High-precision clustering method for data fusion in wireless sensor network

Abstract

The high-precision clustering method for data fusion in a wireless sensor network comprises the following steps: step 1, establishing a high-data-similarity clustering algorithm optimization model and constructing a multi-objective optimization function; step 2, performing preliminary clustering; step 3, calculating the membership of a sensor node to multiple clusters by using a fuzzy C-means method and determining the final attribution of the node; step 4, selecting a cluster head node; step 5, dynamically adjusting instead of re-clustering according to the network condition; step 6, when only a few nodes need to adjust the attribution due to monitoring data changes, modeling the resource competition between clusters as a non-cooperative game and deciding the attribution of the node according to a utility function; step 7, when a local region needs to be adjusted, only the part that needs to be adjusted is locally re-clustered; and step 8, when the network monitoring data changes greatly, step 1 is re-executed to restore the optimal clustering state of the network. The present application significantly reduces the network communication energy consumption and prolongs the network life cycle.

Description

Technical Field

[0001] This invention belongs to the field of data fusion technology in wireless sensor networks, and specifically relates to a high-precision clustering method for data fusion in wireless sensor networks. Background Technology

[0002] For wireless sensor networks, extending network lifetime is a core issue due to the limited energy and difficulty in replacing nodes. Clustering technology, by dividing network nodes into several clusters with cluster heads responsible for data aggregation and forwarding, can effectively reduce communication energy consumption. Data fusion technology, by aggregating data within a cluster at the cluster head, eliminates redundant information and reduces the amount of data transmitted. The core premise of data fusion is that the data of nodes within a cluster is highly similar. Only by organizing nodes with similar data characteristics into the same cluster can data fusion eliminate redundant information to the maximum extent. However, in practical applications, the data distribution within the monitoring area often exhibits complex spatial heterogeneity. Traditional clustering algorithms typically cluster based on the geographical location of nodes, implicitly assuming that spatially adjacent nodes have similar data, but this assumption often fails in multi-parameter monitoring scenarios.

[0003] Pan Yuheng et al. proposed a clustering routing algorithm based on measurement data similarity in "A WSN Clustering Routing Method Based on Measurement Data Similarity" (Journal of Sensor Technology, 2023, 36(3): 481-489). This algorithm utilizes historical measurement data of nodes and measures data similarity by calculating the root mean square error (RMSE) of measurement data between nodes. An adaptive clustering algorithm is designed to group nodes with high measurement similarity into a cluster. Within each cluster, one node is selected to transmit data, while other nodes remain dormant to conserve energy. This method uses data similarity to guide cluster partitioning during the clustering stage, thus initially realizing the clustering idea based on data features. However, although this method proposes a clustering idea based on data correlation, it uses a simple RMSE as a similarity measure, only performing a simple sum of squares calculation on each dimension of data. This fails to address the scale difference problem of multidimensional data and does not consider the correlation between dimensions. When the dimensions and numerical ranges of data in different dimensions differ significantly, or when multiple dimensions are strongly correlated, the dimension with the larger numerical range may dominate the similarity calculation results, or excessive information redundancy may lead to measurement distortion, affecting the accuracy of similarity judgment. Most existing technologies cluster based on geographical location, implicitly assuming that spatially adjacent nodes have similar data. Furthermore, some clustering methods based on data similarity only address single-dimensional data or use simple Euclidean distance metrics, failing to address the scale differences and inter-dimensional correlations in multi-dimensional data. Existing methods employ a cluster-then-fusion approach, treating clustering and fusion as independent processes, failing to fully utilize statistical characteristics to guide cluster partitioning, resulting in a disconnect between clustering results and data fusion requirements. In addition, existing methods lack effective mechanisms to handle dynamic network changes; periodic global re-clustering is costly, and simple local adjustments are insufficient to maintain data similarity, leading to high network energy consumption and short lifespans. Summary of the Invention

[0004] To overcome the shortcomings of the existing technology, the present invention aims to provide a high-precision clustering method for data fusion in wireless sensor networks. This method solves the problems of high heterogeneity of data within clusters, low data fusion accuracy, and high communication energy consumption caused by geographical location-based clustering when clustering wireless sensor networks. Under the premise of ensuring high fusion accuracy, it maximizes the elimination of redundant data, automatically distinguishes and locates different data ranges, significantly reduces network communication energy consumption, and extends the network life cycle.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A high-precision clustering method for data fusion in wireless sensor networks includes the following steps; Step 1: Establish a high data similarity clustering algorithm optimization model and construct a multi-objective optimization function; Step 2: Based on the distributed environment, a high similarity clustering algorithm is used for initial clustering; Step 3: Boundary node processing based on fuzzy clustering: The clustering results from Step 2 are adjusted again, and the membership degree of the wireless sensor node to multiple clusters is calculated using the fuzzy C-means method to determine the final affiliation of the node. Step 4: Energy-aware adaptive cluster head election strategy: After the clusters are determined, the cluster head node is selected by comprehensively considering factors such as the remaining energy of the nodes, communication cost, and centrality. Step 5: Adaptive clustering management mechanism: Real-time monitoring of node status within a cluster, and dynamic adjustment based on network conditions rather than re-clustering; Step 6: Game-based local adjustment strategy: When the proportion of nodes whose affiliation needs to be adjusted due to changes in monitoring data is lower than the preset threshold, ranging from 10% to 20%, these nodes will model inter-cluster resource competition as a non-cooperative game. The node and the cluster heads of other clusters that may accept the node will decide the affiliation of the node based on the utility function. Step 7: Local reclustering strategy: When a local region needs to be adjusted, only the part that needs to be adjusted is reclustered locally; Step 8: Network-wide Reconstruction Trigger Mechanism: When network monitoring data changes on a large scale, the network topology is restructured globally, Step 1 is re-executed, and the optimal clustering state of the network is restored. Step 1 defines a two-dimensional planar monitoring area of ​​area M×M, in which there are N randomly deployed sensor nodes, each node collecting K-dimensional data vectors; the N nodes are divided into L clusters, a high data similarity clustering algorithm optimization model is established, and a multi-objective optimization function is constructed. The specific steps are as follows: Step 1.1, Define variables: Define a clustering matrix A ∈ {0, 1} (N×L) and cluster head matrix B∈{0,1} (N×L) The clustering scheme is characterized in that each node belongs to exactly one cluster, and each cluster has exactly one cluster head. Where a ij The elements in matrix A represent the partitioning relationship between nodes and clusters; b ij The elements in matrix B represent the partitioning relationship between cluster head nodes and ordinary nodes; Step 1.2, Define performance metrics: (1): To accurately measure the similarity of node data in high-dimensional space, Mahalanobis distance is used to measure the similarity of node data. i Data x i To node s j Data x j The Mahalanobis distance is: Where Σ is the global covariance matrix, which can automatically eliminate differences in dimensions and take into account the correlation between dimensions; (2): To accurately measure the communication distance between nodes, node s i Position (posx) i ,posy i ) to node s j Position (posx) j ,posy j The communication distance is: Specifically, if it is cluster C j Any member node s i The communication distance to the cluster head is defined as: Where (posx) CHj ,posy CHj ) is cluster C j cluster head s CHj Location; (3): Define the sum of squared errors within a cluster based on Mahalanobis distance to directly measure the data dispersion of the cluster and achieve tight aggregation of data dimensions: v j It is the weighted average of the data vectors of all nodes within the cluster; (4): Define the sum of energy consumption as the energy consumption of the entire communication process from the node within the cluster to the cluster head and from the cluster head to the base station, so as to minimize the communication energy. Minimize energy consumption as much as possible in the distributed clustering and cluster head selection stages: Where E Tx (l, d) represents the total energy consumed by the sensor node in sending a data packet of length l bits to a receiver at a distance d, E DA Energy consumption for processing 1 bit of data fusion, C j l is the sum of the data received by the cluster head within the cluster. f It is the merged data; Step 1.3: Define the objective function and constraints. The objective function comprehensively considers both data-dimensional and energy-dimensional objectives. The data-dimensional objective is to minimize the sum of squared Mahalanobis distances within the cluster, ensuring tight clustering of data within the cluster. The energy-dimensional objective is to minimize the energy consumption of communication within the cluster and the transmission energy consumption from the cluster head to the base station. Data accuracy and communication efficiency are balanced through weighting coefficients. .

[0006] Furthermore, the constraints are specifically as follows: (1): Completeness perspective: Every node in the network must be assigned to a cluster, and each node can only belong to one cluster; (2): Non-empty cluster constraint: Each cluster must contain at least one node, and empty clusters cannot exist; ; (3): Cluster head uniqueness constraint: Each cluster has one and only one cluster head; (4): Cluster head affiliation constraint: The cluster head must belong to the cluster; ; (5): Spatial proximity constraint: To prevent nodes that are too far apart from each other from being assigned to the same cluster, thereby avoiding communication interruption or excessive energy consumption within the cluster and ensuring the reachability of communication within the cluster; .

[0007] Step 2 specifically involves: Step 2.1, First stage coarse screening: Dimensionality reduction and broadcasting of K-dimensional data, local standardization of data at each node; setting a threshold, adding nodes with a distance less than the threshold to the candidate set; Where z i，k It is node s i Normalization on the Kth dimension, μ k It is node s i The average value of the Kth dimension of all neighbor node data within the communication range, σ k It is node s i The variance of the Kth dimension of the data of all neighboring nodes within the communication range; Step 2.2, the second stage of fine screening; using feature signature technology, the standardized data vectors are normalized into unit vectors, satisfying... Then, multiple random hyperplanes are used for projection, and a binary signature Sig(F) is generated based on the projection results. i ); Calculate the Hamming distance; if the Hamming distance is less than the threshold, then pass the screening. ; The hash function hl(F) i Defined as: ; Step 2.3, Distributed Clustering Process: Each node collects K-dimensional data, calculates feature signatures, and broadcasts them to neighboring nodes; performs double screening to form a set of similar nodes; calculates the cluster head competition weight, and the node with the highest weight is declared as the cluster head; ordinary nodes select the most similar cluster head to join, completing the initial clustering.

[0008] Step 3 specifically involves: Step 3.1, Preliminary membership calculation; taking into account three factors, namely data similarity, communication cost and load balancing, calculate the membership degree of each node to each cluster; , Where S is the data similarity, C is the communication cost, and B is the load balancing coefficient; Step 3.2, Boundary node identification; calculate the membership difference, i.e., the difference between the largest and second largest membership. If the difference is less than the threshold, it indicates that the node belongs to multiple clusters to a similar degree and is identified as a boundary node. Step 3.3, Fuzzy Clustering Optimization of Boundary Nodes; Fuzzy clustering optimization is performed only on the set of boundary nodes, taking into account both data spatial distance and geographic spatial distance; in The square of the normalized Euclidean distance between node i and the data center of cluster x reflects the data similarity; The communication cost (based on transmission distance) between node i and the data center of cluster x represents the communication cost between node i and cluster x. Update the membership degree and iterate until convergence; select the cluster with the largest membership degree as the final destination to complete the precise assignment of boundary nodes.

[0009] Step 4 specifically involves: Step 4.1, Multifactor Evaluation Scheme: Where f data To indicate the representativeness of the data, membership degree is used as a measure; f comm Location centrality is represented by the number of neighboring nodes and the average distance; f ener Measured by the ratio of the current energy to the cluster average energy; f BS It indicates the proximity to the base station, measured by the reciprocal of the distance to the base station; Step 4.2, Calculate the overall weight; perform a weighted summation of the four factors to obtain the overall weight; Step 4.3, Distributed Cluster Head Election: Nodes set random backoff timers, with the timer duration inversely proportional to the overall weight; the node whose timer expires first broadcasts a cluster head announcement, and other nodes cancel their timers upon receiving the announcement, confirm the cluster head, and join; this mechanism does not require centralized control and can make autonomous decisions using only local information.

[0010] Step 5 specifically involves: To address network state changes in dynamic scenarios, an adaptive clustering management mechanism is proposed. Nodes are classified into hot nodes, adjacent cluster heads, and unaffected nodes. Hot nodes are those that trigger adjustment events, adjacent cluster heads are those adjacent to the cluster containing the hot node, and unaffected nodes are those unaffected by the adjustment. The adjustment scope is limited to a local area to avoid network-wide information exchange. Hot nodes report their own status to the cluster head and wait for the cluster head's next instruction.

[0011] Step 6 specifically involves: Step 6.1: Construct a candidate cluster head set. The candidate cluster heads of hot zone nodes are selected from adjacent cluster heads, and must meet the conditions of geographical distance, remaining energy and load. Step 6.2: When deciding whether to accept a new node, the cluster head needs to weigh the benefits of data similarity, the incremental communication cost, and the incremental load cost; the cluster head benefit function U CH Defined as: Where S is the data similarity benefit, C is the communication cost increment, and P is the load cost increment; Step 6.3, Distributed negotiation mechanism: The hot zone node sends a migration request to the current cluster head, and the cluster head forwards the request to the neighboring cluster heads; the neighboring cluster heads calculate the benefit Un and send an acceptance permission or rejection; the node selects the cluster head that maximizes the benefit and completes the migration; Where E represents energy status gain, C represents communication quality gain, and Q represents load stability gain; Step 6.4, Incremental Maintenance Algorithm: Quickly update the data center, standard deviation vector, and feature signature using an online update formula; when a node joins or leaves the cluster, there is no need to recalculate the statistics of all nodes, only need to update them according to the incremental formula.

[0012] Step 7 specifically involves: Step 7.1, Trigger condition judgment; when the number of hot zone nodes exceeds the preset threshold of cluster size, ranging from 20% to 40%, local re-clustering is adopted; this threshold is dynamically adjusted according to the network status; Step 7.2, Local Re-clustering Process: The cluster head sends a re-clustering message to the hot zone nodes. Only nodes that receive this message will participate in the subsequent local re-clustering activities. For nodes that do not receive the re-clustering message, even if they receive re-clustering related messages from the hot zone nodes in the subsequent process, they will directly ignore the message. Step 7.3: Calculate the feature signature of the hot zone node that received the re-clustering message, and perform similarity judgment based on the first stage coarse screening and the second stage fine screening in Step 2; use the clustering algorithm of Step 2 and Step 3 to re-cluster only for hot zone areas; Step 7.4: After the cluster is formed, the cluster head election is carried out in step 4. The new cluster head registers with the base station and establishes inter-cluster routing.

[0013] Step 8 specifically involves: Step 8.1, Calculate network health indicators; comprehensively consider topology change indicators and inter-cluster distance indicators; The topology change index M reflects the degree of change in network topology, while the inter-cluster distance index Dis reflects the degree of deviation between the current cluster structure and the initial optimal cluster structure. Step 8.2, full network reconstruction trigger judgment; when the network health is lower than the threshold, it means that the network structure has seriously deviated from the optimal state, and the base station broadcasts a full network reconstruction instruction; all nodes clear their current cluster affiliation, execute the complete clustering algorithm from step 1 to step 4, and rebuild the cluster structure.

[0014] The beneficial effects of this invention are: First, it significantly improves the accuracy of data fusion across the entire network. Through the high data similarity clustering algorithm optimization model established in step 1, using data feature similarity as the primary clustering basis, a multi-objective optimization function is constructed to maximize intra-cluster similarity and inter-cluster differences. In step 2, a hierarchical clustering method is used to ensure high homogeneity of data within clusters. Furthermore, in step 3, boundary node processing based on fuzzy clustering utilizes the FCM algorithm to calculate the membership degree of boundary nodes to each cluster, avoiding the mixing of heterogeneous data that leads to fusion errors, further improving the similarity of data between clusters. Compared to traditional geographic location-based clustering methods that ignore data correlation leading to high intra-cluster data heterogeneity, the data-driven clustering strategy of this invention allows for high accuracy through simple weighted averaging during data fusion, effectively reducing data volume and fusion errors while maintaining accuracy.

[0015] Second, extend network lifetime. Step 4 uses an energy-aware adaptive cluster head election strategy to select the cluster head with the lowest overall energy consumption, preventing low-energy nodes from prematurely dying. Step 5 uses an adaptive clustering management mechanism to monitor node status in real time. When only a few nodes need adjustment, a game-theoretic local adjustment strategy (step 6) is adopted. When nodes are clustered together in hotspots, a local re-clustering strategy (step 7) is used, re-clustering only hotspots to avoid additional energy consumption. When network-wide monitoring data changes significantly and performance degrades severely, step 8 triggers a full network reconstruction to restore the optimal clustering state and maintain high data transmission quality. This rapid adaptive clustering management technology avoids frequent network-wide information interaction, significantly reducing communication energy consumption.

[0016] Third, the distributed implementation is user-friendly. The algorithm design fully considers the resource-constrained sensor nodes, enabling autonomous decision-making using only local information and one-hop neighborhood information, without centralized control. In the high-similarity clustering method in step 2, nodes only need to exchange data feature vectors with their one-hop neighbors and determine candidate clusters by calculating the similarity matrix locally, without the need for network-wide information aggregation. The management mechanism in step 5 adopts distributed threshold judgment, with each node independently monitoring whether its own data has shifted, while the data center, standard deviation vector, and feature signature within the cluster are periodically broadcast by the cluster head, actively requesting cluster adjustment when data shifts. The game-based local adjustment strategy in step 6 models inter-cluster resource competition as a non-cooperative game, with nodes autonomously deciding whether to switch clusters based on their local utility function, without centralized calculation and control by the base station. The local re-clustering strategy in step 7 only reorganizes the clusters of the detected abnormal local areas and their neighboring clusters, while the cluster structure of unaffected areas remains unchanged, avoiding network-wide information interaction. The network-wide reconstruction triggering mechanism in step 8 only initiates network-wide reconstruction when multiple areas simultaneously report severe degradation signals and reach a preset threshold, ensuring the necessity and timeliness of reconstruction decisions. Compared to centralized algorithms that require aggregating the entire network topology information to the base station, resulting in communication overhead and the risk of single point of failure, the distributed design of this invention allows each node to communicate only with neighboring nodes and cluster heads, greatly reducing decision delay, and ensuring that local network failures do not affect the operation of other areas. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings.

[0019] like Figure 1 As shown, a high-precision clustering method for data fusion in wireless sensor networks includes the following steps; Step 1, Establishment of high data similarity clustering algorithm optimization model: Within a two-dimensional planar monitoring area of ​​area M×M, there are N randomly deployed sensor nodes, each collecting a K-dimensional data vector. We divide the N nodes into L clusters and establish a multi-objective optimization model, comprehensively considering both the data and energy dimensions. The specific steps include: Step 1.1, Define variables: Define a clustering matrix A ∈ {0, 1} (N×L) and cluster head matrix B∈{0,1} (N×L) The clustering scheme represents a system where each node belongs to exactly one cluster, and each cluster has exactly one cluster head.

[0020] Step 1.2, Define performance metrics: (1) To accurately measure the similarity of node data in high-dimensional space, Mahalanobis distance is theoretically used to measure the similarity of node data. i Data x i To node s j Data x j The Mahalanobis distance is: Where Σ is the global covariance matrix, it can automatically eliminate differences in dimensions and take into account the correlation between dimensions. Mahalanobis distance is an ideal choice for measuring the similarity of high-dimensional data.

[0021] (2) To accurately measure the communication distance between nodes, node s i Position (posx) i ,posy i ) to node s j Position (posx) j ,posy j The communication distance is: Specifically, if it is cluster C j Any member node s i The communication distance to the cluster head is defined as: (3) To achieve tight aggregation of data dimensions, the sum of squared errors within a cluster based on Mahalanobis distance is defined to directly measure the data dispersion of the cluster: v j It is the weighted average of the data vectors of all nodes within the cluster.

[0022] (4) To minimize communication energy, the sum of energy consumption is defined as the energy consumption of the entire communication process from the node within the cluster to the cluster head and from the cluster head to the base station: Step 1.3: Define the objective function and constraints. The objective function comprehensively considers both data-dimensional and energy-dimensional objectives. The data-dimensional objective is to minimize the sum of squared Mahalanobis distances within clusters, ensuring tight clustering of data within clusters. The energy-dimensional objective is to minimize the energy consumption of communication within clusters and the transmission energy consumption from the cluster head to the base station. Weighting coefficients are used to balance data accuracy and communication efficiency.

[0023] , Constraints: (1) Completeness perspective: Each node in the network must be assigned to a cluster, and each node can only belong to one cluster.

[0024] (2) Non-empty cluster constraint: Each cluster should contain at least one node, and empty clusters cannot exist.

[0025] , (3) Cluster head uniqueness constraint: Each cluster has one and only one cluster head.

[0026] (4) Cluster head affiliation constraint: The cluster head must belong to the cluster.

[0027] ; (5) Spatial proximity constraint: To prevent nodes that are too far apart from each other from being assigned to the same cluster, thereby avoiding communication interruption or excessive energy consumption within the cluster, and ensuring the reachability of communication within the cluster.

[0028] Step 2, High Similarity Clustering Method: To address the high computational complexity of Mahalanobis distance and the potential singularity of the covariance matrix, a dual-screening mechanism is proposed to significantly reduce computational complexity while maintaining accuracy. In step 2, a hierarchical clustering method is employed to ensure high homogeneity of data within clusters. Specifically, the following steps are included: Step 2.1, First Stage Coarse Screening. Dimensionality reduction and broadcasting are performed on the K-dimensional data. Nodes locally standardize the data. A threshold is set, and nodes with a distance less than the threshold are added to the candidate set.

[0029] Where z i，k It is node s iNormalization on the Kth dimension, μ k It is node s i The average value of the Kth dimension of all neighbor node data within the communication range, σ k It is node s i The variance of the Kth dimension of all neighbor node data within the communication range Step 2.2, Second Stage Fine Screening. Feature signature technology is used to normalize the standardized data vectors into unit vectors, satisfying... Then, multiple random hyperplanes are used for projection, and a binary signature Sig(F) is generated based on the projection results. i Calculate the Hamming distance; if the Hamming distance is less than a threshold, then pass the screening.

[0030] ; The hash function hl(F) i Defined as: ; Step 2.3, Distributed Clustering Process. Each node collects K-dimensional data, calculates a feature signature, and broadcasts it to its neighboring nodes. A double-selection process is performed to form a set of similar nodes. The cluster head competition weight is calculated, and the node with the highest weight is declared the cluster head. Ordinary nodes select the most similar cluster head to join, completing the initial clustering.

[0031] Step 3: Boundary node processing based on fuzzy clustering: By processing the boundary nodes based on fuzzy clustering in Step 3, the membership degree of the boundary nodes to each cluster is calculated using the FCM algorithm, which avoids the mixing of heterogeneous data and causes fusion errors, and further improves the data similarity between clusters.

[0032] To address the situation where boundary nodes may be similar to multiple clusters simultaneously, a fuzzy clustering method is employed for optimization. Specifically, the following steps are included: Step 3.1, Preliminary Membership Calculation. Taking into account three factors—data similarity, communication cost, and load balancing—the membership degree of each node to each cluster is calculated.

[0033] , Where S is the data similarity, C is the communication cost, and B is the load balancing coefficient.

[0034] Step 3.2, Boundary Node Identification. Calculate the membership difference, i.e., the difference between the largest and second-largest membership. If the difference is less than the threshold, it indicates that the node belongs to multiple clusters to a similar degree, and it is identified as a boundary node.

[0035] Step 3.3, Fuzzy Clustering Optimization of Boundary Nodes. Fuzzy clustering optimization is performed only on the set of boundary nodes, taking into account both data spatial distance and geographic spatial distance.

[0036] in The square of the normalized Euclidean distance between node i and the data center of cluster x reflects the data similarity; The communication cost (based on transmission distance) between node i and the data center of cluster x represents the communication cost between node i and cluster x. Update the membership degree and iterate until convergence. Select the cluster with the highest membership degree as the final destination, thus completing the precise assignment of boundary nodes.

[0037] Step 4, Energy-Aware Adaptive Cluster Head Election Strategy: To achieve energy balance and extend network lifetime, a multi-factor evaluation cluster head election strategy is proposed. This includes the following steps: Step 4.1, Multifactor Evaluation Scheme: Where f data To indicate the representativeness of the data, membership degree is used as a measure; f comm Location centrality is represented by the number of neighboring nodes and the average distance; f ener Measured by the ratio of the current energy to the cluster average energy; f BS It indicates the proximity to the base station, measured by the reciprocal of the distance to the base station.

[0038] Step 4.2, Calculate the overall weight. The four factors are weighted and summed to obtain the overall weight.

[0039] Step 4.3, Distributed Cluster Head Election. Nodes set random backoff timers, with the timer duration inversely proportional to the overall weight. The node whose timer expires first broadcasts a cluster head announcement. Other nodes, upon receiving the announcement, cancel their timers, confirm the cluster head, and join. This mechanism requires no centralized control; it can make autonomous decisions using only local information.

[0040] Step 5, Adaptive Cluster Management Mechanism: The management mechanism in Step 5 adopts a distributed threshold judgment. Each node independently monitors whether its own data has shifted, while the data center, standard deviation vector, and feature signature within the cluster are periodically broadcast by the cluster head. When data shifts, the cluster head actively requests cluster adjustment. To address network state changes in dynamic scenarios, an adaptive clustering management mechanism is proposed. Nodes are classified into hot nodes, adjacent cluster heads, and unaffected nodes. Hot nodes are those that trigger adjustment events, adjacent cluster heads are those cluster heads adjacent to the hot node's cluster, and unaffected nodes are those unaffected by the adjustment. The adjustment scope is limited to a local area, avoiding network-wide information exchange. Hot nodes report their own status to the cluster head and await further instructions from the cluster head.

[0041] Step 6, game-based local adjustment strategy: Step 6, game-based local adjustment strategy, models inter-cluster resource competition as a non-cooperative game. Nodes decide autonomously whether to switch clusters based on their local utility function, without the need for centralized calculation and control by the base station; Step 7, local re-clustering strategy, only reorganizes the local areas where anomalies are detected and their neighboring clusters, while the cluster structure of unaffected areas remains unchanged, avoiding information interaction across the entire network. To address the issue of data drift among a small number of nodes, a game-theoretic node migration strategy is proposed. This strategy includes the following steps: Step 6.1: Construct a candidate cluster head set. Candidate cluster heads for hot zone nodes are selected from adjacent cluster heads, and must satisfy geographical distance, remaining energy, and load conditions.

[0042] Step 6.2: When deciding whether to accept a new node, the cluster head needs to weigh the benefits of data similarity, the incremental communication cost, and the incremental load cost. The cluster head benefit function UCH is defined as: Where S represents the data similarity benefit, C represents the communication cost increment, and P represents the load cost increment.

[0043] Step 6.3, Distributed Negotiation Mechanism. The hot zone node sends a migration request to the current cluster head, which forwards the request to neighboring cluster heads. Neighboring cluster heads calculate the benefit Un and send an acceptance or rejection message. The node selects the cluster head that maximizes the benefit and completes the migration.

[0044] Where E represents energy status gain, C represents communication quality gain, and Q represents load stability gain.

[0045] Step 6.4, Incremental Maintenance Algorithm. The data center, standard deviation vector, and feature signature are quickly updated using an online update formula. When a node joins or leaves the cluster, there is no need to recalculate the statistics of all nodes; only updates based on the incremental formula are required.

[0046] Step 7, Local Re-clustering Strategy: For situations with a large number of hot zone nodes, a local re-clustering strategy is proposed. This includes the following steps: Step 7.1, Trigger Condition Determination. When the number of hot zone nodes exceeds a certain proportion of the cluster size, local re-clustering is adopted. This threshold is dynamically adjusted according to the network status.

[0047] Step 7.2, Local Re-clustering Process. The cluster head sends a re-clustering message to the hotspot nodes, and the hotspot nodes calculate feature signatures and perform similarity judgment. The clustering algorithms in Steps 2 and 3 are used, and re-clustering is only performed on hotspot areas. After clusters are formed, cluster head election is performed, the new cluster head registers with the base station, and inter-cluster routing is established.

[0048] Step 8, Full Network Reconstruction Trigger Mechanism: The full network reconstruction trigger mechanism in Step 8 will only initiate full network reconstruction when multiple regions simultaneously report severe degradation signals and reach a preset threshold, ensuring the necessity and timeliness of the reconstruction decision.

[0049] To address scenarios involving significant changes in network topology, a full network reconstruction triggering mechanism is proposed. This mechanism includes the following steps: Step 8.1, Calculation of network health indicators. This involves comprehensively considering topology change indicators and inter-cluster distance indicators.

[0050] The topology change index M reflects the degree of change in network topology, while the inter-cluster distance index Dis reflects the degree of deviation between the current cluster structure and the initial optimal cluster structure.

[0051] Step 8.2, Full Network Reconstruction Trigger Judgment. When the network health score falls below the threshold, it indicates that the network structure has seriously deviated from the optimal state, and the base station broadcasts a full network reconstruction command. All nodes clear their current cluster affiliation, execute the complete clustering algorithm from Steps 1 to 4, and re-establish the cluster structure.

[0052] A clustering optimization model based on intrinsic data similarity rather than geographical location aims to minimize the data fusion error across the entire network while also taking into account communication energy consumption.

[0053] Based on a dual screening mechanism of standardized Euclidean distance and feature signature, the computational complexity of Mahalanobis distance is reduced, and the singularity problem of covariance matrix is ​​solved.

[0054] The collaborative workflow of dual screening mechanism, fuzzy clustering and cluster head election strategy enables high-precision, low-energy distributed clustering.

[0055] The fast adaptive clustering management technology includes a game-theoretic local adjustment strategy, an incremental maintenance algorithm, and a network-wide reconstruction triggering mechanism.

[0056] In farmland monitoring scenarios, nodes need to monitor multiple parameters such as soil temperature, humidity, pH value, and nutrient content. This invention can cluster data based on data similarity, improving data fusion accuracy and providing accurate data support for decisions such as irrigation or fertilization, thereby enhancing resource utilization efficiency.

[0057] In industrial environments, nodes collect parameters such as equipment vibration, temperature, and noise in real time. This invention can identify regions of data heterogeneity, achieve high-precision monitoring, and support equipment status early warning and environmental safety monitoring.

[0058] In urban environmental monitoring, nodes collect parameters such as PM2.5, NO2, temperature, and humidity. This invention can handle spatially heterogeneous data distributions, improve monitoring accuracy, and support pollution source location and environmental quality assessment.

[0059] In scenarios such as intelligent transportation and intelligent lighting, a large number of sensors communicate via WSNs. This invention can reduce communication power consumption, extend network lifespan, and reduce maintenance costs.

[0060] In the event of natural disasters or emergencies, emergency monitoring networks need to be deployed rapidly and operated long-term. The dynamic adjustment mechanism of this invention can adapt to network changes, pinpoint the scope of the event, and ensure the real-time nature and reliability of monitoring data.

Claims

1. A high-precision clustering method for data fusion in wireless sensor networks, characterized in that, Includes the following steps; Step 1: Establish a high data similarity clustering algorithm optimization model and construct a multi-objective optimization function; Step 2: Based on the distributed environment, a high similarity clustering algorithm is used for initial clustering; Step 3: Adjust the clustering results from Step 2 again, and use the fuzzy C-means method to calculate the membership degree of the wireless sensor node to multiple clusters to determine the final affiliation of the node; Step 4: After the clustering is determined, the cluster head node is selected by comprehensively considering factors such as the remaining energy of the nodes, communication cost, and centrality. Step 5: Monitor the status of nodes within the cluster in real time, and dynamically adjust rather than re-cluster based on network conditions; Step 6: When the proportion of nodes whose affiliation needs to be adjusted due to changes in monitoring data is lower than the preset threshold, these nodes will model inter-cluster resource competition as a non-cooperative game. The node and the cluster heads of other clusters that may accept the node will decide the affiliation of the node based on the utility function. Step 7: When a local area needs adjustment, only the part that needs adjustment should be re-clustered locally; Step 8: When network monitoring data changes on a large scale, perform a global reorganization of the network topology, re-execute Step 1, and restore the optimal clustering state of the network.

2. The high-precision clustering method for data fusion in wireless sensor networks according to claim 1, characterized in that, Step 1 defines a two-dimensional planar monitoring area of ​​area M×M, containing N randomly deployed sensor nodes, each collecting a K-dimensional data vector; the N nodes are divided into L clusters, a high data similarity clustering algorithm optimization model is established, and a multi-objective optimization function is constructed. The specific steps are as follows: Step 1.1, Define the clustering matrix A∈{0,1} (N×L) and cluster head matrix B∈{0,1} (N×L) The clustering scheme is characterized by each node belonging to a cluster, and each cluster having a cluster head. Where a ij The elements in matrix A represent the partitioning relationship between nodes and clusters; b ij The elements in matrix B represent the partitioning relationship between cluster head nodes and ordinary nodes; Step 1.2, Define performance metrics: (1): Mahalanobis distance is used to measure the similarity of node data, which measures the similarity of node data in high-dimensional space, node s i Data x i To node s j Data x j The Mahalanobis distance is: Where Σ is the global covariance matrix, which can automatically eliminate differences in dimensions and take into account the correlation between dimensions; (2): Node s i Position (posx) i ,posy i ) to node s j Position (posx) j ,posy j The communication distance is: If it is cluster C j Any member node s i The communication distance to the cluster head is defined as: Where (posx) CHj ,posy CHj ) is cluster C j cluster head s CHj Location; (3): Define the sum of squared errors within a cluster based on Mahalanobis distance to directly measure the data dispersion of the cluster and achieve tight aggregation of data dimensions: v j It is the weighted average of the data vectors of all nodes within the cluster; (4): The sum of energy consumption is defined as the energy consumption of the entire communication process from the node within the cluster to the cluster head and from the cluster head to the base station: Where E Tx (l, d) represents the total energy consumed by the sensor node in sending a data packet of length l bits to a receiver at a distance d, E DA Energy consumption for processing 1 bit of data fusion, C j l is the sum of the data received by the cluster head within the cluster. f It is the merged data; Step 1.3: Taking into account both data dimension objectives and energy dimension objectives, define a multi-objective optimization function and constraints. The data dimension objective is to minimize the sum of squared Mahalanobis distances within the cluster to ensure tight clustering of data within the cluster. The energy dimension objective is to minimize the energy consumption of communication within the cluster and the transmission energy consumption from the cluster head to the base station. The weighting coefficients are used to balance data accuracy and communication efficiency. 。 3. The high-precision clustering method for data fusion in wireless sensor networks according to claim 2, characterized in that, The specific constraints are as follows: (1): Each node in the network must be assigned to a cluster, and each node can only belong to one cluster; (2): Each cluster must contain at least one node, and empty clusters cannot exist; ； (3): Each cluster has one and only one cluster head; (4): The cluster head must belong to the cluster; ； (5): To prevent nodes that are too far apart in space from being assigned to the same cluster, thereby avoiding communication interruption or excessive energy consumption within the cluster and ensuring the reachability of communication within the cluster; 。 4. The high-precision clustering method for data fusion in wireless sensor networks according to claim 2, characterized in that, Step 2 specifically involves: Step 2.1, First stage coarse screening: Dimensionality reduction and broadcasting of K-dimensional data, local standardization of data at each node; setting a threshold, adding nodes with a distance less than the threshold to the candidate set; Where z i，k It is node s i Normalization on the Kth dimension, μ k It is node s i The average value of the Kth dimension of all neighbor node data within the communication range, σ k It is node s i The variance of the Kth dimension of the data of all neighboring nodes within the communication range; Step 2.2, the second stage of fine screening; Feature signature technology is used to normalize the standardized data vector into a unit vector, satisfying... Then, multiple random hyperplanes are used for projection, and a binary signature Sig(F) is generated based on the projection results. i ); Calculate the Hamming distance; if the Hamming distance is less than a threshold, then pass the screening. ； The hash function hl(F) i Defined as: ； Step 2.3, Distributed Clustering Process; Each node collects K-dimensional data, calculates a feature signature, and broadcasts it to neighboring nodes; a double screening is performed to form a set of similar nodes. Calculate the cluster head competition weights, and the node with the highest weight is declared as the cluster head; ordinary nodes select the most similar cluster head to join, completing the initial clustering.

5. The high-precision clustering method for data fusion in wireless sensor networks according to claim 4, characterized in that, Step 3 specifically involves: Step 3.1: Taking into account three factors, namely data similarity, communication cost and load balancing, calculate the membership degree of each node to each cluster; ， Where S is the data similarity, C is the communication cost, and B is the load balancing coefficient; Step 3.2: Calculate the membership difference, which is the difference between the largest and second-largest membership. ; If the difference is less than the threshold, it means that the node belongs to multiple clusters to a similar degree and is identified as a boundary node. Step 3.3: Perform fuzzy clustering optimization only on the set of boundary nodes, taking into account both data spatial distance and geographic spatial distance; in The square of the normalized Euclidean distance between node i and the data center of cluster x reflects the data similarity; The communication cost between node i and the data center of cluster x; Update the membership degree and iterate until convergence; select the cluster with the largest membership degree as the final destination to complete the precise assignment of boundary nodes.

6. The high-precision clustering method for data fusion in wireless sensor networks according to claim 5, characterized in that, Step 4 specifically involves: Step 4.1, Multifactor Evaluation Scheme: Where f data To indicate the representativeness of the data, membership degree is used as a measure; f comm Location centrality is represented by the number of neighboring nodes and the average distance; f ener Measured by the ratio of the current energy to the cluster average energy; f BS It indicates the proximity to the base station, measured by the reciprocal of the distance to the base station; Step 4.2, Calculate the overall weight; perform a weighted summation of the four factors to obtain the overall weight; Step 4.3, Distributed Cluster Head Election: Nodes set a random backoff timer, the timer duration of which is inversely proportional to the overall weight; the node whose timer expires first broadcasts a cluster head announcement, and other nodes cancel their timers, confirm the cluster head, and join after receiving the announcement.

7. The high-precision clustering method for data fusion in wireless sensor networks according to claim 6, characterized in that, Step 5 specifically involves: To address the changing network states in dynamic scenarios, an adaptive clustering management mechanism is proposed. Nodes are classified into hot zone nodes, adjacent cluster heads, and unaffected nodes. Hot zone nodes are those that trigger adjustment events, adjacent cluster heads are those that are adjacent to the cluster where the hot zone node is located, and unaffected nodes are those that are not affected by the adjustment. The adjustment scope is limited to a local area to avoid information exchange across the entire network. Hot zone nodes report their status to the cluster head and wait for the cluster head's next instruction.

8. The high-precision clustering method for data fusion in wireless sensor networks according to claim 7, characterized in that, Step 6 specifically involves: Step 6.1: Construct a candidate cluster head set. The candidate cluster heads of hot zone nodes are selected from adjacent cluster heads, and must meet the conditions of geographical distance, remaining energy and load. Step 6.2: When deciding whether to accept a new node, the cluster head needs to weigh the benefits of data similarity, the incremental communication cost, and the incremental load cost; the cluster head benefit function U CH Defined as: Where S is the data similarity benefit, C is the communication cost increment, and P is the load cost increment; Step 6.3: The hot zone node sends a migration request to the current cluster head, and the cluster head forwards the request to the neighboring cluster heads; The adjacent cluster heads calculate the benefit Un and send an acceptance or rejection message; the node selects the cluster head that maximizes the benefit and completes the migration. Where E represents energy status gain, C represents communication quality gain, and Q represents load stability gain; Step 6.4: Use the online update formula to quickly update the data center, standard deviation vector, and feature signature; when a node joins or leaves the cluster, there is no need to recalculate the statistics of all nodes, only to update them according to the incremental formula.

9. The high-precision clustering method for data fusion in wireless sensor networks according to claim 8, characterized in that, Step 7 specifically involves: Step 7.1, Trigger condition judgment; when the number of hot zone nodes exceeds the preset threshold of cluster size, ranging from 20% to 40%, local re-clustering is adopted; this threshold is dynamically adjusted according to the network status; Step 7.2, Local Re-clustering Process: The cluster head sends a re-clustering message to the hot zone nodes. Only nodes that receive this message will participate in the subsequent local re-clustering activities. For nodes that do not receive the re-clustering message, even if they receive re-clustering related messages from the hot zone nodes in the subsequent process, they will directly ignore the message. Step 7.3: Calculate the feature signature of the hot zone node that received the re-clustering message, and perform similarity judgment based on the first stage coarse screening and the second stage fine screening in Step 2; use the clustering algorithm of Step 2 and Step 3 to re-cluster only for hot zone areas; Step 7.4: After the cluster is formed, the cluster head election is carried out in step 4. The new cluster head registers with the base station and establishes inter-cluster routing.

10. The high-precision clustering method for data fusion in wireless sensor networks according to claim 9, characterized in that, Step 8 specifically involves: Step 8.1, Calculate network health indicators; comprehensively consider topology change indicators and inter-cluster distance indicators; The topology change index M reflects the degree of change in network topology, while the inter-cluster distance index Dis reflects the degree of deviation between the current cluster structure and the initial optimal cluster structure. Step 8.2, full network reconstruction trigger judgment; when the network health is lower than the threshold, it means that the network structure has seriously deviated from the optimal state, and the base station broadcasts a full network reconstruction instruction; all nodes clear their current cluster affiliation, execute the complete clustering algorithm from step 1 to step 4, and rebuild the cluster structure.

Images