Sensor network unsupervised spatio-temporal feature learning method and system for incomplete scene

Abstract

This invention provides an unsupervised spatiotemporal feature learning method and system for sensor networks in incomplete scenarios. Applied to the field of wireless sensor network technology, the method includes: acquiring an observation data matrix with missing values ​​and a corresponding missing indicator matrix based on a constructed sensor network; inputting the observation data matrix with missing values ​​and the missing indicator matrix into a sequence module to extract multi-scale temporal correlation representations, generating temporal embedding representations of nodes; inputting the temporal embedding representations into a grouped graph neural network module, combining graph structure priors and an adaptive graph learning mechanism to establish spatial dependencies between nodes in the sensor network, and updating the grouped representations of nodes in the sensor network through clustering consistency constraints, outputting a spatiotemporal feature representation. This invention's method can gradually converge to a stable representation through single-sample iterative optimization without relying on a large number of training samples.

Description

Technical Field

[0001] This invention relates to the field of wireless sensor network technology, and in particular to an unsupervised spatiotemporal feature learning method and system for sensor networks in incomplete scenarios. Background Technology

[0002] With the development of ubiquitous sensing technology, wireless sensor networks composed of large-scale distributed sensor nodes have been widely used in environmental monitoring, industrial automation, and smart cities. However, wireless sensor networks typically operate in open environments for extended periods, making them susceptible to hardware failures, link fading, and energy constraints, leading to gaps in observation data across both time and space. Data gaps not only disrupt the statistical consistency of time series data and introduce error accumulation in sequence analysis, but also increase the uncertainty in estimating inter-node dependencies. Furthermore, the limited number of effective samples available to the system in open environments means that the problems of data gaps and limited sample size coexist. Limited samples are insufficient to support the generation of reliable statistical models, while data gaps further reduce the completeness and usability of the data on top of the insufficient sample size.

[0003] Existing research primarily focuses on either the problem of missing data or the problem of limited sample size, making it difficult to effectively address the real-world scenario where both exist simultaneously in the operation of wireless sensor networks (this invention defines this situation as multiple information incompleteness). Furthermore, the applicability of related methods is further limited when large-scale pre-training data is lacking to support knowledge transfer. When missing data and limited sample size coexist, their combined effect imposes a dual constraint on the model's representation learning ability and generalization performance: missing data weakens the model's ability to capture spatiotemporal structure and introduces noise interference, while limited sample size reduces the stability of parameter estimation and increases the risk of overfitting.

[0004] Therefore, there is an urgent need to develop unsupervised spatiotemporal feature learning methods and systems for sensor networks in incomplete scenarios to solve the above problems. Summary of the Invention

[0005] This invention provides an unsupervised spatiotemporal feature learning method and system for sensor networks in incomplete scenarios. It solves the technical problems in the prior art, such as data deficiency weakening the model's ability to capture spatiotemporal structure and introducing noise interference, and limited sample size reducing the stability of parameter estimation and increasing the risk of overfitting.

[0006] According to a first aspect of the present invention, an unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios is provided, wherein the incomplete scenarios include scenarios where sensor network observation data has random missing values ​​and scenarios where the effective training sample size is limited, the method comprising: Based on the constructed sensor network, an observation data matrix with missing values ​​and the corresponding missing value indication matrix are obtained; The observation data matrix with missing values ​​and the missing indicator matrix are input into the sequence module to extract multi-scale temporal correlation representations and generate temporal embedding representations of nodes. The time-embedded representation is input into the grouped graph neural network module. Combining graph structure prior and adaptive graph learning mechanism, spatial dependencies between nodes in the sensor network are established. The grouped representation of nodes in the sensor network is updated by clustering consistency constraints, and spatiotemporal feature representation is output. Specifically, the sequence-to-sequence module and the grouped graph neural network module are trained using the effective training samples.

[0007] Furthermore, the observation data matrix with missing values ​​and the missing indicator matrix are input into the sequence module to extract multi-scale temporal correlation representations and generate temporal embedding representations of the nodes. Specifically, this includes the following steps: Local convolutional units are used to preprocess the observation data matrix and the missing indicator matrix with missing values ​​to obtain the updated observation data matrix and the updated missing indicator matrix. The updated observation data matrix and the updated missing indicator matrix are input into the time-domain feature extractor to output time-domain features, and into the frequency-domain feature extractor to output frequency-domain features. The time-domain features and frequency-domain features are fused through concatenation, linear mapping, and combination operations to generate a time-embedded representation; The time-domain feature extractor is used to capture local dynamic changes and short-term dependencies using multi-layer one-dimensional convolution; the frequency-domain feature extractor is used to map the updated observation data matrix to the frequency domain through discrete Fourier transform, and to perform nonlinear transformation on the real and imaginary parts of the mapped spectrum using gated linear units.

[0008] Furthermore, by combining graph structure priors with adaptive graph learning mechanisms, spatial dependencies between nodes in the sensor network are established, specifically including: Randomly initialize the learnable parameter matrix, and construct a graph structure based on the learnable parameter matrix; A spatial adjacency matrix is ​​constructed based on the relationships between various sensor nodes in the sensor network; Calculate the graph structure regularization term based on the graph structure and the spatial adjacency matrix; The graph structure regularization term includes a dynamic weight factor, which is obtained based on the current training round number and the total training round number.

[0009] Furthermore, the node representations in the sensor network are updated by grouping them through clustering consistency constraints, and the updated node embedding representation matrix is ​​output. This process includes the following steps: Based on the time embedding representation, observation data matrix, and missing indicator matrix, the reconstructed complete data is calculated, and the complete data is iteratively updated to obtain the updated complete data. Clustering algorithms are used to perform cluster analysis on the updated complete data, and cluster group sets are output. Node embedding is performed on each cluster in the cluster group sets to obtain a feature matrix. Linear transformation parameters are assigned to each feature matrix to obtain a cluster-level feature matrix. The cluster-level feature matrix is ​​mapped to the global node space by the index recovery operator, and an updated node embedding representation matrix is ​​generated by performing graph convolution operation. The index recovery operator is used to represent the correspondence between the global node space and the local index space within the cluster.

[0010] Furthermore, the clustering consistency constraint also includes: The cluster centers from the previous training round are used to constrain the grouping representation update process in the current round, so as to reduce the distance between node representations within the same cluster and increase the difference between node representations in different clusters. The cluster centers are updated using the following formula: ; in, Indicates the updated cluster centers. This represents the set of nodes corresponding to the k-th cluster. This indicates the number of nodes in the corresponding node set. This represents the embedding representation value corresponding to the i-th sensor node in the node embedding representation matrix.

[0011] Furthermore, it also includes: A two-stage alternating optimization strategy is used to update the parameters and graph structure of the sequence-to-sequence module and the grouped graph neural network module until convergence. The two-stage alternating optimization strategy includes alternating the execution of graph structure optimization with fixed parameters for the sequence-to-sequence module and the grouped graph neural network module, and parameter updates for the sequence-to-sequence module and the grouped graph neural network module with fixed graph structure.

[0012] Furthermore, the two-stage alternating optimization strategy specifically includes: Phase 1: Fix the parameters of the sequence-to-sequence module and the grouped graph neural network module, and optimize the graph structure with the goal of minimizing reconstruction error, cluster consistency loss and graph structure regularization term; Second stage: Fix the graph structure and update the parameters of the sequence-to-sequence module and the grouping graph neural network module with the goal of minimizing reconstruction error and cluster consistency loss; Repeat the first and second phases until the overall optimization objective function of the sequence-to-sequence module and the grouped graph neural network module converges.

[0013] According to a second aspect of the present invention, the present invention provides an unsupervised spatiotemporal feature learning system for sensor networks in incomplete scenarios, comprising: The data acquisition module is used to acquire the observation data matrix with missing values ​​and the corresponding missing indicator matrix based on the constructed sensor network. An embedded representation module, connected to the data acquisition module, is used to input the observation data matrix with missing values ​​and the missing indicator matrix into the sequence to extract multi-scale temporal correlation representations and generate temporal embedded representations of nodes. The clustering representation module, connected to the embedding representation module, is used to input the temporal embedding representation into the grouped graph neural network module. Combining graph structure priors and adaptive graph learning mechanisms, it establishes spatial dependencies between nodes in the sensor network and updates the grouped representations of nodes in the sensor network through clustering consistency constraints, outputting spatiotemporal feature representations.

[0014] According to a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the method as described in the first aspect.

[0015] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, characterized in that a computer program is stored thereon; the computer program is executed by a processor to implement the method as described in the first aspect.

[0016] The beneficial effects of this invention are: This invention constructs a sequence-to-sequence module to characterize multi-scale temporal correlations under the condition of missing inputs; then, based on the temporal embedding representation, a grouped graph neural network is introduced to establish spatial dependencies between nodes.

[0017] This invention uses graph topology priors to limit the search space of graph structures and improves the structural stability of the embedding space through cluster consistency constraints. At the same time, it designs an alternating optimization strategy for graph structures and network parameters to avoid the amplification of noise correlations during the iteration process, thereby improving the overall training stability of the model in scenarios with multiple incomplete information.

[0018] This invention represents wireless sensor network observation data as a unified spatiotemporal graph. By treating the model architecture as an implicit structural prior that characterizes the inherent spatiotemporal correlation of the data, and embedding the parameter optimization process into the inference stage, the model can gradually converge to obtain a stable representation without relying on a large number of training samples through single-sample iterative optimization.

[0019] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0020] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein: Figure 1 A flowchart of the unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios provided by an embodiment of the present invention is shown; Figure 2 The flowcharts of the sequence-to-sequence module and the grouped graph neural network module provided in the embodiments of the present invention are shown; Figure 3 A flowchart illustrating the optimization process of the graph structure provided in an embodiment of the present invention is shown; Figure 4 A block diagram of an unsupervised spatiotemporal feature learning system for sensor networks in incomplete scenarios, provided in an embodiment of the present invention, is shown. Figure 5 A block diagram of an electronic device according to an embodiment of the present invention is shown; Figure 6 The MSE values ​​of different algorithms under different missing rates provided in the embodiments of the present invention on the TJ dataset are shown. Figure 7 The MAE values ​​of different algorithms under different missing rates provided in this embodiment of the invention are shown on the TJ dataset. Figure 8 The MAE values ​​of different algorithms under different missing rates provided in the embodiments of the present invention are shown on the Intel dataset; Figure 9 The MSE values ​​of different algorithms under different missing rates provided in the embodiments of the present invention are shown on the Intel dataset; Figure 10 This diagram illustrates the RoG-STGNN representation similarity results under different missing rates on the TJ dataset, as provided in an embodiment of the present invention. Figure 11 This diagram illustrates the similarity results of RoG-STGNN representations under different missing rates on the Intel dataset, as provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0023] This invention provides an unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios. It is applicable to scenarios where sensor network observation data has random missing values ​​and where the size of the effective training sample is limited. (See also...) Figures 1 to 3 This includes the following steps: S1. Based on the constructed sensor network, obtain the observation data matrix with missing values ​​and the corresponding missing value indication matrix.

[0024] Consider by A wireless sensor network consisting of sensor nodes can be represented as a graph. ,in, Represents a set of sensor nodes. This represents a set of spatial relationships between sensor nodes. In a length of... Within the time window, the observation data matrix collected by each sensor node can be represented as: .

[0025] Due to factors such as hardware failure, communication interference, and energy consumption limitations, the observation data matrix is ​​prone to random missing data. A missing data indicator matrix is ​​constructed and denoted as... Any item in this matrix can be represented as , Time indicates sensor node At any moment The observations are available. The time period indicates that the item is missing.

[0026] S2. Input the observation data matrix with missing values ​​and the missing indicator matrix into the Sequence to Sequence (Seq2Seq) module to extract multi-scale temporal correlation representations and generate temporal embedding representations of nodes.

[0027] In real-world deployments, only limited time segments are typically available for analysis, resulting in small sample sizes and missing data, creating scenarios with multiple layers of incomplete information. Under these conditions, supervised spatiotemporal representation methods that rely on large-scale, complete samples for training struggle to obtain stable and reliable representations.

[0028] How can we obtain a robust and discriminative node embedding representation matrix without relying on large-scale pre-training data, given only a single short-term window of samples and with random missing samples? '( Corresponding node The embedded representation allows it to maintain the spatiotemporal correlation structure in the original data while remaining stable against missing patterns and changes in sample size.

[0029] This invention constructs the following mapping function: ; in, Represents the spatial adjacency matrix of nodes. Represents a mapping function. 'Indicates the node embedding representation matrix.

[0030] The mapping function must satisfy the following conditions simultaneously: 1. The learned embedding representations can recover the latent complete structure of the original observations under missing conditions to minimize reconstruction error, thereby ensuring reconstruction consistency; 2. The embedding representations obtained under different missing rates remain highly consistent to ensure stability; 3. The learned embedding representation can accurately distinguish the intrinsic attributes of node-collected data through a simple linear classifier, demonstrating discriminative power; 4. The model parameters can converge through single-sample iterative optimization without relying on large-scale data training, thus ensuring its feasibility with small samples.

[0031] Based on the above considerations, this invention proposes an unsupervised spatiotemporal representation learning algorithm (RoG-STGNN algorithm) for scenarios with multiple information incompleteness. The overall architecture is as follows: Figure 2 The diagram shows the first stage of training, where the parameters of each module are fixed and the graph structure parameters are optimized. It includes all of the following: When the number of samples is limited, time series representations cannot extract stable temporal dependencies under conditions of incomplete observation. To address this, this invention constructs a Seq2Seq module for missing data perception.

[0032] The Seq2Seq module includes the following steps: Local convolutional units are used to preprocess the observation data matrix and the missing indicator matrix with missing values ​​to obtain the updated observation data matrix and the updated missing indicator matrix. The updated observation data matrix and the updated missing indicator matrix are input into the time-domain feature extractor to output time-domain features, and into the frequency-domain feature extractor to output frequency-domain features. The time-domain features and frequency-domain features are fused through concatenation, linear mapping, and combination operations to generate a time-embedded representation; The time-domain feature extractor is used to capture local dynamic changes and short-term dependencies using multi-layer one-dimensional convolution; the frequency-domain feature extractor is used to map the updated observation data matrix to the frequency domain through discrete Fourier transform, and to perform nonlinear transformation on the real and imaginary parts of the mapped spectrum using gated linear units.

[0033] Specifically, the following detailed explanation is provided: The observation data matrix with missing values ​​and the missing value indication matrix are jointly input into the Sequence (Seq2Seq) module. This module includes a local convolutional unit, a time-frequency dual-branch network constructed based on one-dimensional convolution and discrete Fourier transform, and a time-frequency joint modeling unit. Through the aforementioned structure, the missing patterns are dynamically perceived while establishing complex time dependencies, thereby effectively suppressing the propagation of missing noise in the deep network and outputting a time-level embedding.

[0034] (1) Local convolutional unit A local convolutional unit is introduced to preprocess the observation data matrix with missing values. This unit employs partial convolution and a missing indicator matrix update mechanism, ensuring that only available observations participate in the calculation during each sliding window, preventing the bias caused by missing values ​​from affecting subsequent feature learning. See the following formulas, which represent the embedding and missing indicator matrix update processes, respectively: ; ; in, Represents sensor nodes The embedding obtained after local convolution is the updated observation data matrix; This represents the corresponding value of sensor node i in the observation data matrix; Represents sensor nodes The corresponding value in the missing indicator matrix; sum(·) represents summation. To and Vectors with the same shape and all elements being 1. This indicates element-wise multiplication. This represents the updated missing indicator matrix. Represents the local convolution weight matrix. This represents the local convolution bias vector.

[0035] (2) Time-frequency dual-branch network The network includes parallel time-domain feature extractors and frequency-domain feature extractors.

[0036] Temporal feature extractors utilize multiple layers of one-dimensional convolutions to capture local dynamic changes and short-term dependencies. As the number of layers increases, its receptive field gradually expands, thereby progressively integrating trend information over a longer time span. The temporal feature extractor outputs temporal features, denoted as... .

[0037] The frequency domain feature extractor uses the discrete Fourier transform to update the observation data matrix. Mapping to the frequency domain, we obtain the spectral representation. ,in, Indicates taking the real part, This indicates taking the imaginary part. This indicates the performance of a Discrete Fourier Transform. Spectral representation can explicitly reveal periodicity and global patterns, providing additional structural clues to the model when the sample size is insufficient.

[0038]

[0039] ; ; in, This indicates the result of the real-world control transformation. This represents the result of the virtual departmental control transformation. This represents the frequency domain characteristics of the output. This indicates the solution to the inverse Fourier transform.

[0040] The above formula indicates the use of multi-layer gate control units. and For the real part respectively With the imaginary part Nonlinear transformations are performed to suppress noise and highlight discriminative frequency components.

[0041] Since the original input is a real-valued vector, only the real part of the inverse Fourier transform is retained as the frequency domain feature output by the frequency domain feature extractor. .

[0042] (3) Time-frequency joint modeling unit Temporal and frequency domain features are fused using a time-frequency joint modeling unit to generate a temporal embedding representation. See the following formula: ; ; in, Represents the gate vector, This represents the modeling weight matrix. This represents the modeling bias vector, and m||n represents concatenating m and n.

[0043] Will and The features are concatenated, and then linearly mapped to generate a gating vector. Finally, the temporal embedding representation is obtained by combining time and frequency feature representations. The temporal embedding representation takes into account instantaneous fluctuations, short-term changes, long-term trends, and periodic patterns. Features at different scales complement each other, improving the stability of the representation.

[0044] S3. Input the time embedding representation into the grouped graph neural network module, combine graph structure prior and adaptive graph learning mechanism to establish spatial dependencies between nodes in the sensor network, and update the grouped representation of nodes in the sensor network through clustering consistency constraints to output spatiotemporal feature representation; wherein, the sequence-to-sequence module and the grouped graph neural network module are trained using the effective training samples.

[0045] The temporal embedding representation is input into the Grouped Graph Neural Network (GGNN) module. This module does not treat the graph structure as known or static information, but rather constructs a graph topological prior using the spatial locations of nodes, adaptively inferring the learnable graph structure. After grouped linear transformation, the node features are combined with the graph structure to perform graph convolution operations to generate updated node representations.

[0046] Dependencies between nodes are difficult to estimate directly and accurately in scenarios with multiple incomplete information. To address this, this invention designs a Grouped Graph Neural Network (GGNN), which combines graph structure priors and cluster consistency constraints to stabilize graph structure learning and node representation learning.

[0047] Specifically, GGNN can be divided into adaptive graph learning units and cluster-based node grouping and representation update units.

[0048] (1) The adaptive graph learning unit includes the following steps: Randomly initialize the learnable parameter matrix, and construct a graph structure based on the learnable parameter matrix; A spatial adjacency matrix is ​​constructed based on the relationships between various sensor nodes in the sensor network; Calculate the graph structure regularization term based on the graph structure and the spatial adjacency matrix; The graph structure regularization term includes a dynamic weight factor, which is obtained based on the current training round number and the total training round number.

[0049] Specifically, to characterize the potential relationships between nodes, a learnable parameter matrix is ​​introduced. And construct a graph structure , Represents the normalized exponential function, This represents the activation function. This indicates that we are looking for the transpose of E.

[0050] However, under conditions of multiple incomplete information, graph structure inference relying entirely on data-driven methods is prone to introducing noisy edges. Based on the aforementioned considerations, this invention incorporates the spatial adjacency matrix as a graph topology prior into the graph structure learning process and sets dynamic weighting factors. See the following formula: ; ; in, Indicates the current training round number. Indicates the total number of training rounds. Represents the spatial adjacency matrix. To find the norm, This represents a graph structure regularization term. By constructing a graph structure regularization term, the search space of the graph structure is limited, preventing the introduction of noisy edges when the sample size is limited and data is missing. The model is dynamically adjusted during the training process, enabling it to fully utilize the prior graph structure to achieve stable optimization in the early stages of training, while gradually enhancing its data-driven adaptive graph learning capabilities in the later stages of training.

[0051] (2) The cluster-based node grouping and representation update unit includes the following steps: Based on the time embedding representation, observation data matrix, and missing indicator matrix, the reconstructed complete data is calculated, and the complete data is iteratively updated to obtain the updated complete data. Clustering algorithms (such as k-means) are used to perform cluster analysis on the updated complete data, and the cluster group sets are output. For each cluster in the clustering group set, node embedding is performed to obtain a feature matrix, and a linear transformation parameter is assigned to each feature matrix to obtain a cluster-level feature matrix; The cluster-level feature matrix is ​​mapped to the global node space by the index recovery operator, and an updated node embedding representation matrix is ​​generated by performing graph convolution operation. The index recovery operator is used to represent the correspondence between the global node space and the local index space within the cluster.

[0052] To enhance the model's ability to characterize node heterogeneity and reduce the risk of overfitting under limited sample size, this invention constructs a parameter sharing mechanism based on clustering, as detailed below: The updated complete data can be obtained using the following formula. : ; in, This represents the temporal embedding representation or the complete data obtained from the previous iteration. A fixed number of iterations can be set, and the iteration calculation is performed again for each new training round to obtain the updated complete data.

[0053] The k-means algorithm is used in the original data space containing the observed data matrix to obtain the cluster group set. See the following formula: ; in, This refers to the k-means algorithm, which clusters groups of data. The Middle The set of nodes corresponding to each cluster is represented as follows: , This represents the total number of clusters. It is a feature matrix composed of the embeddings of nodes within each cluster. Assigning linear transformation parameters The cluster-level feature matrix is ​​obtained. See the following formula: ; Furthermore, the index recovery operator is used. Cluster-level feature matrix Mapping back to the global node space ensures the consistency of node indices and guarantees the correctness of subsequent graph convolutions.

[0054] The index recovery operator is a 0-1 sparse matrix used to establish a correspondence between the local index space (the intra-cluster arrangement of nodes) and the global index space (the positional arrangement of all sensor nodes). When the global index is... The nodes are The first in When there are nodes, Otherwise, it is 0. See the following formula: ; The matrix has exactly one 1 in each column and at most one 1 in each row, ensuring the uniqueness and mutual exclusion of the node index mapping.

[0055] Through matrix multiplication Cluster-level feature matrix Upgraded to A global node space of dimension, such that the nodes belonging to Each node receives an updated representation at its corresponding location in the global index space, while other locations are zero. Summing over all clusters recovers the complete and indexed node representation matrix of the global index space, providing aligned input for subsequent graph convolutions.

[0056] Specifically, a graph convolution operation is performed to obtain the updated node embedding representation matrix. , i.e., spatiotemporal feature representation, see the following formula: ; in, express +I degree matrix, Represents a unit diagonal matrix. This represents the activation function.

[0057] To overcome the ill-posedness of node representation and the instability of the optimization process under multiple incomplete information conditions, this invention introduces a clustering consistency constraint, as shown in the following formula: ; in, This represents the cluster corresponding to node i. This represents a clustering consistency constraint. This indicates the number of nodes in the corresponding node set. Let c represent the updated cluster centers, and c represent the set of updated cluster centers. This represents the embedding representation value corresponding to the j-th sensor node in the updated node embedding representation matrix. Specifically, it is continuously updated using the following formula, and the updated cluster centers are used to constrain the updating of node representations in the next round of training: ; in, This represents the embedding representation value corresponding to the i-th sensor node in the updated node embedding representation matrix.

[0058] Referring to the formula above, the cluster centers obtained in the previous round... It is used to constrain the optimization process of the current round of node representation, that is, by narrowing the distance between node representations within the same cluster, while increasing the representational differences between node representations in different clusters, the discriminativeness of node distribution in the representation space of different clusters is improved.

[0059] Through the above process, a structural relationship is established between the original data space and the representation space, as well as between adjacent training rounds, forming a consistency constraint system that spans space and stages.

[0060] During the node representation update process, a clustering consistency constraint is introduced to avoid overfitting the model with limited information.

[0061] This invention uses the following formula as the overall optimization objective function: ; in, Indicates reconstruction error. Represents clustering constraints. Represents the prior regularity of graph structure. and All represent weighting factors. First, the magnitudes of each loss term in the early stages of training are statistically analyzed, and the weighting factors are adjusted accordingly to keep them at similar magnitudes after weighting. Preliminary experiments have verified that this can be set to... , , Represents network parameters, This indicates the portion of the observation data that was actually observed, where missing values ​​are set to 0 or ignored.

[0062] ; Furthermore, in scenarios with multiple incomplete information, simultaneously optimizing the graph structure and network parameters can lead to instability in the optimization process. To address this issue, this invention designs an alternating optimization strategy that separates graph structure learning from parameter optimization.

[0063] This strategy improves training stability and avoids initial noise amplification by alternately performing graph structure optimization with fixed network parameters and network weight updates with fixed graph structure. Ultimately, the obtained representation can be used to reconstruct the original complete data and achieve node classification.

[0064] Specifically, a two-stage alternating optimization strategy is used to update the parameters and graph structure of the sequence-to-sequence module and the grouped graph neural network module until convergence. The two-stage alternating optimization strategy includes alternating the execution of graph structure optimization with fixed parameters for the sequence-to-sequence module and the grouped graph neural network module, and parameter updates for the sequence-to-sequence module and the grouped graph neural network module with fixed graph structure.

[0065] The two-stage alternating optimization strategy specifically includes: Phase 1: Fix the parameters of the sequence-to-sequence module and the grouped graph neural network module, and optimize the graph structure with the goal of minimizing reconstruction error, cluster consistency loss and graph structure regularization term; Second stage: Fix the graph structure and update the parameters of the sequence-to-sequence module and the grouping graph neural network module with the goal of minimizing reconstruction error and cluster consistency loss; Repeat the first and second phases until the overall optimization objective function of the sequence-to-sequence module and the grouped graph neural network module converges.

[0066] Figure 3 This demonstrates the overall graph structure learning process, where the graph structure is optimized using the following formula as the objective function: ; To control the sparsity of the adjacency matrix and avoid introducing noisy edges, a predefined threshold is introduced, and edges with values ​​below the predefined threshold are removed.

[0067] Since the threshold truncation operation is not differentiable, a straight-through estimator (STE) is used to approximate the gradient during backpropagation to ensure that the gradient can be effectively backpropagated to the graph structure, thereby achieving end-to-end training.

[0068] Next, the graph structure is fixed, and the model parameters are optimized using the following formula as the objective function.

[0069] ; By iterating alternately between structure learning and parameter updates, the model can gradually optimize the graph structure while avoiding instability in representation learning caused by drastic structural changes.

[0070] This invention provides a specific embodiment, as detailed below: Referring to Table 1, the following two datasets are first described: 1) The Intel Berkeley Research Lab dataset (Intel dataset for short) originates from an indoor wireless sensor network and is used to monitor environmental conditions. Temperature data from 38 nodes between February 28th and March 21st of a certain year were selected, resampled at 30-minute intervals, and the window mean was used as the observed value. 2) The China Surface Basic Meteorological Observation Dataset (TJ Dataset) was provided by 301 meteorological stations in a certain region and includes various meteorological elements such as temperature, air pressure, and humidity. This invention uses hourly temperature data from the entire year for experiments.

[0071] Table 1. Node classification accuracy of RoG-STGNN under the conditions of dataset, annotation ratio, and missing rate.

[0072] To verify the reconstruction performance under different missing rates, a comparative experiment was conducted with the proposed RoG-STGNN algorithm, which included DCRNN, AGCRN, TTS-IMP, GRU-D, HD-TTS, and HPoolGCL algorithms. The proposed algorithm used a fixed input sequence length of 36 time steps and evaluated the reconstruction accuracy of each algorithm under different missing rates (5%, 15%, 20%, and 35%). The RoG-STGNN algorithm employed the ADAM optimizer, with 200 training epochs, an initial learning rate of 0.0125, and a decay of 0.9 times every 20 epochs. Considering that each loss term has a different numerical scale, the proposed algorithm first statistically analyzed the magnitude of each loss term in the initial training phase and adjusted the weights accordingly to maintain a similar order of magnitude after weighting. All experiments were performed on a computing platform equipped with a 12th Gen Intel(R) Core(TM) i7-12700K processor.

[0073] See Figures 6 to 9 Smaller MSE and MAE values ​​indicate higher prediction accuracy. Clearly, the RoG-STGNN algorithm of this invention achieves lower prediction errors across all missing percentages, significantly outperforming other existing mainstream spatiotemporal time series models. While the reconstruction errors of all models increase with the missing percentage, the RoG-STGNN algorithm exhibits a smaller error increase, demonstrating stronger robustness in high-missing-rate scenarios, especially in... Figure 7 Although the MAE value is not the lowest, its growth rate with increasing missing rate is the smallest; it maintains optimal performance across different types of sensor time series data, verifying the advantages of the algorithm of this invention over existing algorithms in spatiotemporal missing data filling tasks.

[0074] In another embodiment, to verify that the representation learned by the RoG-STGNN algorithm is a stable decisional representation rather than overfitting noise, this invention designed a repeatability experiment: With a fixed sequence length of 36, the algorithm of this invention was run twice under different missing rates to obtain two sets of feature representations (denoted as z_missing_rate_1 and z_missing_rate_2). The latent feature similarity between the two was calculated using centered kernel alignment (CKA). The closer the CKA value is to 1, the higher the consistency between the two representations. Figure 5 shows the CKA similarity matrices on the TJ dataset and the Intel dataset. Obviously, on two completely different real datasets, the latent representations obtained by two independent trainings under the same missing rate have a CKA similarity close to 1, proving that the algorithm is not affected by random initialization or training noise, and can learn consistent spatiotemporal features in each training, without the problem of drastic fluctuations in training results. Moreover, as the proportion of missing data continues to increase, the cross-missing-rate feature similarity still maintains a high value, indicating that even if a large amount of observed data is lost, the algorithm of this invention can still capture a unified and unchanging underlying spatiotemporal pattern, and will not learn messy and meaningless noise features due to severe missing data.

[0075] In summary, the RoG-STGNN algorithm of this invention not only has lower prediction error, but also a more stable and robust internal feature extraction mechanism.

[0076] Based on the above technical solution, this invention constructs a sequence-to-sequence module to characterize multi-scale temporal correlations under the condition of missing inputs; subsequently, based on the temporal embedding representation, a grouped graph neural network is introduced to establish spatial dependencies between nodes.

[0077] This invention uses graph topology priors to limit the search space of graph structures and improves the structural stability of the embedding space through cluster consistency constraints. At the same time, it designs an alternating optimization strategy for graph structures and network parameters to avoid the amplification of noise correlations during the iteration process, thereby improving the overall training stability of the model in scenarios with multiple incomplete information.

[0078] This invention represents wireless sensor network observation data as a unified spatiotemporal graph. By treating the model architecture as an implicit structural prior that characterizes the inherent spatiotemporal correlation of the data, and embedding the parameter optimization process into the inference stage, the model can gradually converge to obtain a stable representation without relying on a large number of training samples through single-sample iterative optimization.

[0079] This invention also provides an unsupervised spatiotemporal feature learning system 400 for sensor networks in incomplete scenarios, see [link to related document]. Figure 4 ,include: The data acquisition module 410 is used to acquire an observation data matrix with missing values ​​and a corresponding missing indicator matrix based on the constructed sensor network. The embedding representation module 420, connected to the data acquisition module 410, is used to input the observation data matrix with missing values ​​and the missing indicator matrix into the sequence to extract the multi-scale temporal correlation representation and generate the temporal embedding representation of the node. The clustering representation module 430, connected to the embedding representation module 420, is used to input the temporal embedding representation into the grouped graph neural network module. Combining graph structure prior and adaptive graph learning mechanism, it establishes the spatial dependency relationship between nodes in the sensor network, and updates the grouped representation of nodes in the sensor network through clustering consistency constraints, and outputs spatiotemporal feature representation.

[0080] For other details, please refer to the methods described above; they will not be repeated here.

[0081] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0082] The acquisition, storage, and application of user personal information involved in the technical solution of this invention all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0083] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0084] Figure 5A schematic block diagram of an electronic device 500 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0085] Electronic device 500 includes a computing unit 501, which can perform various appropriate actions and processes according to a computer program stored in ROM 502 or a computer program loaded into RAM 503 from storage unit 508. RAM 503 can also store various programs and data required for the operation of electronic device 500. The computing unit 501, ROM 502, and RAM 503 are interconnected via bus 504. I / O interface 505 is also connected to bus 504.

[0086] Multiple components in electronic device 500 are connected to I / O interface 505, including: input unit 506, such as keyboard, mouse, etc.; output unit 507, such as various types of monitors, speakers, etc.; storage unit 508, such as disk, optical disk, etc.; and communication unit 509, such as network card, modem, wireless transceiver, etc. Communication unit 509 allows electronic device 500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0087] The computing unit 501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above. For example, in some embodiments, the methods may be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 500 via ROM 502 and / or communication unit 509. When the computer program is loaded into RAM 503 and executed by the computing unit 501, one or more steps of the unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios described above may be performed. Alternatively, in other embodiments, computing unit 501 may be configured, by any other suitable means (e.g., by means of firmware), to perform an unsupervised spatiotemporal feature learning method for the sensor network in the incomplete scenario.

[0088] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0089] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A sensor network unsupervised spatio-temporal feature learning method for incomplete scenarios, the incomplete scenarios including scenarios with random missing values in sensor network observation data and scenarios with limited effective training sample size, characterized in that, include: Based on the constructed sensor network, an observation data matrix with missing values ​​and the corresponding missing value indication matrix are obtained; The observation data matrix with missing values ​​and the missing indicator matrix are input into the sequence module to extract multi-scale temporal correlation representations and generate temporal embedding representations of nodes. The time-embedded representation is input into the grouped graph neural network module. Combining graph structure prior and adaptive graph learning mechanism, spatial dependencies between nodes in the sensor network are established. The grouped representation of nodes in the sensor network is updated by clustering consistency constraints, and spatiotemporal feature representation is output. Specifically, the sequence-to-sequence module and the grouped graph neural network module are trained using the effective training samples.

2. The unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios according to claim 1, characterized in that, The observation data matrix with missing values ​​and the missing indicator matrix are input into the sequence module to extract multi-scale temporal correlation representations and generate temporal embedding representations of nodes. Specifically, this includes the following steps: Local convolutional units are used to preprocess the observation data matrix and the missing indicator matrix with missing values ​​to obtain the updated observation data matrix and the updated missing indicator matrix. The updated observation data matrix and the updated missing indicator matrix are input into the time-domain feature extractor to output time-domain features, and into the frequency-domain feature extractor to output frequency-domain features. The time-domain features and frequency-domain features are fused through concatenation, linear mapping, and combination operations to generate a time-embedded representation; The time-domain feature extractor is used to capture local dynamic changes and short-term dependencies using multi-layer one-dimensional convolution; the frequency-domain feature extractor is used to map the updated observation data matrix to the frequency domain through discrete Fourier transform, and to perform nonlinear transformation on the real and imaginary parts of the mapped spectrum using gated linear units.

3. The unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios according to claim 1, characterized in that, By combining graph structure priors and adaptive graph learning mechanisms, spatial dependencies between nodes in a sensor network are established, specifically including: Randomly initialize the learnable parameter matrix, and construct a graph structure based on the learnable parameter matrix; A spatial adjacency matrix is ​​constructed based on the relationships between various sensor nodes in the sensor network; Calculate the graph structure regularization term based on the graph structure and the spatial adjacency matrix; The graph structure regularization term includes a dynamic weight factor, which is obtained based on the current training round number and the total training round number.

4. The unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios according to claim 3, characterized in that, The node representation in the sensor network is updated by grouping nodes according to cluster consistency constraints, and the updated node embedding representation matrix is ​​output. The specific steps include: Based on the time embedding representation, observation data matrix, and missing indicator matrix, the reconstructed complete data is calculated, and the complete data is iteratively updated to obtain the updated complete data. Clustering algorithms are used to perform cluster analysis on the updated complete data, and cluster group sets are output. Node embedding is performed on each cluster in the cluster group sets to obtain a feature matrix. Linear transformation parameters are assigned to each feature matrix to obtain a cluster-level feature matrix. The cluster-level feature matrix is ​​mapped to the global node space by the index recovery operator, and an updated node embedding representation matrix is ​​generated by performing graph convolution operation. The index recovery operator is used to represent the correspondence between the global node space and the local index space within the cluster.

5. The unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios according to claim 4, characterized in that, The clustering consistency constraint also includes: The cluster centers from the previous training round are used to constrain the grouping representation update process in the current round, so as to reduce the distance between node representations within the same cluster and increase the difference between node representations in different clusters. The cluster centers are updated using the following formula: ； in, Indicates the updated cluster centers. This represents the set of nodes corresponding to the k-th cluster. This indicates the number of nodes in the corresponding node set. This represents the embedding representation value corresponding to the i-th sensor node in the node embedding representation matrix.

6. The unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios according to claim 5, characterized in that, Also includes: A two-stage alternating optimization strategy is used to update the parameters and graph structure of the sequence-to-sequence module and the grouped graph neural network module until convergence. The two-stage alternating optimization strategy includes alternating the execution of graph structure optimization with fixed parameters for the sequence-to-sequence module and the grouped graph neural network module, and parameter updates for the sequence-to-sequence module and the grouped graph neural network module with fixed graph structure.

7. The unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios according to claim 6, characterized in that, The two-stage alternating optimization strategy specifically includes: Phase 1: Fix the parameters of the sequence-to-sequence module and the grouped graph neural network module, and optimize the graph structure with the goal of minimizing reconstruction error, cluster consistency loss and graph structure regularization term; Second stage: Fix the graph structure and update the parameters of the sequence-to-sequence module and the grouping graph neural network module with the goal of minimizing reconstruction error and cluster consistency loss; Repeat the first and second phases until the overall optimization objective function of the sequence-to-sequence module and the grouped graph neural network module converges.

8. An unsupervised spatiotemporal feature learning system for sensor networks in incomplete scenarios, used to implement the unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to acquire the observation data matrix with missing values ​​and the corresponding missing indicator matrix based on the constructed sensor network. An embedded representation module, connected to the data acquisition module, is used to input the observation data matrix with missing values ​​and the missing indicator matrix into the sequence to extract multi-scale temporal correlation representations and generate temporal embedded representations of nodes. The clustering representation module, connected to the embedding representation module, is used to input the temporal embedding representation into the grouped graph neural network module. Combining graph structure priors and adaptive graph learning mechanisms, it establishes spatial dependencies between nodes in the sensor network and updates the grouped representations of nodes in the sensor network through clustering consistency constraints, outputting spatiotemporal feature representations.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement an unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program; the computer program is executed by a processor to implement an unsupervised spatiotemporal feature learning method for sensor networks in incomplete scenarios as described in any one of claims 1 to 7.

Images