Method for adaptive spatio-temporal alignment of perception information in water disaster scene

Abstract

The application provides a water disaster field sensing information adaptive space-time alignment method, and the specific steps are as follows: step one, water disaster field environment monitoring Internet of Things simulation building; step two, simulation data preprocessing; step three, constructing a water disaster field sensing information adaptive space-time alignment model; and step four, realizing water disaster field sensing data space-time alignment. The method can be applied to the space-time alignment of multi-source heterogeneous sensing information of a water disaster field and has generality.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to an adaptive spatiotemporal alignment method for sensing information in water disaster sites. Background Technology

[0002] Current aquatic disaster site perception systems primarily rely on heterogeneous Internet of Things (IoT) to achieve comprehensive information collection and transmission. Data types encompass multi-dimensional and multi-modal data, including optical images, sonar, audio, and time-series numerical data. Due to significant differences in acquisition principles, hardware parameters, sampling frequencies, and transmission mechanisms among various sensors, coupled with the influence of complex hydrological environments such as underwater turbidity changes, light refraction, water body color, underwater geological structures, and turbulent currents, the collected data generally suffers from quality issues such as low signal-to-noise ratio, sparse effective features, severe noise superposition, and insufficient completeness. This makes the data prone to anomalies and defects, a core challenge distinguishing it from stable land-based data collection scenarios.

[0003] Against this backdrop, the use of water disaster monitoring data faces two prominent challenges: First, spatiotemporal misalignment of multi-source data. Underwater sensors are prone to dynamic drift due to water flow, resulting in unstable deployment. Simultaneously, severe attenuation and random latency fluctuations in underwater wireless transmission channels lead to issues such as timestamp asynchrony, inconsistent sampling intervals, and inconsistent spatial coordinate benchmarks among different devices, significantly increasing the difficulty of spatiotemporal alignment of multi-source data. Second, data loss exhibits dynamic and time-varying characteristics. When the water environment is stable, external interference is weak, data acquisition is relatively good, and the loss rate is relatively low. However, after a sudden emergency, drastic changes in water flow, unstable links, and intensified environmental disturbances can cause the loss rate to rise rapidly. The random switching between high and low loss scenarios and the significant differences in data status necessitate that models possess dynamic loss adaptation capabilities for complex water environments.

[0004] Currently, the industry uses the Transformer model for spatiotemporal alignment. However, a fixed regularization weight λ is commonly used as a global hyperparameter to adjust the relative contributions of reconstruction loss and Fourier completion loss. This approach fails to adapt to the dynamic missing data characteristics of aquatic disaster data, exhibiting significant adaptation defects. Therefore, the regularization weight λ of the Fourier completion loss (FIL) is chosen as the core hyperparameter to control the strength of the model's frequency domain sparsity constraint, directly determining the model's accuracy in compensating for missing data and its noise suppression capability. In scenarios with extremely large fluctuations in the missing rate, in low-missing-rate scenarios, a fixed λ value that is too large will over-constrain effective features in the frequency domain, resulting in overly smooth completed data and the loss of subtle hydrological changes and detailed disaster field features. In high-missing-rate scenarios, a fixed λ value that is too small will fail to effectively suppress strong noise, data distortion, and other interference, making it difficult to accurately reconstruct the true disaster field data distribution.

[0005] To address the issues of random fluctuations in data loss rates and constantly changing spatiotemporal misalignment in dynamic aquatic disaster sites, it is necessary to construct a spatiotemporal alignment and completion model with dynamic sensing and adaptive adjustment capabilities. This model can adaptively adjust key parameters based on the data loss rate, achieving accurate spatiotemporal alignment and effective completion of multi-source monitoring data. This improves the model's stability, adaptability, and completion accuracy in complex aquatic environments. It provides continuous, complete, and spatiotemporally consistent multi-source sensing data support for operational needs such as aquatic disaster situation assessment, emergency command, and decision analysis. Summary of the Invention

[0006] The purpose of this invention is to address the problems in existing technologies by proposing an adaptive spatiotemporal alignment method for aquatic disaster site perception information. This method is applicable to spatiotemporal completion of data with different missing rates.

[0007] This invention is achieved through the following technical solution: This invention proposes an adaptive spatiotemporal alignment method for water disaster site perception information, the specific steps of which are as follows:

[0008] Step 1: Establishing an IoT simulation platform for environmental monitoring at aquatic disaster sites;

[0009] Step 2: Simulation data preprocessing;

[0010] Step 3: Construct an adaptive spatiotemporal alignment model for water disaster site perception information;

[0011] Step 4: Achieve spatiotemporal alignment of disaster site perception data in water areas.

[0012] Furthermore, in step one, the IoT network for monitoring the water disaster site is simulated using the Simulink visualization tool in MATLAB. Temperature sensors, water depth sensors, flow velocity sensors, and turbidity sensors are simulated using modules in Simulink, and the simulation data is saved.

[0013] Furthermore, in step two, timestamps and location information are extracted from the simulation data, the timestamps are converted to UTC time, and coordinate transformation is performed on the data in the non-standard coordinate system. The transformation is calculated to complete the data coordinate system one.

[0014] Furthermore, in step three, an adaptive spatiotemporal alignment model for water disaster site perception information is constructed. In the ImputeFormer model framework, an adaptive Fourier completion loss weight generation mechanism with missing rate perception is designed. The reconstruction loss and Fourier completion loss are weighted and fused to obtain the overall loss function.

[0015] Furthermore, in step four, the model first fuses temporal encoding and node features through a spatiotemporal input embedding module to generate an initial hidden representation. Then, it sequentially passes through a temporal projection attention module, a Transformer encoder, and a spatial embedding attention module to complete temporal dependency modeling and multi-sensor spatial feature aggregation. The fused spatiotemporal representation is used to generate the completion result, achieving spatiotemporal alignment of multi-source data. The model calculates the sample missing rate based on the missing mask and generates dynamic Fourier loss weights by combining globally adjusted hyperparameters. The reconstruction loss and the dynamic weighted Fourier completion loss are combined into a total loss function. The model parameters are updated through backpropagation until training converges. The water disaster field perception data to be processed and its missing mask are input into the trained model. The model predicts the completion value of the original missing position and together with the existing observations, they form a complete spatiotemporal sequence. Finally, the aligned water disaster field perception data is output.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] This invention provides a novel solution for spatiotemporal alignment of water disaster site data. Compared to existing inventions, its innovations are as follows: It proposes a combination of Imputeformer and a missing rate-aware adaptive Fourier completion loss weight generation mechanism, which can dynamically adjust according to different data missing states, thereby improving computational accuracy and efficiency. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0019] Figure 1 This is a schematic diagram of the adaptive spatiotemporal alignment method for water disaster site perception information described in this invention;

[0020] Figure 2 This is a simulation design diagram of the Internet of Things for monitoring the water disaster site environment according to the present invention;

[0021] Figure 3 This is a schematic diagram of the adaptive spatiotemporal alignment model for water disaster site perception information of the present invention. Detailed Implementation

[0022] 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.

[0023] See Figures 1-2 - Figure 3 This invention proposes an adaptive spatiotemporal alignment method for water disaster site perception information. Figure 1 The flowchart of the adaptive spatiotemporal alignment method for water disaster site perception information provided by the present invention is shown. The specific steps of the method are as follows:

[0024] Step 1: Establishing an IoT simulation platform for environmental monitoring at aquatic disaster sites;

[0025] In step one, the temperature sensor operates by adding random noise to the incremental trend adjustment. The trend adjustment has a 70% probability of maintaining the direction; otherwise, it changes randomly. Each change is between 0.05 and 0.2 degrees Celsius, plus ±0.05 degrees of noise. The initial temperature is 15°C, the highest temperature is 21°C, and the lowest temperature is 13°C. The formula is as follows:

[0026]

[0027] The water depth sensor uses a cosine function to simulate tides, superimposed with Gaussian noise, and limits the single change to no more than 0.5 m / s. The tidal part uses a cosine function, which may simulate periodic changes, such as the periodicity of tides. The noise represents environmental disturbances, and the continuity constraint ensures that the data will not change abruptly, which is consistent with the inertia of the actual water body. The reference water depth is 5 meters, the tidal amplitude is 2 meters, the tidal frequency is 0.1 Hz, and the disturbance amplitude is 0.15 meters.

[0028] The flow velocity sensor uses a cosine function superimposed with random disturbances to ensure that each change does not exceed 0.2 meters and maintains a consistent trend. The cosine function may simulate periodic flows, such as tidal flows, while the noise represents turbulent or transient flows. Limiting the amplitude of changes ensures data continuity and conforms to fluid dynamics. The initial flow velocity is 0, the minimum static velocity is -0.1 m / s, and the maximum turbulent velocity is 1.5 m / s. The flow velocity variation trend is as follows: .

[0029] The turbidity sensor uses trend adjustment, with a 60% probability of maintaining direction, 30% random variation, and 10% stability. The variation is increased with Gaussian noise, and the amplitude of a single variation is limited. This simulates the trend and random perturbation of turbidity changes, which may reflect the effects of sediment suspension or biological activity. The initial turbidity is 45 NTU, the minimum turbidity is 0 NTU, the maximum turbidity is 100 NTU, and the random noise amplitude is set to 0.3.

[0030] Step 2: Simulation data preprocessing;

[0031] The system retrieves the timestamp and location information, converts the timestamp to UTC time, performs coordinate transformation on data in non-standard coordinate systems, calculates the transformation, and completes the data coordinate system unification.

[0032] Step 3: Construct an adaptive spatiotemporal alignment model for water disaster site perception information;

[0033] Spatiotemporal input embedding module:

[0034] The model's spatiotemporal input embedding module is responsible for integrating the original input, temporal encoding, and node embeddings into a hidden state that the model can process. Its input dimension expansion and projection can be represented as follows:

[0035]

[0036] The Unsqueeze(·) operation changes the dimension from R to R. N×T Mapped to R N×T×1 , ∈R N×T×D As the initial hidden representation, this operation allows the model to learn data-related weights to aggregate information from other time points. The update process is as follows:

[0037]

[0038] In the formula, ℱℓ(·): R T×D →R T×T This represents the data-driven information interaction function of the ℓth layer.

[0039] The model has a timestamp encoding mechanism. Considering the relatively limited time span of the input sequence, it focuses only on intraday time features and adopts a sinusoidal positional encoding scheme to inject intraday time information into each time series. The calculation method is as follows:

[0040]

[0041] Where, p t This is the intraday time index corresponding to the t-th time step in the sequence. The daily time-scale mapping coefficients are obtained by splicing sine codes. Cosine encoding Finally, the timestamp encoding vector Ut:t+T∈R is obtained. T×2 .

[0042] The spatiotemporal input embedding of a single node in the model is as follows:

[0043]

[0044] in, As input for subsequent spatiotemporal interaction modules.

[0045] Time-projection attention module:

[0046] The model first projects initial features into a dense representation space through attention operations on low-dimensional vectors, and then randomly initializes a learnable and gradient-differentiable shared projection vector. Then hide the state Mapped to the projection space through attention interaction with the projection query vector:

[0047]

[0048] In the formula W is the projected feature vector. Q W K W V These are the weights for the linear mapping.

[0049] By distributing the information in the compact representation across the full-length sequence, the complete sequence can be reconstructed, and the projected features are restored to the original space as follows:

[0050]

[0051] The above process will be integrated into the Transformer encoder to complete the feature update in the time dimension. The Transformer encoder is integrated as follows:

[0052]

[0053] Z i,(ℓ+1) The interpolation features are output by the ℓth layer time interaction module.

[0054] Spatial Embedded Attention Module:

[0055] Spatial correlation matrix calculation is the core module of spatial attention:

[0056]

[0057] Spatial attention output updated to:

[0058]

[0059] Low-complexity spatial incidence matrix approximation:

[0060]

[0061] Where σ(·) is a shorthand for the softmax function.

[0062] Fourier interpolation loss module

[0063] Let the original observation mask be M, where M i =1 indicates that the i-th spatiotemporal variable element has a real observed value, M i =0 indicates that the element is missing at that location. For a single sample, the missing rate is defined as the ratio of the number of missing elements to the total number of elements in the sample:

[0064] The formula for calculating the missing rate is:

[0065]

[0066] Where r is the missing rate of the samples (ranging from [0,1]), and N missing For the original missing mask M missing The number of missing elements, N, is marked in the table. total Let N be the total number of elements in the sample. If the sample contains T time steps, N space stations, and C variables, then N... total =TNC.

[0067] Missing rate based on observation mask:

[0068]

[0069] The missing rate is determined by the missing information of the original input samples themselves, and does not depend on manually assigned fixed weights.

[0070] Dynamic loss weights:

[0071]

[0072] Where, λ dynamic α is the weight of the dynamic Fourier completion loss; α is the global adjustment hyperparameter (empirical range [0.03, 0.07]); ε is a minimal constant to prevent the denominator from being zero; and r is the sample missing rate calculated in the formula.

[0073] Since the original input sequence contains missing values, it cannot be directly subjected to a Fourier transform. First, a complete sequence is constructed. Let X be the actual observed sequence. The complete completion sequence is defined as the completion result output by the model:

[0074] Complete sequence definition

[0075]

[0076] in, This represents a complete sequence composed of both true observations and model-completed values. This definition means that true observations are retained at the original observation locations, while model-predicted values ​​are used at the original missing locations. This avoids directly performing frequency domain transformation on the sequence containing missing values ​​and ensures that frequency domain constraints apply to the final complete spatiotemporal sequence.

[0077] After obtaining the complete sequence, a Fourier transform is performed on it, and the mean amplitude in the frequency domain is used as the Fourier completion loss. The formula for the Fourier completion loss is as follows:

[0078]

[0079] in, Indicates Fourier transform, Mean(·) represents the frequency domain amplitude, and Mean(·) represents averaging the frequency domain amplitude. This loss is used to constrain the frequency domain structure of the completed sequence, so that the model can still generate a stable completion result even when the missing values ​​are severe.

[0080] This paper retains the reconstruction loss from the original ImputeFormer as the main supervision term. The reconstruction loss is calculated during the training phase at locations where the model is artificially occluded and has real observations, and is used to constrain the model's predicted values ​​to be close to the true values. The reconstruction loss formula is:

[0081]

[0082] Among them, M rec This is the reconstruction mask during the training phase, where ⊙ represents element-wise multiplication.

[0083] The total loss function under dynamic weights is:

[0084]

[0085] in To reconstruct the loss, To compensate for the Fourier loss.

[0086] Step 4: Achieve spatiotemporal alignment of disaster site perception data in water areas;

[0087] In step four, the masked water area perception data is input into the constructed adaptive spatiotemporal alignment model. The original data is first fused with temporal coding and node features by the spatiotemporal input embedding module to generate initial hidden representations. The resulting representations are then fed into the temporal projection attention module to complete feature projection reconstruction, and the temporal features are updated using the Transformer encoder. The temporally optimized features are then input into the spatial embedding attention module, where spatial attention and approximate matrix operations are used to aggregate multi-sensor spatial features. The intermediate features that fuse spatiotemporal information output the completion results, completing the spatiotemporal alignment of multi-source data. Simultaneously, the features are fed into the Fourier interpolation loss module, which calculates the sample missing rate based on the missing mask and obtains dynamic loss weights by combining fixed hyperparameters. The total loss function is constructed by combining the reconstruction loss and Fourier completion loss using dynamic weights. The parameters of each module of the model are iteratively updated using backpropagation, and the training is repeated until the loss converges. The water area disaster perception data to be processed and its missing mask are input into the trained model, which predicts the completion values ​​of the original missing positions and forms a complete spatiotemporal sequence together with the existing observations. Finally, the aligned water area disaster perception data is output.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. An adaptive spatiotemporal alignment method for water disaster site perception information, characterized in that, The specific steps of the method are as follows: Step 1: Establishing an IoT simulation platform for environmental monitoring at aquatic disaster sites; Step 2: Simulation data preprocessing; Step 3: Construct an adaptive spatiotemporal alignment model for water disaster site perception information; Step 4: Achieve spatiotemporal alignment of disaster site perception data in water areas.

2. The method according to claim 1, characterized in that: The Simulink visualization tool in MATLAB is used to simulate an IoT network for environmental monitoring in aquatic disaster sites. Simulink modules are used to simulate temperature sensors, water depth sensors, flow velocity sensors, and turbidity sensors, and the simulation data is saved.

3. The method according to claim 2, characterized in that: The simulation data is extracted for timestamps and location information. The timestamps are converted to UTC time. For data in non-standard coordinate systems, coordinate transformation is performed. The transformation is calculated to complete the data coordinate system unification.

4. The method according to claim 3, characterized in that: An adaptive spatiotemporal alignment model for water disaster site perception information is constructed. In the ImputeFormer model framework, an adaptive Fourier completion loss weight generation mechanism with missing rate perception is designed. The reconstruction loss and Fourier completion loss are weighted and fused to obtain the overall loss function.

5. The method according to claim 4, characterized in that: The model first fuses temporal encoding and node features through a spatiotemporal input embedding module to generate initial hidden representations. Then, it sequentially passes through a temporal projection attention module, a Transformer encoder, and a spatial embedding attention module to complete temporal dependency modeling and multi-sensor spatial feature aggregation. The fused spatiotemporal representations are used to generate completion results, achieving spatiotemporal alignment of multi-source data. The model calculates the sample missing rate based on the missing mask and generates dynamic Fourier loss weights by combining globally adjusted hyperparameters. The reconstruction loss and the dynamic weighted Fourier completion loss are combined into a total loss function, and the model parameters are updated through backpropagation until training converges. The water disaster field perception data to be processed and its missing mask are input into the trained model. The model predicts the completion values ​​of the original missing positions and together with the existing observations, they form a complete spatiotemporal sequence. Finally, the aligned water disaster field perception data is output.

Images