi-physical field coupling compensation on the grid sensing data;

[0019] Perform cross-modal fusion on the compensated sensing data with meteorological radar and satellite remote sensing data to generate a coupling tensor;

[0020] Perform dynamic feature distillation on the coupling tensor to obtain a six-dimensional feature matrix with compressed dimensions, where the dimensions of the six-dimensional feature matrix include timestamp, spatial grid, hydrology, meteorology, geological characteristics, and compensation factor dimensions.

[0021] Performing cross-modal fusion on the compensated sensing data with meteorological radar and satellite remote sensing data includes:

[0022] Construct a hydrological feature sub-tensor;

[0023] Project the meteorological radar reflectivity tensor and the remote sensing deformation tensor to the hydrological feature space of the hydrological feature sub-tensor through a decomposition algorithm to generate a coupling tensor.

[0024] The dynamic feature distillation includes a forward propagation path and a backward propagation path, specifically including:

[0025] In the forward propagation path, use hydrological features, meteorological features, and geological features as query vectors, key vectors, and value vectors respectively to calculate the dynamic correlation degree between different feature channels and generate a weight distribution map;

[0026] Perform Hadamard product operation on the weight distribution map and the coupling tensor to output a preliminary optimized feature subspace;

[0027] In the backward propagation path, construct a feature selector based on gradient sensitivity, screen the feature subspace through the feature selector, and apply a hard truncation to the feature dimensions with gradient magnitudes lower than the threshold to generate a sparsification mask matrix;

[0028] Perform bitwise multiplication on the sparsification mask matrix and the feature sub-control to output a six-dimensional feature matrix.

[0029] The first processing network includes:

[0030] A spatio-temporal feature extraction layer for jointly modeling the spatio-temporal characteristics of the multi-modal feature matrix and outputting a feature mapping matrix, where the spatio-temporal joint modeling includes a spatio-temporal convolution kernel group dynamically adjusted based on timestamps and spatial grids. The spatio-temporal convolution kernel group slides along the time dimension to capture the evolution trend of hydrological parameters and expands along the spatial dimension to model the flood propagation correlation between basin nodes;

[0031] A branch feature extraction layer for processing the feature mapping matrix through a neural network model and outputting branch features;

[0032] A feature output layer for adaptively allocating weights to each feature dimension through an attention mechanism. The attention mechanism includes a channel attention mechanism and a spatial attention mechanism, and outputs the branch features after dynamic channel and spatial weighting through an activation function.

[0033] The branch feature extraction layer includes a hydrological branch, a meteorological branch, and a geological branch, where:

[0034] The hydrological branch extracts multi-scale runoff features through a dilated convolution network;

[0035] The meteorological branch captures the spatio-temporal propagation pattern of precipitation through a bidirectional gated recurrent unit;

[0036] The geological branch models the spatial topological relationship of the soil permeability in the basin through a graph convolution network.

[0037] The second processing network includes:

[0038] A data mapping layer for modeling the dynamic relationship between environmental data and the output of the first processing network through a physically-driven non-linear mapping mechanism to generate an influence weight matrix;

[0039] A feature reconstruction layer for reconstructing each dimension of the output of the first processing network according to the influence weight matrix;

[0040] An output layer for performing consistency correction on the reconstructed features through a bidirectional collaborative optimization mechanism, and non-linearly coupling the features through the domain and spatial domain according to the corrected features to calculate the disaster assessment result.

[0041] A multi-modal feature fusion system for a wireless sensor network for flood disasters, the system includes a data acquisition module, a data processing module, and a result feedback module, where:

[0042] The data acquisition module is used to obtain multi-dimensional sensing data of the basin through a time window, where the length of the time window is adaptively adjusted according to the precipitation intensity change rate, and the window sliding step size has a non-linear mapping relationship with the basin runoff propagation speed;

[0043] The data processing module is used to construct a multi-modal feature matrix according to the output of the data acquisition module, and input the multi-modal feature matrix into a preset intelligent AI model to output a disaster assessment result;

[0044] The result feedback module is used to feedback the output of the data processing module to the monitoring platform to generate a dynamic risk map of the basin and send a graded early warning signal.

[0045] The data processing module includes:

[0046] A matrix generation unit, which is used to map and organize multi-dimensional sensing data according to spatio-temporal relationships into a multi-modal feature matrix;

[0047] A feature extraction unit, configured with a first processing network, to extract features from the multi-modal feature matrix;

[0048] An output compensation unit, configured with a second processing network, to perform environmental factor compensation on the output of the first processing network through a compensation factor in the multi-modal feature matrix.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] By constructing a multi-modal feature matrix, the present invention not only realizes the spatio-temporal mapping of multi-modal heterogeneous sensing data such as hydrology, meteorology, and geology to a unified three-dimensional geographic coordinate system, but also overcomes the problems of scale difference, heterogeneity, and spatio-temporal asynchrony during data fusion through a multi-physical field coupling compensation and cross-modal fusion mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] By reading the detailed description of the non-limiting embodiments with reference to the following drawings, other features, objects, and advantages of the present invention will become more apparent:

[0052] Figure 1 It is a schematic flowchart of a multi-modal feature fusion method for a wireless sensor network for flood disasters in Embodiment 1 of the present invention;

[0053] Figure 2 It is a network structure diagram of the intelligent AI model in Embodiment 1 of the present invention;

[0054] Figure 3 It is a structure diagram of the first processing network in Embodiment 1 of the present invention;

[0055] Figure 4 It is a structure diagram of the branch feature extraction layer in Embodiment 1 of the present invention;

[0056] Figure 5 It is a structure diagram of the second processing network in Embodiment 1 of the present invention;

[0057] Figure 6This is the module diagram of the multi-modal feature fusion system of the wireless sensor network for flood disasters in Embodiment 2 of the present invention. Detailed implementation manners

[0058] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0059] Embodiment 1:

[0060] Please refer to Figure 1 , an embodiment provided by the present invention: a multi-modal feature fusion method for a wireless sensor network for flood disasters, and the specific steps of the method are as follows:

[0061] S1: Obtain multi-dimensional sensing data of the basin through a time window;

[0062] In this embodiment, first, multi-dimensional sensing data of different monitoring nodes (such as soil humidity, precipitation intensity, water level, wind speed, etc.) in the basin is obtained through a wireless sensor network. Through the time window mechanism, the data within each period of time is aggregated together for subsequent processing. Specifically, the length of the time window is adaptively adjusted according to the change rate of precipitation intensity, and the data acquisition time is dynamically optimized according to the trend of precipitation change to ensure that the critical moments of flood disasters are captured. At the same time, according to the topographic and hydrological characteristics of the basin, the data update frequency is precisely adjusted. Precise disaster prediction is achieved, and the timeliness and relevance of the data are improved.

[0063] S2: Construct a multi-modal feature matrix;

[0064] In this embodiment, the obtained multi-dimensional sensing data is pre-processed (such as spatio-temporal alignment, standardization, etc.), and then a multi-modal feature matrix is constructed according to the spatio-temporal relationship between the data. Specifically, the data is first mapped to the three-dimensional geographical coordinate system of the basin and subjected to spatio-temporal unified grid processing. These gridded sensing data will undergo multi-physical field coupling compensation to uniformly fuse the data collected by different sensors (such as soil humidity, precipitation, meteorological data, etc.). Then, cross-modal data fusion is performed to combine the compensated sensing data with meteorological radar and satellite remote sensing data to generate a coupling tensor. Finally, through dynamic feature distillation processing, a six-dimensional feature matrix is obtained, covering dimensions of timestamp, spatial grid, hydrology, meteorology, geological characteristics, and compensation factors.

[0065] S3: Input the multi-modal feature matrix into a preset intelligent AI model to output a disaster assessment result;

[0066] In this embodiment, the constructed multi-modal feature matrix is input into a preset intelligent AI model. Please refer to Figure 2, The network structure diagram of the intelligent AI model in the embodiment of the present invention. The model includes a first processing network and a second processing network. The first processing network extracts spatio-temporal features from the feature matrix through a neural network to capture the key features of the disaster occurrence. The second processing network compensates the output of the first processing network based on the compensation factors in the multi-modal feature matrix, and realizes more refined disaster assessment by considering the specific meteorological, geological and hydrological environments of the basin.

[0067] S4: Feed back the disaster assessment result to the monitoring platform to generate a dynamic risk map of the basin and send a graded early warning signal;

[0068] In this embodiment, the disaster assessment result output by the intelligent AI model will be fed back to the monitoring platform. The platform uses these results to generate a dynamic risk map of the basin and send corresponding early warning signals according to different disaster risk levels. The risk map can dynamically display the prediction results of the disaster occurrence, including the possible impact range and time scale, to help relevant decision-makers take emergency measures in a timely manner.

[0069] In this embodiment, the disaster assessment result output by the intelligent AI model will be fed back to the monitoring platform. The platform uses these results to generate a dynamic risk map of the basin and send corresponding early warning signals according to different disaster risk levels. The risk map can dynamically display the prediction results of the disaster occurrence, including the possible impact range and time scale, to help relevant decision-makers take emergency measures in a timely manner.

[0070] In the field of actual flood disaster monitoring, although the existing technology can obtain local hydrological, meteorological or geological data respectively through a single type of sensor (such as a hydrological sensor, a weather station or a remote sensing satellite, etc.), due to the spatio-temporal heterogeneity of the data and the differences in monitoring methods, it is difficult for the existing technology to achieve a refined, dynamic and global description of the disaster occurrence and development process. For example, when the traditional technology uses a wireless sensor network to obtain soil moisture and water level data at different monitoring points in the basin, due to the lack of a unified mapping and fusion of the spatio-temporal relationship between heterogeneous data, it is difficult to determine the accurate correlation between the local disaster characteristics at a specific moment and the flood evolution process of the entire basin; further, when independent technologies such as meteorological radar or satellite remote sensing provide large-scale regional data, due to insufficient spatial resolution or timeliness, it is also difficult for the traditional technology to accurately project this large-scale information into the data space of micro-scale monitoring nodes. This gap between data scales and patterns greatly limits the accuracy and real-time performance of the existing disaster warning technology.

[0071] In this embodiment, by constructing a multimodal feature matrix, not only the spatiotemporal mapping of multimodal heterogeneous sensor data such as hydrology, meteorology, and geology to a unified three-dimensional geographic coordinate system is achieved, but also the scale difference, heterogeneity, and spatiotemporal asynchrony problems in data fusion in the prior art are overcome through multi-physical field coupling compensation and cross-modal fusion mechanism.

[0072] Specifically, the method proposed in this application first maps the data uniformly to three-dimensional spatial coordinates through gridding, and then uses cross-modal tensor fusion technology to efficiently fuse large-scale monitoring data from meteorological radar and satellite remote sensing with local node data to form a coupled tensor with precise fusion scale and multi-dimensional information. Furthermore, through feature distillation technology, the six-dimensional feature matrix formed efficiently retains the spatiotemporal characteristics of key data, thereby significantly improving the accuracy of disaster prediction.

[0073] Furthermore, through the forward propagation and back propagation paths in dynamic feature distillation, adaptive optimization of weight distribution between multi-source data is achieved. This method can spontaneously mine and strengthen the implicit dynamic correlation between data, avoiding the errors caused by artificially setting data weights in traditional methods. First, in the forward propagation path, a weight distribution map is constructed through the dynamic correlation of hydrological, meteorological, and geological features; then, in the back propagation path, feature selection is performed based on gradient sensitivity, and data channels with weak correlation are dynamically sparsely processed, creatively reducing the interference caused by redundant data, thereby ensuring that the final feature matrix only retains effective features that have actual contributions to disaster prediction, so that the effectiveness of data feature fusion reaches the best state.

[0074] In this embodiment, the length and window sliding step are calculated based on the LSTM prediction model, and the specific steps are as follows:

[0075] S1.1: Obtain historical precipitation intensity data, and predict the precipitation intensity change trend in the future based on the historical precipitation intensity data;

[0076] In this embodiment, first, historical precipitation intensity data is obtained through a wireless sensor network or meteorological monitoring equipment. The data may include precipitation intensity information for the past several days or hours, and after spatio-temporal alignment, these data are used to train the LSTM model. The LSTM model performs excellently in time series data modeling and can capture long-term dependencies and short-term fluctuations in precipitation intensity data. To improve the prediction accuracy, the historical precipitation intensity data is normalized by the sliding window method, outliers are removed, and weighted according to factors such as the geographical characteristics of the basin and seasonal changes. After training the LSTM model, the changing trend of precipitation intensity in the future for a period of time can be accurately predicted. Through the output of the LSTM model, the changing trajectory of precipitation intensity in the future for a period of time can be obtained, and information such as future possible precipitation extremes and the changing range of precipitation intensity can be predicted, providing a basis for subsequent calculation of the time window length and sliding step size.

[0077] S1.2: Calculate the maximum changing range of the precipitation intensity according to the changing trend of the precipitation intensity;

[0078] Specifically, the changing range of precipitation intensity in the future time period is quantified to obtain the maximum value of the changing range. This maximum changing range reflects the range in which the precipitation intensity may change drastically, and it can provide a more accurate basis for setting the time window. For example, if the predicted change in precipitation intensity is relatively drastic, a longer time window needs to be set to ensure that the entire process of the change can be fully captured. On the contrary, if the change in precipitation intensity is small, a shorter time window can be selected. By calculating the maximum changing range, the problem of the time window being too long or too short can be avoided, ensuring that the length of the time window is more in line with the actual meteorological conditions.

[0079] S1.3: Set the length of the time window according to the maximum changing range;

[0080] Specifically, the maximum changing range is used as a key parameter for the length of the time window to ensure that the time window can include all possible changes in precipitation intensity.

[0081] In this embodiment, the initial time window is weighted by the maximum changing range to provide a suitable time scale for flood disaster monitoring. For example, when the change in precipitation intensity is large, the time window will be extended to fully capture the large fluctuations in precipitation intensity; while when the change in precipitation is small, the window length will be shortened to reduce the interference of irrelevant data. The specific weighting method can be determined by those skilled in the art through a large number of repeated experiments.

[0082] S1.4: Input the changing trend of the precipitation intensity into a preset modeling model to output the window sliding step size, where the modeling model is established based on the terrain, precipitation pattern, and runoff propagation characteristics of the basin;

[0083] Specifically, based on the topographic features of the basin (such as slope, river distribution, vegetation cover, etc.) and the spatio-temporal patterns of precipitation, the modeling model simulates how precipitation intensity propagates within the basin under different sliding step lengths. The runoff propagation characteristics of the basin, the spatial distribution of precipitation, and the drainage capacity of different regions will all affect the setting of the sliding step length. By simulating the actual situations under different step lengths, the modeling model can output the most suitable window sliding step length, ensuring the continuity and real-time nature of the monitoring data and adapting to different geographical conditions and climate characteristics.

[0084] The specific steps of S2 are as follows:

[0085] S2.1: Map the multi-dimensional sensing data to the three-dimensional geographical coordinate system of the basin to generate spatio-temporally unified grid sensing data, where the multi-dimensional sensing data includes heterogeneous sensing data distributed at water body, soil, and meteorological monitoring nodes;

[0086] Specifically, the spatio-temporal distribution of the sensing data is heterogeneous, and due to different monitoring devices, the data obtained may involve different spatial scales and time granularities. Therefore, to ensure that various types of data can be fused and processed within a unified framework, these data must be converted into point data in the three-dimensional coordinate system through spatial mapping. For example, a soil moisture sensor may be located at a certain location in the basin, while a weather station may be located at another location, and a water level sensor is at yet another position. By using GIS (Geographic Information System) technology, these heterogeneous data are uniformly mapped to the three-dimensional geographical coordinate system of the basin, ensuring that the data collected by different sensors can be processed within the same spatio-temporal framework. This mapping not only ensures the spatial consistency of the data but also the spatio-temporal unity in the subsequent fusion process.

[0087] S2.2: Perform multi-physical-field coupling compensation on the grid sensing data;

[0088] Specifically, the grid sensing data after gridification, especially hydrological, meteorological, and geological data, are usually affected by multiple physical fields. Natural factors such as meteorological changes, geological characteristics, and basin topography affect the data, which may cause systematic deviations in the data. For example, within the same basin, due to local climate changes or topographic differences, the hydrological data of different monitoring nodes may vary. A water level sensor located upstream of a certain highland is affected by precipitation differently from a sensor located in a low-lying area. In addition, the change in soil moisture may be affected by factors such as seasonal climate, groundwater level, and soil type. To eliminate these effects and improve the accuracy of the data, multi-physical-field coupling compensation must be carried out.

[0089] In this embodiment, first, a multi-physical field mathematical model is established by using the geographical characteristics, meteorological data, and hydrological characteristics of the basin. The model includes a meteorological field (such as precipitation, temperature, humidity, etc.), a geological field (such as soil type, stratigraphic structure, groundwater level, etc.), and a hydrological field (such as water level changes, flow velocity, etc.). Through Geographic Information System (GIS) and remote sensing technology, the spatial distribution data of these physical fields are collected.

[0090] Furthermore, after obtaining the basic data of each physical field, the gridded sensing data is preliminarily analyzed to identify the error sources. For example, by comparing the hydrological sensing data in different geographical regions, it can be found that some data have systematic errors due to differences in climate factors or geological conditions.

[0091] Furthermore, after identifying the errors by comparing the results obtained from physical field modeling with the sensing data, the data is compensated using numerical optimization methods. For example, if there are deviations in the hydrological sensor data, the compensation method will combine the terrain and meteorological conditions of the basin to adjust the data to a more realistic range. This compensation process takes into account the interactive effects between physical fields, such as the accumulation and drainage of precipitation by the terrain, and the infiltration characteristics of the soil for precipitation.

[0092] Furthermore, by minimizing the difference between the actual sensor data and the predicted data of the physical field model, the optimal compensation value is found. For hydrological data, the error between the water level sensor readings and the water level at the corresponding position in the model can be calculated by the least squares method and then adjusted.

[0093] Preferably, the compensated data is updated in real time and dynamically adjusted through a feedback mechanism. For example, in practical applications, meteorological changes are dynamic, so the compensation model will continuously adjust the compensation strategy according to the latest meteorological data and real-time monitoring results. This dynamic compensation method can ensure that data errors in long-term monitoring are corrected in real time.

[0094] S2.3: Construct a hydrological feature sub-tensor, and project the meteorological radar reflectivity tensor and the remote sensing deformation tensor onto the hydrological feature space of the hydrological feature sub-tensor through a decomposition algorithm to generate a coupled tensor;

[0095] Specifically, the key to constructing the hydrological feature sub-tensor lies in extracting the sub-space related to hydrological features from multi-source data. Hydrological features usually include data related to hydrological dynamics such as precipitation intensity, flow rate, water level changes, soil moisture, etc. To extract these hydrological features from sensor data, first, key feature values representing the hydrological changes in the basin need to be extracted from existing hydrological data (such as water level sensors, precipitation intensity sensors, etc.). Then, data dimensionality reduction methods are used to reduce the dimensionality of the data and retain the features that can best reflect hydrological changes. By projecting the original multi-dimensional data into a lower-dimensional space, a hydrological feature sub-tensor containing the main hydrological features is generated.

[0096] Furthermore, the hydrological feature sub-tensor is fused with the meteorological radar reflectivity tensor and the remote sensing deformation tensor through non-negative matrix factorization. Through this decomposition algorithm, the meteorological radar reflectivity tensor and the remote sensing deformation tensor are decomposed into several factors in multiple dimensions such as time and space. These factors represent the potential structures and relationships between various dimensions. The goal of the decomposition is to represent each tensor as the product of several low-rank matrices, so that the low-dimensional representation that can best reflect the relationships between various types of data can be extracted. For example, the meteorological radar reflectivity tensor can be decomposed into a set of factors related to spatial features, and these factors represent the meteorological reflection characteristics of different geographical regions; the remote sensing deformation tensor can extract factors representing surface deformation during different time periods through decomposition. The extraction process of these factors ensures that the internal structure of each type of data is fully reflected.

[0097] Furthermore, the projection process is actually to map the meteorological radar and remote sensing data in the hydrological feature sub-space, converting their spatio-temporal features into a feature space representation closely related to hydrological changes. To achieve the projection, first, the similarity between the meteorological radar reflectivity tensor and the remote sensing deformation tensor and the hydrological feature sub-tensor needs to be calculated. The calculation of similarity can be achieved by calculating the inner product, correlation, or other measurement criteria. Then, using these calculated similarities, the meteorological radar and remote sensing deformation data are mapped into the hydrological feature sub-space through linear or non-linear mapping. The core purpose of the projection process is to fuse the large-scale meteorological data and remote sensing data with the hydrological data within the basin in a common space, so as to effectively reflect the correlation between different data sources and generate a unified coupling tensor. The generated coupling tensor will simultaneously contain multi-dimensional information such as hydrology, meteorology, and remote sensing, and can comprehensively and accurately describe the disaster process within the basin.

[0098] S2.4: Perform dynamic feature distillation on the coupling tensor to obtain a six-dimensional feature matrix with compressed dimensions, where the dimensions of the six-dimensional feature matrix include timestamp, spatial grid, hydrology, meteorology, geological characteristics, and compensation factor dimensions;

[0099] Specifically, the coupled tensor after cross-modal fusion usually contains data in multiple dimensions, which may have high dimensions and redundant information, resulting in high computational and storage costs. To effectively reduce the computational complexity and retain useful information, a dynamic feature distillation technique is used to compress the dimensions of the tensor. Through the forward propagation and backward propagation processes, dynamic feature distillation automatically identifies the feature dimensions that have the most impact on disaster prediction and removes redundant or irrelevant information, thus compressing the tensor into a six-dimensional feature matrix.

[0100] The dynamic feature distillation includes a forward propagation path and a backward propagation path. The specific steps of S2.4 are as follows:

[0101] S2.4.1: In the forward propagation path, hydrological features, meteorological features, and geological features are used as query vectors, key vectors, and value vectors respectively to calculate the dynamic correlation degree between different feature channels and generate a weight distribution map.

[0102] Specifically, first, hydrological, meteorological, and geological features are input into a self-attention mechanism model as query vectors (Query), key vectors (Key), and value vectors (Value) respectively. Each feature corresponds to a different feature channel. The feature vectors of each feature channel are mapped into a high-dimensional vector space. These high-dimensional vectors represent the abstract representations of each feature and retain their potential contributions to the final decision-making in disaster prediction through the mapping. The self-attention mechanism learns the dynamic dependence relationships between different feature channels by calculating the correlation degrees between these feature channels.

[0103] Furthermore, the query vector, key vector, and value vector are extracted from different feature spaces respectively. First, the similarity between each pair of query vectors and key vectors is calculated through the inner product to obtain the correlation scores between each pair of features. These similarity metrics represent the degree of correlation between different features and can reflect their influence on the same disaster warning target. For example, certain meteorological features may have a high similarity with hydrological features, indicating that their impacts on flood disasters are closely related; while the similarity between other geological features and meteorological features may be low, indicating that their prediction impacts on this disaster are small.

[0104] Furthermore, based on the calculated similarities, the model uses these similarities to generate a weight distribution map. The weight distribution map is a matrix where the value at each position represents the association strength between a specific feature and other features. These weight values can be regarded as the relative importance of features in disaster prediction. For example, if the similarity between hydrological features and meteorological features is high, the weight between them will also be large, indicating that they have a stronger influence on disaster prediction; while relatively weak features (such as the low correlation between certain geological features and hydrological features) will be assigned smaller weights.

[0105] S2.4.2: Perform a Hadamard product operation on the weight distribution map and the coupling tensor, and output a preliminarily optimized feature subspace;

[0106] Specifically, the obtained weight distribution map will be subjected to a Hadamard product operation with the coupling tensor, that is, perform element-wise multiplication on the weights of each feature channel and the elements of the corresponding coupling tensor. Weight each dimension of the coupling tensor to optimize the representation of the feature space. Through the Hadamard product operation, the dynamic weights of the feature channels can be effectively applied to the coupling tensor, strengthening the features most relevant to disaster prediction and weakening other less relevant features. The Hadamard product is an element-wise multiplication operation that can ensure that the contribution degree of each feature matches its importance in disaster prediction, so that the finally obtained feature subspace better meets the needs of actual disaster prediction.

[0107] S2.4.3: In the backpropagation path, construct a feature selector based on gradient sensitivity, screen the feature subspace through the feature selector, apply a hard truncation to the feature dimensions with a gradient magnitude lower than the specified value, and generate a sparsity mask matrix;

[0108] Specifically, the feature selector measures the importance of each feature by calculating the change rate of each feature during the gradient descent process. If a certain feature contributes less during the entire training process, resulting in a lower gradient magnitude, then these features are hard-truncated through the feature selector and removed from the feature space. This operation generates a sparsity mask matrix, indicating which feature dimensions are not important in disaster prediction and can be discarded. This process is similar to feature selection, but here it is screened based on gradient sensitivity to ensure that features with less contribution in disaster prediction do not affect the learning effect of the model.

[0109] S2.4.4: Perform a bitwise multiplication on the sparsity mask matrix and the feature subspace control to output a six-dimensional feature matrix;

[0110] Specifically, through bitwise multiplication, the sparsification mask matrix sets the feature dimensions that were previously hard-truncated to zero, thus finally outputting a refined six-dimensional feature matrix, ensuring that only those features highly relevant to disaster prediction are retained in the feature matrix, thereby reducing the computational and storage burdens and enhancing the model's efficiency in processing large-scale data.

[0111] Please refer to Figure 3 , the first processing network structure diagram of the embodiment of the present invention. The design of the first processing network is to achieve efficient processing of the multimodal feature matrix, especially by constructing a spatio-temporal feature extraction layer, a branch feature extraction layer, and a feature output layer, to improve the accuracy and timeliness of flood disaster prediction. The first processing network includes:

[0112] A spatio-temporal feature extraction layer for performing spatio-temporal joint modeling on the multimodal feature matrix and outputting a feature mapping matrix, where the spatio-temporal joint modeling includes a spatio-temporal convolution kernel group dynamically adjusted based on timestamps and spatial grids. The spatio-temporal convolution kernel group slides along the time dimension to capture the evolution trend of hydrological parameters and expands along the spatial dimension to model the flood propagation correlation between basin nodes;

[0113] Specifically, the spatio-temporal convolution kernel group is a key component of the spatio-temporal feature extraction layer. To achieve spatio-temporal joint modeling, the spatio-temporal convolution kernel not only slides along the time dimension to capture the evolution trend of hydrological parameters but also expands along the spatial dimension to model the flood propagation correlation between basin nodes.

[0114] In this embodiment, the spatio-temporal convolution kernel group consists of multiple convolution kernels, and each convolution kernel has two dimensions, time and space, respectively responsible for capturing temporal features and spatial features.

[0115] Furthermore, along the time dimension, the convolution kernel slides along the time axis to extract the evolution trend of hydrological parameters over time. To capture hydrological changes over a long time span, the size and stride of the convolution kernel can be dynamically adjusted according to the actual data. For example, hydrological data may have a long change cycle (such as seasonal precipitation changes), so the sliding stride of the convolution kernel can be set to a longer time span to better reflect the long-term change trend. In this way, the convolution kernel can automatically learn the temporal pattern of hydrological data, providing a basis for predicting the occurrence of floods.

[0116] Furthermore, in the spatial dimension, the convolutional kernel slides along the spatial grid to establish the flood propagation correlation between monitoring nodes in the basin. The mutual influence and propagation relationship between each monitoring node (such as water level, soil moisture, meteorological data, etc.) in the basin need to be modeled by the spatial convolutional kernel. For example, precipitation in a certain area may affect the soil moisture or water level changes in neighboring areas, and the mutual relationship between these data is captured by the spatial convolutional kernel. To accurately model the spatial propagation pattern in the basin, the size and shape of the convolutional kernel can be adjusted according to the geographical characteristics of the basin. Through spatial convolution, the model can learn the dependence relationships between different spatial positions, so as to better predict the flood propagation path.

[0117] Furthermore, after the spatio-temporal convolutional kernel group performs a convolutional operation on the input multi-modal feature matrix, it outputs a feature map matrix. This matrix is the final output of the spatio-temporal feature extraction layer and contains the temporal and spatial information in the multi-modal data. Each element in the feature map matrix represents the data feature at a specific time and spatial position, such as precipitation, water level, or soil moisture at a certain time point and spatial position.

[0118] The branch feature extraction layer is used to process the feature map matrix through a neural network model and output branch features;

[0119] In this embodiment, the input feature map matrix will be further processed through a neural network model to extract representative branch features. This branch feature extraction layer usually includes multiple neural network modules, and each module focuses on mining different types of useful information from the feature matrix.

[0120] The feature output layer is used to adaptively allocate the weights of each feature dimension through an attention mechanism. The attention mechanism includes a channel attention mechanism and a spatial attention mechanism, and outputs the branch features after dynamic channel and spatial weighting through an activation function;

[0121] Specifically, the channel attention mechanism is mainly used to adjust the weights of different feature channels. Through this mechanism, the model can adaptively assign different weights to different features, so as to highlight those features that have a greater impact on disaster prediction. For example, hydrological features (such as water level, precipitation, etc.) are usually more important than meteorological or geological features in flood disaster prediction, so the system will assign higher weights to hydrological features according to the actual situation of the data.

[0122] Furthermore, the spatial attention mechanism is used to adjust weights according to the importance of different geographical regions. For example, in some regions, relatively intense precipitation or water level changes may occur. At this time, the features of this region should receive higher attention, while the features of other regions can appropriately reduce their weights. The spatial attention mechanism can adaptively adjust the importance of features according to spatial changes, ensuring that the system can focus on the data in key regions.

[0123] Furthermore, after dynamic channel and spatial weighting, all features will undergo a non-linear mapping through an activation function to generate the final output features. These output features will be used as the input for subsequent disaster assessment and prediction models, ensuring that the system can issue timely disaster warnings based on high precision.

[0124] Please refer to Figure 4 , the structural diagram of the branch feature extraction layer in the embodiment of the present invention. The design of the branch feature extraction layer aims to efficiently process the input multi-modal feature matrix through multiple feature branches and extract key hydrological, meteorological, and geological features. Each branch focuses on extracting features in different fields, thereby providing rich and multi-dimensional information for subsequent disaster prediction. Specifically, the hydrological branch, meteorological branch, and geological branch respectively adopt three different network architectures, namely the Dilated Convolutional Network, Bidirectional Gated Recurrent Unit (BiGRU), and Graph Convolutional Network (GCN), to process the features in their respective fields. The branch feature extraction layer includes a hydrological branch, a meteorological branch, and a geological branch, where:

[0125] The hydrological branch extracts multi-scale runoff features through a dilated convolutional network;

[0126] Specifically, the main task of the hydrological branch is to extract multi-scale runoff features from the input hydrological data. Runoff features refer to the trends of hydrological data such as precipitation, flow rate, and water level changing over time and space. To capture these changes, the hydrological branch adopts a dilated convolutional network. Dilated convolution is different from standard convolution. It can skip some input values during the convolution operation, thereby expanding the receptive field without increasing the number of parameters. In this way, dilated convolution can capture the long-term dependencies in the data without increasing the computational complexity. Process features at different scales. For example, in hydrological data, runoff features may change rapidly in a short period (such as sudden heavy precipitation) or slowly evolve over a long period (such as seasonal changes). Through dilated convolution, the network can automatically capture these multi-scale change features and provide more detailed hydrological information for subsequent flood disaster prediction.

[0127] The meteorological branch captures the spatio-temporal propagation pattern of precipitation through a bidirectional gated recurrent unit;

[0128] Specifically, the main task of the meteorological branch is to capture the spatio-temporal propagation pattern in precipitation data. To this end, the meteorological branch uses a bidirectional gated recurrent unit to model the spatio-temporal evolution of precipitation intensity. BiGRU is a variant of the gated recurrent unit that can consider both the forward and backward dependencies of sequential data. In meteorological data, the spatio-temporal pattern of precipitation is affected not only by past precipitation conditions but also by the reverse influence of future precipitation patterns. BiGRU can effectively capture these spatio-temporal dependencies and provide more comprehensive prediction information.

[0129] Furthermore, the input at each time step includes not only the precipitation data at the current time step but also the forward and backward historical information, which enables the network to more accurately predict future precipitation trends. Especially in flood disaster prediction, the temporal relationship of precipitation amounts is crucial, and the use of bidirectional GRU enables the system to better capture the temporal fluctuations and possible anomalies of precipitation.

[0130] The geological branch models the spatial topological relationship of the soil permeability in the basin through a graph convolutional network;

[0131] Specifically, the main task of the geological branch is to extract features related to flood disasters from geological data, especially soil permeability and the spatial topological structure of the basin. To effectively process this spatial data, the geological branch adopts a graph convolutional network. It can handle the relationships between nodes and edges in graph data, and the soil permeability data of the basin can be naturally represented as a graph structure, where different regions in the basin are the nodes of the graph, and the connections between the nodes represent the mutual relationships between different regions (e.g., the propagation path of water flow). GCN can model the influence of the soil permeability within the basin on water infiltration and flood propagation by learning the connection relationships between these nodes.

[0132] Furthermore, the features of each node (such as soil permeability) are updated according to the information of adjacent nodes, so as to capture the influence of the spatial topological relationship on flood propagation. GCN can effectively process complex basin geological structures and help predict the water flow dynamics in different regions.

[0133] Please refer to Figure 5 , the second processing network structure diagram of the embodiment of the present invention. The role of the second processing network is to further optimize the output features of the first processing network, combine environmental data for modeling and adjustment, and finally output accurate disaster assessment results. It adopts technologies such as a physically driven non-linear mapping mechanism, dimension-by-dimension reconstruction, and a bidirectional collaborative optimization mechanism. The second processing network includes:

[0134] A data mapping layer, which is used to model the dynamic relationship between environmental data and the output of the first processing network through a physically-driven non-linear mapping mechanism, and generate an influence weight matrix;

[0135] In this embodiment, the main task of the data mapping layer is to model the dynamic relationship between the influence from environmental data and the output of the first processing network through a physically-driven non-linear mapping mechanism, and generate an influence weight matrix. First, environmental data such as meteorological data (wind speed, temperature, precipitation, etc.), hydrological data (flow rate, water level change, etc.) and geological data (soil permeability, terrain, etc.) are multi-dimensional data collected by sensors. There is a certain spatio-temporal correlation between these environmental data and the output features of the first processing network. Therefore, through the physically-driven non-linear mapping mechanism, the influence of these environmental factors can be effectively mapped to the feature space of the output of the first processing network.

[0136] Specifically, the physically-driven non-linear mapping mechanism takes into account the interaction of multiple physical factors and uses non-linear functions to capture the non-linear influence of environmental changes on the disaster assessment results. The purpose of this process is to combine the environmental data with the spatio-temporal features of the network output to generate an influence weight matrix, which reflects the importance of each feature under different environmental factors. In this way, the model can more accurately identify which environmental factors have a greater impact on the occurrence of disasters, and make the prediction of the model more accurate through weight adjustment.

[0137] A feature reconstruction layer, which is used to perform per-dimensional reconstruction on the output of the first processing network according to the influence weight matrix;

[0138] In this embodiment, the task of the feature reconstruction layer is to perform per-dimensional reconstruction on the output of the first processing network according to the generated influence weight matrix. The influence weight matrix provides the importance of each feature dimension under different environmental conditions, and the feature reconstruction layer uses this weight information to adjust the feature matrix. Specifically, feature reconstruction is achieved by weighting the output features of the first processing network to ensure that important features obtain higher weights in disaster assessment, while secondary features are appropriately weakened. This process is actually to dynamically adjust each feature dimension to optimize the model's attention to key features.

[0139] Specifically, during the feature reconstruction process, the feature matrix is weighted per dimension according to the influence weight matrix to enhance the influence of key features in disaster prediction. For example, if the precipitation in a certain area has a greater impact on the occurrence of floods, then during reconstruction, the corresponding feature dimension in this area will be assigned a higher weight, thus playing a greater role in subsequent disaster assessment. Through this per-dimensional reconstruction, the feature reconstruction layer can effectively optimize the input features, ensure that the model can fully consider environmental impacts, and generate more accurate disaster prediction results.

[0140] An output layer, which is used to perform consistency correction on the reconstructed features through a two-way collaborative optimization mechanism, and based on the corrected features, non-linearly couple the features through the domain and spatial domains to calculate the disaster assessment result;

[0141] In this embodiment, the main function of the output layer is to perform consistency correction on the reconstructed features through a two-way collaborative optimization mechanism and finally generate a disaster assessment result. The core idea of the two-way collaborative optimization mechanism is to optimize the prediction results of the model in two directions: on the one hand, optimize the accuracy of the output features through forward propagation; on the other hand, use backpropagation to perform consistency correction on the output features to ensure a high degree of consistency between the prediction results and the actual environmental data.

[0142] Specifically, first correct the reconstructed features to ensure the consistency of all features in disaster assessment. This correction not only considers the mutual dependence between features but also combines the impact of environmental data on the occurrence of disasters to ensure the accuracy and reliability of the final result. The corrected features will be non-linearly coupled through the domain and spatial domains to further refine the spatial distribution characteristics of disaster prediction. Specifically, domain coupling considers the logical relationship between features, while spatial domain coupling considers the mutual influence in terms of geographical location. These two coupling methods can help the model accurately predict the occurrence, development, and impact range of disasters in a multi-dimensional and multi-scale space.

[0143] Embodiment 2:

[0144] Please refer to Figure 6 , the present invention provides an embodiment: a multi-modal feature fusion system for a wireless sensor network facing flood disasters, the system includes a data acquisition module, a data processing module, and a result feedback module, wherein:

[0145] The data acquisition module is used to obtain multi-dimensional sensing data of the basin through a time window, wherein the length of the time window is adaptively adjusted according to the precipitation intensity change rate, and the window sliding step size has a non-linear mapping relationship with the basin runoff propagation speed;

[0146] The data processing module is used to construct a multi-modal feature matrix according to the output of the data acquisition module and input the multi-modal feature matrix into a preset intelligent AI model to output a disaster assessment result;

[0147] The result feedback module is used to feedback the output of the data processing module to the monitoring platform to generate a dynamic risk map of the basin and send a graded warning signal.

[0148] The data processing module includes:

[0149] A matrix generation unit, configured to map and organize multi-dimensional sensing data into a multi-modal feature matrix according to spatio-temporal relationships;

[0150] A feature extraction unit, configured with a first processing network, to perform feature extraction on the multi-modal feature matrix;

[0151] An output compensation unit, configured with a second processing network, to perform environmental factor compensation on the output of the first processing network through a compensation factor in the multi-modal feature matrix.

[0152] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A multi-modal feature fusion method for wireless sensor networks facing flood disasters, characterized in that, The multi-modal feature fusion method for the wireless sensor network includes: Obtaining multi-dimensional sensing data of the watershed through a time window, and constructing a multi-modal feature matrix, where the length of the time window is adaptively adjusted according to the precipitation intensity change rate, and the window sliding step size has a non-linear mapping relationship with the watershed runoff propagation speed; Inputting the multi-modal feature matrix into a preset intelligent AI model to output a disaster assessment result, where the intelligent AI model includes a first processing network and a second processing network. The first processing network extracts features from the multi-modal feature matrix, and the second processing network compensates for environmental factors in the output of the first processing network through the compensation factor in the multi-modal feature matrix; Feeding back the disaster assessment result to the monitoring platform to generate a dynamic risk map of the watershed and send a graded warning signal.

2. The multimodal feature fusion method for a wireless sensor network facing flood disasters according to claim 1, wherein The length and window sliding step size are calculated based on the LSTM prediction model, specifically including: Obtaining historical precipitation intensity data and predicting the precipitation intensity change trend in a future period according to the historical precipitation intensity data; Calculating the maximum change amplitude of the precipitation intensity according to the precipitation intensity change trend; Setting the length of the time window according to the maximum change amplitude; Inputting the precipitation intensity change trend into a preset modeling model to output the window sliding step size, where the modeling model is established based on the terrain, precipitation pattern and runoff propagation characteristics of the watershed.

3. The multi-modal feature fusion method of the wireless sensor network for flood disasters according to claim 1, wherein The construction of the multi-modal feature matrix includes: Mapping the multi-dimensional sensing data to the three-dimensional geographic coordinate system of the watershed to generate spatio-temporally unified grid sensing data, where the multi-dimensional sensing data includes heterogeneous sensing data distributed at water body, soil and meteorological monitoring nodes; Performing multi-physical field coupling compensation on the grid sensing data; Performing cross-modal fusion on the compensated sensing data with meteorological radar and satellite remote sensing data to generate a coupling tensor; Performing dynamic feature distillation on the coupling tensor to obtain a six-dimensional feature matrix with compressed dimensions, where the dimensions of the six-dimensional feature matrix include timestamp, spatial grid, hydrology, meteorology, geological characteristics and compensation factor dimensions.

4. The multi-modal feature fusion method for a wireless sensor network for flood disasters according to claim 3, wherein Performing cross-modal fusion of the compensated sensing data with meteorological radar and satellite remote sensing data includes: Constructing a hydrological feature sub-tensor; Projecting the meteorological radar reflectivity tensor and the remote sensing deformation tensor to the hydrological feature space of the hydrological feature sub-tensor through a decomposition algorithm to generate a coupling tensor.

5. The multimodal feature fusion method for a wireless sensor network for flood disasters according to claim 3, wherein The dynamic feature distillation includes a forward propagation path and a backward propagation path, specifically including: In the forward propagation path, using hydrological features, meteorological features and geological features as query vectors, key vectors and value vectors respectively, calculating the dynamic correlation degree between different feature channels, and generating a weight distribution map; Performing a Hadamard product operation on the weight distribution map and the coupling tensor to output a preliminarily optimized feature subspace; In the backward propagation path, constructing a feature selector based on gradient sensitivity, screening the feature subspace through the feature selector, and applying a hard truncation to the feature dimensions with a gradient magnitude lower than a certain value to generate a sparsification mask matrix; Performing a bitwise multiplication of the sparsification mask matrix and the feature subspace control to output a six-dimensional feature matrix.

6. The multimodal feature fusion method for a wireless sensor network for flood disasters according to claim 1, wherein The first processing network includes: A spatio-temporal feature extraction layer for performing spatio-temporal joint modeling on the multi-modal feature matrix and outputting a feature mapping matrix, where the spatio-temporal joint modeling includes a spatio-temporal convolutional kernel group dynamically adjusted based on timestamps and spatial grids. The spatio-temporal convolutional kernel group slides along the time dimension to capture the evolution trend of hydrological parameters and expands along the spatial dimension to model the flood propagation correlation between basin nodes; A branch feature extraction layer for processing the feature mapping matrix through a neural network model and outputting branch features; A feature output layer for adaptively allocating weights for each feature dimension through an attention mechanism. The attention mechanism includes a channel attention mechanism and a spatial attention mechanism, and outputs the branch features after dynamic channel and spatial weighting through an activation function.

7. The multimodal feature fusion method for a wireless sensor network for flood disasters according to claim 6, characterized in that, The branch feature extraction layer includes a hydrological branch, a meteorological branch, and a geological branch, where: The hydrological branch extracts multi-scale runoff features through a dilated convolutional network; The meteorological branch captures the spatio-temporal propagation pattern of precipitation through a bidirectional gated recurrent unit; The geological branch models the spatial topological relationship of the soil permeability in the basin through a graph convolutional network.

8. The multimodal feature fusion method for a wireless sensor network for flood disasters according to claim 6, wherein The second processing network includes: A data mapping layer for modeling the dynamic relationship between environmental data and the output of the first processing network through a physically-driven non-linear mapping mechanism to generate an influence weight matrix; A feature reconstruction layer for performing per-dimensional reconstruction on the output of the first processing network according to the influence weight matrix; An output layer for performing consistency correction on the reconstructed features through a bidirectional collaborative optimization mechanism, and non-linearly coupling the features through the domain and spatial domain according to the corrected features to calculate the disaster assessment result.

9. A multi-modal feature fusion system for a wireless sensor network for flood disasters, which is used to implement the multi-modal feature fusion method for a wireless sensor network for flood disasters described in any one of claims 1-8, and is characterized in that, The system includes a data acquisition module, a data processing module, and a result feedback module, where: The data acquisition module is used to obtain multi-dimensional sensing data of the basin through a time window, where the length of the time window is adaptively adjusted according to the precipitation intensity change rate, and the window sliding step size has a non-linear mapping relationship with the basin runoff propagation speed; The data processing module is used to construct a multi-modal feature matrix according to the output of the data acquisition module and input the multi-modal feature matrix into a preset intelligent AI model to output a disaster assessment result; The result feedback module is used to feedback the output of the data processing module to the monitoring platform to generate a dynamic risk map of the basin and send a graded warning signal.

10. The multi-modal feature fusion system of the wireless sensor network for flood disasters according to claim 9, characterized in that, The data processing module includes: A matrix generation unit for mapping and organizing multi-dimensional sensing data into a multi-modal feature matrix according to spatio-temporal relationships; A feature extraction unit configured with a first processing network to extract features from the multi-modal feature matrix; An output compensation unit configured with a second processing network to perform environmental factor compensation on the output of the first processing network through a compensation factor in the multi-modal feature matrix.

Images