Iot device intelligent management control method based on cloud edge collaboration

Abstract

The application discloses a cloud-edge collaborative-based intelligent management control method for Internet of Things devices, and belongs to the technical field of Internet of Things device management, and comprises the following steps: an edge node collects signal strength sequences, data packet round-trip delay sequences and connection keeping time sequences of Internet of Things devices under jurisdiction, aligns and maps the sequences into a three-dimensional channel state matrix, and generates a local communication characteristic spectrum through singular value decomposition; a cloud center receives a plurality of local communication characteristic spectra, synthesizes a global characteristic tensor, calculates the communication load contribution degree of each edge node, and feeds back the communication load contribution degree; and the edge node adjusts the heartbeat packet sending interval and the data retransmission timeout according to the communication load contribution degree. The method uses spectral feature extraction of multi-dimensional communication sequences and global load contribution quantification, realizes adaptive dynamic adjustment of communication parameters, reduces signaling overhead and retransmission storm risk on the premise of guaranteeing connectivity reliability, and improves network resource utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of IoT device management technology, specifically to an intelligent management and control method for IoT devices based on cloud-edge collaboration. Background Technology

[0002] In IoT device management, edge nodes typically send heartbeat packets at fixed intervals to maintain device status monitoring and use preset timeout durations to handle data retransmissions. This static configuration cannot be dynamically adjusted according to changes in the communication environment. When network load increases or channel quality fluctuates, fixed heartbeat packets exacerbate channel contention and energy waste, while excessively short retransmission timeouts trigger a large number of invalid retransmissions under instantaneous congestion, further deteriorating the network condition. Conversely, overly conservative fixed parameters lead to delayed fault detection and wasted resources. Existing cloud-edge collaboration solutions decentralize some management functions to the edge, but edge nodes mainly report scalar statistical values ​​such as average received signal strength and average latency to the cloud center. This type of data loses distribution information in the time and device dimensions, making it difficult to describe the structured characteristics of the communication status. The cloud center can only make rough judgments based on coarse statistics and cannot accurately distinguish the differences in communication load contributions of each edge node, limiting global collaborative optimization capabilities.

[0003] Faced with a vast number of IoT devices operating under diverse channel conditions, current technologies lack the means to transform heterogeneous, multi-source sequences, such as signal strength, packet round-trip time, and connection hold duration, into compact and discriminative communication feature representations. This results in the inability to effectively separate the dominant patterns from transient noise in local communication, leading to high redundancy and low information density in uploaded data. Simultaneously, there is a lack of a mechanism in the cloud to integrate local communication features from multiple edge nodes to construct a global view and to finely quantify the load contribution of each node. Therefore, it is necessary to address how to extract spectral features reflecting the essential spatiotemporal structure of communication from multidimensional time series at the edge, and how to utilize distributed spectral features in the cloud center to assess communication load contribution and drive adaptive parameter adjustments. Summary of the Invention

[0004] This invention provides an intelligent management and control method for IoT devices based on cloud-edge collaboration. It aims to solve the problems in existing methods, such as the difficulty of adapting static configuration of communication parameters to dynamic network conditions, the lack of communication structure representation capabilities in edge-reported information, and the inability of the cloud to accurately quantify the load contribution of edge nodes. By extracting structured spectral features of multidimensional communication sequences from edge nodes and synthesizing global feature tensors and calculating load contribution from the cloud center, the method achieves dynamic adaptive adjustment of heartbeat packet transmission interval and data retransmission timeout. Under the premise of ensuring device connectivity reliability, it balances network load, reduces signaling overhead and retransmission costs, and improves the resource utilization efficiency and management flexibility of IoT communication systems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an intelligent management and control method for IoT devices based on cloud-edge collaboration. This method collects communication state vectors of the IoT devices under its jurisdiction through edge nodes. These communication state vectors include signal strength sequences, data packet round-trip delay sequences, and connection hold-up duration sequences. The edge nodes align the communication state vectors according to time labels and map them to a two-dimensional state matrix. The two-dimensional state matrix is ​​then diagonalized to generate a local communication feature spectrum. As a technical solution of the present invention, the edge nodes fill the signal strength sequences, data packet round-trip delay sequences, and connection hold-up duration sequences into three channel layers of the two-dimensional state matrix to form a three-dimensional channel state matrix. Singular value decomposition is performed on each channel layer to extract the maximum singular value and the corresponding left and right singular vectors. The maximum singular values ​​of the three channel layers are combined into a principal eigenvalue vector. Simultaneously, the outer product matrix of the left and right singular vectors is reconstructed into a channel feature matrix, and then concatenated along the depth direction to generate a spectral component matrix. Preferably, if missing values ​​occur during the time label alignment process, linear interpolation is used to fill them in to ensure the integrity of the state matrix.

[0006] Edge nodes upload their local communication feature spectra to the cloud center. After receiving multiple local communication feature spectra from various edge nodes, the cloud center performs global feature tensor synthesis. The cloud center arranges all principal eigenvalue vectors in edge node order to generate a node-eigenvalue matrix, concatenates all spectral component matrices along the first dimension to generate a global spectral tensor, and uses the node-eigenvalue matrix as a weighting coefficient matrix to perform weighted compression on the first dimension of the global spectral tensor, generating a three-dimensional global feature tensor. When the number of edge nodes is less than a preset threshold, the cloud center pads the node-eigenvalue matrix with zero rows to ensure the matrix dimension meets processing requirements, thereby enhancing the stability of the global feature tensor. This global feature tensor can fuse communication feature information from different edge nodes, reflecting the spatiotemporal distribution and correlation characteristics of communication states at the overall network level.

[0007] The cloud center calculates the communication load contribution of each edge node based on the global feature tensor and the local communication feature spectrum of each edge node. Specifically, the cloud center calculates the mean vector and covariance matrix of the global feature tensor along the first dimension, and obtains the difference vector between the principal eigenvalue vector and the mean vector of the target edge node. The difference vector is then weighted using the inverse of the covariance matrix to obtain the Mahalanobis distance value, which characterizes the degree of deviation in the communication state of the target edge node. This Mahalanobis distance value is directly used as the communication load contribution. The cloud center feeds back the communication load contribution to the corresponding edge nodes, enabling each edge node to know its relative position and contribution level in the overall network load structure.

[0008] Edge nodes adaptively adjust the heartbeat packet sending interval and data retransmission timeout time with IoT devices based on their received communication load contribution. The edge node obtains the current number of connected devices and the current data throughput. It multiplies the communication load contribution by the number of connected devices to obtain a first adjustment factor, and divides the communication load contribution by the current data throughput to obtain a second adjustment factor. Based on preset basic heartbeat sending intervals and basic data retransmission timeout times, the edge node subtracts the product of the first adjustment factor and a first preset coefficient from the basic heartbeat sending interval to obtain the updated heartbeat packet sending interval; it also adds the product of the second adjustment factor and a second preset coefficient to the basic data retransmission timeout time to obtain the updated data retransmission timeout time. This adjustment mechanism allows edge nodes with higher communication load contributions to proactively reduce the heartbeat sending frequency and appropriately extend the retransmission waiting time, thus playing a role in peak shaving and valley filling when network load is unevenly distributed, effectively alleviating network congestion. Nodes with lower contribution contributions maintain a more aggressive heartbeat and retransmission strategy to ensure the real-time performance and reliability of device access.

[0009] Through the aforementioned cloud-edge collaborative processing flow, this method extracts local communication data from edge nodes into a compact feature spectrum, performs global feature fusion and load contribution quantification in the cloud, and then feeds it back to the edge side to drive parameter adaptive adjustment, forming a closed-loop intelligent management system from perception, modeling, analysis to control. This system significantly improves the utilization efficiency of IoT communication resources, reduces redundant signaling overhead, and enhances the network's dynamic adaptability under conditions of device scale growth and communication fluctuations.

[0010] The technical effects and advantages provided by the present invention in the above technical solution are as follows: Edge nodes align signal strength sequences, packet round-trip delay sequences, and connection hold-up sequences to a unified sampling time grid using time labels, and fill them into a three-dimensional channel state matrix with rows representing the time dimension and columns representing the device dimension, fully preserving the spatiotemporal correlation between multidimensional communication data. Singular value decomposition is performed on each channel layer to extract the maximum singular value and its corresponding left and right singular vectors. The maximum singular value reflects the principal component of the communication energy of that channel, the left singular vector characterizes the change pattern in the time dimension, and the right singular vector captures the distribution structure in the device dimension. The maximum singular values ​​of the three channels are combined into a principal eigenvalue vector. Simultaneously, the outer product of the left and right singular vectors is calculated and weighted by the principal eigenvalue, reconstructing the feature matrices of each channel. These matrices are then concatenated along the depth direction to form a spectral component matrix, thus obtaining the local communication feature spectrum composed of the principal eigenvalue vector and the spectral component matrix. This process compresses the original high-dimensional, noisy time-series data into low-dimensional spectral features with communication pattern discriminative capabilities, filtering out instantaneous fluctuations and measurement noise, and highlighting the main coupling structures of different devices in the time dimension. The obtained local communication feature spectrum is robust to changes in the number of IoT devices and sampling duration, providing the cloud with a feature representation that is high in information density and low in redundancy. This enables subsequent global analysis to significantly reduce data transmission and computation costs while preserving key communication differences.

[0011] After the cloud center aggregates the principal eigenvalue vectors and spectral component matrices uploaded from multiple edge nodes, it arranges all principal eigenvalue vectors into node-eigenvalue matrices according to the edge nodes. Then, it concatenates all spectral component matrices along the first dimension to form a global spectral tensor. Using the elements of the node-eigenvalue matrix as weighting coefficients, it weights the spectral component matrices of each channel for each node. The weighted results of the three channels are then added along the depth direction to compress and obtain the compressed spectral component matrices of each node, which are then stacked into a three-dimensional global feature tensor. This global feature tensor integrates the joint distribution of communication patterns of all edge nodes across time, device, and parameter channels. It uses principal eigenvalues ​​for importance weighting to strengthen high-energy feature components and suppress secondary noise patterns. The mean vector and covariance matrix of the global feature tensor along the first dimension are calculated, and the Mahalanobis distance from the principal eigenvalue vector of the target edge node to this mean vector is taken as the communication load contribution. Mahalanobis distance incorporates the correlation between feature dimensions through the covariance matrix, and can more accurately measure the deviation of the target node's communication state from the global average state than Euclidean distance. A higher load contribution indicates a heavier communication burden or a greater deviation in communication quality for the edge node. After receiving the load contribution, the edge node calculates an adjustment factor based on the number of local device connections and the current data throughput, dynamically reducing the heartbeat packet transmission interval and increasing the data retransmission timeout. For edge nodes with high load contributions, shortening the heartbeat interval increases the monitoring density of device online status, while extending the retransmission timeout alleviates retransmission contention caused by local channel congestion. This closed-loop feedback adjustment allows communication management parameters to automatically adapt to real-time communication load conditions, avoiding signaling storms or detection lag caused by fixed configurations during load fluctuations. It balances wireless channel occupancy, effectively suppressing unnecessary energy consumption and retransmission overhead while ensuring connectivity reliability, thus improving the overall communication efficiency and resource utilization of the Internet of Things. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0013] Figure 1 This is a flowchart of an intelligent management and control method for IoT devices based on cloud-edge collaboration; Figure 2 This is a schematic diagram of the global feature tensor synthesis process; Figure 3 This is a flowchart illustrating how the cloud center assesses the communication load contribution of edge nodes based on Mahalanobis distance. Figure 4 It is the distribution curve of the principal feature value vectors of multiple edge nodes; Figure 5This is a scatter plot showing the relationship between the number of local device connections and the contribution of communication load. Detailed Implementation

[0014] 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, 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.

[0015] See Figure 1 This invention provides a cloud-edge collaborative intelligent management and control method for IoT devices, comprising: Edge nodes collect communication status vectors of the IoT devices under their jurisdiction. The communication status vectors include signal strength sequences, data packet round-trip time sequences, and connection hold duration sequences. The edge node aligns the communication state vector with time labels and maps it to a two-dimensional state matrix, and performs a diagonalization transformation on the two-dimensional state matrix to generate the local communication feature spectrum of the edge node. The edge nodes upload the local communication feature spectrum to the cloud center. After receiving multiple local communication feature spectra uploaded by multiple edge nodes, the cloud center performs global feature tensor synthesis processing to generate a global feature tensor of the communication network. The cloud center calculates the communication load contribution of each edge node based on the global feature tensor and the local communication feature spectrum of each edge node, and feeds back the communication load contribution to the corresponding edge node. The edge node adjusts the heartbeat packet sending interval and data retransmission timeout between the edge node and the IoT device based on the communication load contribution.

[0016] Example 1

[0017] In practice, after collecting signal strength sequences, data packet round-trip time sequences, and connection hold duration sequences from the IoT devices under its jurisdiction, the edge nodes first extract the timestamps carried by each signal strength value in the signal strength sequence to obtain the acquisition time corresponding to each signal strength value. The acquisition times of all signal strength values ​​are then uniformly aligned to a pre-defined sampling time grid. The sampling time grid consists of a series of equally spaced time points, and the interval between adjacent time points can be set according to the communication frequency of the IoT devices, for example, 1 second. The alignment method is as follows: for each timestamp in the signal strength sequence, the closest time point in the sampling time grid is found, and the signal strength value is assigned to the grid position corresponding to that time point. If a grid time point does not have a directly corresponding signal strength value, but there are signal strength values ​​at adjacent time points, the signal strength value at that grid time point can be obtained through linear interpolation based on the preceding and following signal strength values, ensuring that a signal strength value exists at every grid time point.

[0018] After time alignment of the signal strength sequence, the edge nodes perform time alignment on the packet round-trip delay sequence and the connection hold-up sequence using the same sampling time grid, ensuring that the values ​​of all three sequences lie on the same time coordinate. The time dimension is represented by the row index of the sampling time grid, with the row index order corresponding to the ascending time point order of the sampling time grid; the device dimension is represented by the column index, with each column corresponding to an IoT device under the jurisdiction of the edge node. The time-aligned signal strength sequence values ​​are then filled into a two-dimensional matrix according to the corresponding time and device indices. The number of rows in this two-dimensional matrix equals the number of time points in the sampling time grid, and the number of columns equals the number of IoT devices under the jurisdiction of the edge node, forming the signal strength channel layer. Similarly, the packet round-trip delay sequence values ​​are filled into a two-dimensional matrix of the same size, forming the packet round-trip delay channel layer; the connection hold-up sequence values ​​are filled into a two-dimensional matrix of the same size, forming the connection hold-up time channel layer. The signal strength channel layer, the data packet round-trip delay channel layer, and the connection hold-up duration channel layer are stacked along the depth direction to form a three-dimensional channel state matrix. The three channel layers of the three-dimensional channel state matrix represent the distribution of signal strength, data packet round-trip delay, and connection hold-up duration in the time and device dimensions, respectively.

[0019] The edge nodes perform singular value decomposition on each channel layer of the 3D channel state matrix. Taking the signal strength channel layer as an example, the signal strength channel layer is treated as a real number matrix. The number of rows is the number of time steps. The number of columns represents the number of devices. For the matrix Perform singular value decomposition to obtain ,in yes An orthogonal matrix whose column vectors are left singular vectors; yes An orthogonal matrix whose column vectors are right singular vectors; yes A diagonal matrix whose diagonal elements are singular values, arranged in descending order. From Extract the largest singular value from the diagonal elements, denoted as . And extract the left singular vector corresponding to the maximum singular value, i.e. The first column vector And the right singular vector corresponding to the maximum singular value, i.e. The first column vector . As the main characteristic value of signal strength.

[0020] Edge nodes use the same method to adjust the round-trip delay channel layer matrix for data packets. Perform singular value decomposition to obtain the maximum singular value. and its corresponding left singular vector and right singular vector , As the main feature value of time delay. For the connection-maintaining duration channel layer matrix. Perform singular value decomposition to obtain the maximum singular value. and its corresponding left singular vector and right singular vector , As the connection-preserving principal characteristic value. The signal strength principal characteristic value. Delay principal eigenvalues And connections maintain principal eigenvalues Arranged in sequence, they form a principal eigenvalue vector of length 3.

[0021] After obtaining the maximum singular value and its corresponding singular vector for each channel layer, the edge nodes reconstruct the channel feature matrix for the signal strength channel layer. The first left singular vector is then calculated. transpose of the first right singular vector The outer product is used to generate the first outer product matrix, and then each element in the first outer product matrix is ​​multiplied by the main eigenvalue of the signal strength. The signal strength channel feature matrix is ​​obtained. The calculation process can be expressed as follows:

[0022] in, Represents the signal strength channel characteristic matrix. It is the main characteristic value of signal strength. It is the left singular vector corresponding to the maximum singular value obtained after the singular value decomposition of the signal strength channel layer, and it is in column vector form. It is the transpose of the right singular vector corresponding to the maximum singular value, in row vector form. The outer product matrix of the left singular vector and the right singular vector has a size of . .

[0023] For the packet round-trip delay channel layer, the edge nodes calculate the second left singular vector. transpose of the second right singular vector The outer product is used to generate a second outer product matrix, and each element of the second outer product matrix is ​​multiplied by the principal eigenvalue of the time delay. The time delay channel feature matrix is ​​obtained. For the connection-maintaining channel layer, the third left singular vector is calculated for the edge nodes. transpose of the third right singular vector The outer product is used to generate a third outer product matrix, and each element of the third outer product matrix is ​​multiplied by the concatenation function to preserve the principal eigenvalues. The connection-preserving channel feature matrix is ​​obtained. .

[0024] Edge nodes will use the signal strength channel feature matrix Delay channel feature matrix and connection-maintaining channel feature matrix Stack along the depth direction to generate a size of The spectral component matrix is ​​defined as follows: the first layer in the depth direction is the signal strength channel feature matrix, the second layer is the time delay channel feature matrix, and the third layer is the connection-preserving channel feature matrix. Edge nodes use the principal eigenvalue vector and the spectral component matrix together as their local communication feature spectrum for subsequent uploading to the cloud center.

[0025] Example 2

[0026] In specific implementation, please refer to Figure 2 The cloud center establishes data connections with multiple edge nodes through a communication network, receiving local communication feature spectra uploaded by each edge node. The local communication feature spectra consist of two parts: a principal eigenvalue vector and a spectral component matrix. The cloud center receives the first principal eigenvalue vector and the first spectral component matrix uploaded by the first edge node. The first principal eigenvalue vector contains the first edge node's signal strength principal eigenvalue, time delay principal eigenvalue, and connection maintenance principal eigenvalue; the first spectral component matrix is ​​a three-dimensional matrix generated by the first edge node, with a size equal to the first time step. Multiply by the number of first devices Multiply by 3. The cloud center receives the second principal feature vector and the second spectral component matrix uploaded by the second edge node. The size of the second spectral component matrix is ​​the second time step. Multiply by the number of second devices Multiply by 3. In this way, the cloud center continues to receive until it receives the Nth principal eigenvalue vector and the Nth spectral component matrix uploaded by the Nth edge node, where N is the total number of edge nodes participating in the collaboration.

[0027] The cloud center arranges all principal eigenvalue vectors in order of edge node number, with edge node numbers ranging from 1 to N. The numbering order matches the order in which the edge nodes were added to the network or a preset identifier order. The three components of the first principal eigenvalue vector are used as the first row, the three components of the second principal eigenvalue vector as the second row, and so on, until the three components of the Nth principal eigenvalue vector are used as the Nth row, generating a node-eigenvalue matrix. The node-eigenvalue matrix has N rows and 3 columns. The row index of the node-eigenvalue matrix corresponds to the edge node number, column index 1 corresponds to the position of the signal strength principal eigenvalue component, column index 2 corresponds to the position of the delay principal eigenvalue component, and column index 3 corresponds to the position of the connection-maintaining principal eigenvalue component.

[0028] The cloud center concatenates all spectral component matrices along the first dimension to generate a global spectral tensor. This is because the sizes of the spectral component matrices at different edge nodes vary. and The process may differ; before stitching, the cloud center first performs size normalization on each spectral component matrix. The normalization process involves setting a uniform number of time steps. and uniform number of devices For time steps less than The spectral component matrix is ​​padded with zeros at the end of the time dimension. For devices with fewer than The spectral component matrix is ​​padded with zero columns at the end in the device dimension. After size normalization, the size of all spectral component matrices is 1. The cloud center uses the normalized first spectral component matrix as the first slice of the global spectral tensor along the first dimension, the normalized second spectral component matrix as the second slice, and so on, stacking all N normalized spectral component matrices along the first dimension to generate the global spectral tensor. The first dimension of the global spectral tensor corresponds to the edge node number and has a size of N; the second dimension corresponds to the matrix row index, representing the normalized time dimension and has a size of [missing value]. The third dimension corresponds to the matrix column index, representing the normalized device dimension, with a dimension size of [missing information]. The fourth dimension corresponds to the channel depth index, with a dimension size of 3. Depth index 1 corresponds to the signal strength channel layer, depth index 2 corresponds to the data packet round-trip delay channel layer, and depth index 3 corresponds to the connection hold time channel layer.

[0029] In some embodiments, the cloud center can extract the element in the a-th row and b-th column of the node-eigenvalue matrix as the b-th feature weight coefficient of the a-th edge node, where the value of a ranges from 1 to N, and the value of b ranges from 1 to 3. The feature weight coefficient is directly taken from the corresponding principal eigenvalue in the node-eigenvalue matrix. Therefore, the three feature weight coefficients corresponding to the a-th edge node are the principal eigenvalue of signal strength in the a-th row and 1-th column of the node-eigenvalue matrix, the principal eigenvalue of time delay in the a-th row and 2-th column, and the principal eigenvalue of connection maintenance in the a-th row and 3-th column, respectively.

[0030] The cloud center extracts the slice corresponding to the a-th edge node from the global spectral tensor, which is the sub-tensor of the global spectral tensor with index a in the first dimension. This sub-tensor has a size of The spectral component matrix is ​​then decomposed along the fourth dimension to obtain three components of size . The matrix at depth index 1 in the fourth dimension is the first spectral component matrix of the a-th edge node, corresponding to the signal strength channel feature matrix; the matrix at depth index 2 is the second spectral component matrix of the a-th edge node, corresponding to the time delay channel feature matrix; and the matrix at depth index 3 is the third spectral component matrix of the a-th edge node, corresponding to the connection-keeping channel feature matrix.

[0031] The cloud center multiplies each element of the first spectral component matrix of the a-th edge node by the first feature weight coefficient of the a-th edge node to obtain the first weighted spectral component matrix of the a-th edge node; it multiplies each element of the second spectral component matrix by the second feature weight coefficient of the a-th edge node to obtain the second weighted spectral component matrix; and it multiplies each element of the third spectral component matrix by the third feature weight coefficient of the a-th edge node to obtain the third weighted spectral component matrix. The dimensions of the three weighted spectral component matrices are all [missing information]. .

[0032] The cloud center adds the first, second, and third weighted spectral component matrices of the a-th edge node along the depth direction, i.e., summing the elements at corresponding positions, to generate the compressed spectral component matrix of the a-th edge node. This process is calculated as follows, for the time dimension index... From 1 to Device dimension index From 1 to Position in the compressed spectral component matrix The element values ​​are determined by the positions of the first weighted spectral component matrix. Element, position of the second weighted spectral component matrix Position of elements and third weighted spectral component matrix The sum of the elements is obtained. The size of the compressed spectral component matrix is... It merges the weighted feature matrices of the three channels into a single two-dimensional matrix.

[0033] The cloud center performs the above weighted summation operation on all N edge nodes to generate N nodes of size [missing value]. The compressed spectral component matrices are then stacked along the first dimension. That is, the first slice of the first dimension contains the compressed spectral component matrix of the first edge node, the second slice contains the compressed spectral component matrix of the second edge node, and so on, ultimately generating a matrix of size [missing information]. The three-dimensional global feature tensor. The first dimension of the global feature tensor corresponds to the edge node number, the second dimension corresponds to the normalized time dimension, and the third dimension corresponds to the normalized device dimension.

[0034] In some implementations, the compressed spectral component matrix of the a-th edge node The generation can be represented as:

[0035] in, This represents the compressed spectral component matrix of the a-th edge node, with size . ; 'a' represents the edge node index; This represents the first feature weight coefficient of the a-th edge node, taken from the element in the a-th row and 1-th column of the node-eigenvalue matrix, which is the main eigenvalue of the signal strength. The second feature weight coefficient of the a-th edge node is taken from the element in the a-th row and 2-th column of the node-eigenvalue matrix, i.e., the delayed principal eigenvalue. The third feature weight coefficient of the a-th edge node is taken from the element in the a-th row and 3-th column of the node-eigenvalue matrix, which is the connection-preserving principal eigenvalue. This represents the first spectral component matrix of the a-th edge node extracted from the global spectral tensor at the fourth-dimensional depth index 1; This represents the second spectral component matrix of the a-th edge node at the fourth dimension depth index 2; This represents the third spectral component matrix of the a-th edge node at the fourth dimension depth index 3. , , All dimensions are 。 operator The sign indicates scalar multiplication of a scalar and a matrix, while the plus sign indicates element-wise addition of the matrices.

[0036] Through the above process, the cloud center synthesizes a three-dimensional global feature tensor that characterizes the state of the entire communication network by utilizing the local communication feature spectra of all edge nodes.

[0037] See Figure 4This figure illustrates the trends of signal strength, delay, and connectivity persistence as a function of edge node numbers in the principal feature value vectors received by the cloud center from multiple edge nodes in Example 2. The horizontal axis represents the edge node number, ranging from 1 to 200, and the vertical axis represents the principal feature values, reflecting the main characteristics of the local communication feature spectrum of the edge nodes. The red solid line with square markers represents the signal strength principal feature value, with an overall range of approximately 5 to 16. The curve exhibits some fluctuation but generally remains in the mid-to-high value range. The green dashed line with triangular markers represents the delay principal feature value, with generally lower values, mainly distributed between 1 and 6. The curve shows smaller fluctuations and a slight upward trend. The purple dotted line with round markers represents the connectivity persistence principal feature value, with a wider range of values, approximately between 5 and 18. The curve shows significant fluctuations and several peak points, indicating that the connectivity persistence characteristics vary considerably among different edge nodes.

[0038] Example 3

[0039] In specific implementation, please refer to Figure 3 The cloud center obtains the generated global feature tensor. The first dimension of the global feature tensor corresponds to the edge node number. For the numbered edge node... For edge nodes, the cloud center extracts the corresponding feature value components from the corresponding positions of the global feature tensor. These feature value components consist of the main feature value of signal strength, the main feature value of time delay, and the main feature value of connectivity preservation, thus constituting the edge node's feature value. The eigenvalue vectors of all edge nodes in the global feature tensor are arranged in numerical order.

[0040] The cloud center calculates the mean vector of the global feature tensor along the first dimension. Each component of the mean vector is obtained by the arithmetic mean of the corresponding feature value components of all edge nodes. Specifically, let the total number of edge nodes be... , No. The main characteristic value of the signal strength of each edge node is The main characteristic value of the time delay is The connection maintains the main feature value. The cloud center calculates the mean of the main characteristic values ​​of signal strength. The mean of the principal eigenvalues ​​of the time delay The connection preserves the mean of the principal eigenvalues. .Will , , Arranged sequentially, forming the mean vector ,in This represents the mean vector with dimension . , This indicates the transpose operation.

[0041] When calculating the covariance matrix in the cloud center, the deviation between the eigenvalue vector and the mean vector of each edge node is first calculated. The deviation vector of each edge node is Covariance matrix for Matrix, its first Line 1 The column elements are calculated by taking the values ​​of all edge nodes at the 1st... The eigenvalue component and the eigenvalue component The sum of the products of the deviations of the eigenvalue components divided by ,in and The range of values ​​is , Corresponding to the main eigenvalue components of the signal strength, Corresponding to the principal eigenvalue components of the time delay, Corresponding connections preserve the principal eigenvalue components. Covariance matrix The values ​​of the elements reflect the degree of variation and correlation between different components.

[0042] The cloud center identifies the target edge node to be evaluated; the target edge node's number is... The cloud center obtains the principal eigenvalue vector from the local communication feature spectrum of the target edge node. This principal eigenvalue vector contains the principal eigenvalues ​​of the target edge node's signal strength. Delay principal eigenvalues And connections maintain principal eigenvalues , forming vectors .

[0043] Cloud center calculates principal eigenvalue vectors With mean vector The difference vector between , Each component of the difference vector represents the degree of deviation of the target edge node from the global average level in the corresponding communication feature dimension.

[0044] The cloud center calculates the Mahalanobis distance using the difference vector and the inverse of the covariance matrix. The calculation process involves first transposing the difference vector... Multiply by the inverse of the covariance matrix Then compare the result with the difference vector. Multiply to obtain the Mahalanobis distance value. This can be expressed as a formula:

[0045] in, Represents the target edge node Mahalanobis distance value; It is a difference vector The transpose of is Row vectors; It is the covariance matrix The inverse matrix of size , The positive definiteness is guaranteed by the distribution characteristics of the eigenvalue vectors of each edge node, when In singular cases, the cloud center uses a generalized inverse matrix instead of an inverse matrix; It is a difference vector, which is Column vector. It is a non-negative real number. The larger the value, the farther the communication feature vector of the target edge node is from the center of the global communication feature distribution, and the higher the degree of deviation of the communication state of the edge node from the overall network.

[0046] The cloud center will calculate the Mahalanobis distance value Directly used as target edge nodes The communication load contribution is a dimensionless value. The cloud center sends the communication load contribution to the target edge node via the downlink communication link, so that the target edge node can perform subsequent adjustments to the heartbeat packet sending interval and data retransmission timeout.

[0047] See Figure 5 In the graph, the horizontal axis represents the number of local device connections at the edge node, measured in "units," ranging from approximately 5 to 120; the vertical axis represents the communication load contribution, ranging from 0 to approximately 15. Each point in the scatter plot represents a measurement of the communication load contribution of an edge node at a given moment. The point's color gradually transitions from dark to light, with the color gradient positively correlated with the communication load contribution value—the higher the value, the lighter the color. The overall trend shows that as the number of local device connections increases, the communication load contribution exhibits a significant upward trend, indicating that the more connected devices, the greater the communication load contribution of the edge node. In the lower connection count range (approximately 5 to 40), the communication load contribution is mainly concentrated between 0 and 6, and the distribution is relatively dense and even. In the higher connection count range (approximately 60 to 120), the communication load contribution increases significantly, with some nodes exceeding 10, and the maximum value approaching 15. The scatter plot distribution shows greater dispersion and volatility.

[0048] Example 4

[0049] In practice, edge nodes maintain a communication connection with the cloud center and receive the communication load contribution calculated and fed back by the cloud center for each edge node. The communication load contribution is a Mahalanobis distance value calculated by the cloud center based on the global feature tensor and the local communication feature spectrum of the edge node, denoted as […]. , It is a non-negative real number.

[0050] The edge node obtains the current number of local device connections. The number of local device connections refers to the total number of IoT devices under the edge node's jurisdiction that maintain a valid communication connection at the current moment, denoted as . Edge nodes maintain a device connection table to count the number of local device connections. Each entry in the device connection table corresponds to a connected IoT device, and the number of entries is the number of connections. .

[0051] The edge node obtains its current data throughput, which refers to the amount of data processed by the edge node per unit time within the current statistical period, denoted as . The unit is bits per second. Edge nodes deploy traffic counters at the network interface layer, sampling the total amount of data sent and received every preset statistical period, and then dividing the total amount of data sent and received by the duration of the preset statistical period to obtain the result. .

[0052] Edge node contribution to communication load Number of local device connections and current data throughput Calculate the first adjustment factor and the second adjustment factor. The first adjustment factor is calculated by taking the communication load contribution... Multiply by the number of local device connections The first adjustment factor is denoted as , . Dimensions and Consistency reflects the overall load level after the communication load contribution is amplified by equipment scale. The second adjustment factor is calculated by adjusting the communication load contribution... Divide by the current data throughput The second adjustment factor is denoted as , . The unit of measurement is per second per bit, which reflects the contribution of communication load carried by a unit of data throughput.

[0053] Edge nodes read preset parameters, including the basic heartbeat sending interval. Basic data retransmission timeout First preset coefficient Second preset coefficient Basic heartbeat sending interval This is the default heartbeat sending interval predefined by the communication protocol, measured in seconds. The value of the basic heartbeat sending interval is set by network operations and maintenance personnel during deployment at the edge nodes based on the communication environment. For example, it can be set to... Seconds. Basic data retransmission timeout. This is the default data retransmission timeout period predefined by the communication protocol, measured in seconds. The value of the basic data retransmission timeout is determined by the communication protocol stack based on the estimated network round-trip delay. For example, it can be set to... Seconds. First preset coefficient. It is a scaling factor that converts the first adjustment factor into a time quantity, with the dimension of seconds. The specific values ​​are determined based on the communication frequency and response sensitivity requirements of the IoT devices under the jurisdiction of the edge node. In scenarios where the edge node has higher requirements for communication response sensitivity, the values ​​will be adjusted accordingly. A smaller value helps prevent excessive contraction of the heartbeat interval, which could lead to a signaling storm. (Second preset coefficient) It is a scaling factor that converts the second adjustment factor into a time quantity, with the dimension of bits. The specific value is determined based on the average data packet size and retransmission efficiency requirements. When the average data packet size is large... The value increases accordingly.

[0054] Edge nodes calculate the updated heartbeat packet sending interval and updated data retransmission timeout The update process is represented by the following formula:

[0055] in, This indicates the updated heartbeat packet sending interval, in seconds; This represents the basic heartbeat sending interval, which is a preset constant. This represents the first preset coefficient, a preset positive real number, used to convert the dimensionless first adjustment factor into a time dimension. The value is determined by the contribution of communication load. Number of connections to local devices Multiplying them together yields the result. This indicates the timeout period for retransmitting the updated data, in seconds. This represents the timeout period for basic data retransmission, which is a preset constant. This represents the second preset coefficient, a preset positive real number, used to convert the dimensions of the second adjustment factor into the time dimension. The second adjustment factor... The value is determined by the contribution of communication load. Divide by the current data throughput get.

[0056] Edge nodes calculate the updated heartbeat packet sending interval When this happens, a lower bound constraint is applied to the calculation results. If Less than the preset minimum heartbeat sending interval Then Forced to be set to Minimum heartbeat sending interval The value is determined based on the hardware processing capabilities of the IoT device; for example, for low-power IoT devices, Set to 5 seconds. This is the timeout period for edge nodes to calculate and retransmit updated data. When this happens, an upper limit constraint is imposed on the calculation results. If The maximum data retransmission timeout is greater than the preset timeout. Then Forced to be set to Maximum data retransmission timeout The value is determined based on the latency tolerance of the upper-layer application; for example, for real-time control applications, Set to 20 seconds.

[0057] The edge node updates the heartbeat packet sending interval parameter to The edge node periodically sends heartbeat packets to its subordinate IoT devices according to the updated heartbeat packet sending interval. The edge node updates the data retransmission timeout parameter to... When an edge node sends a data packet to an IoT device, it starts a retransmission timer. If no corresponding acknowledgment packet is received within the specified time, a data packet retransmission operation is triggered. Edge nodes continuously monitor the communication load contribution reported by the cloud center during each adjustment cycle. When the communication load contribution changes, the above calculation process is re-executed to dynamically adjust the heartbeat packet sending interval and data retransmission timeout.

[0058] Example 5

[0059] In practice, when the edge node performs the alignment operation from the communication state vector to the sampling time grid, it traverses each grid point in the sampling time grid and checks whether the time corresponding to that grid point has an actual sampled value in the signal strength sequence. When the time corresponding to a certain grid point does not have a directly corresponding signal strength value in the signal strength sequence, the edge node determines that the grid point is a missing grid point.

[0060] For missing grid points, the edge nodes search for the most recent sampling point in the signal intensity sequence whose timestamp is before the time of the missing grid point. This is denoted as the forward sampling point, and its timestamp is... The signal strength value is Simultaneously, find the most recent sampling point whose timestamp is after the time of the missing grid point, denoted as the backward sampling point, and the timestamp of the backward sampling point is... The signal strength value is The timestamp of the missing grid point is... ,satisfy .

[0061] The signal strength values ​​of missing grid points are calculated using linear interpolation at the edge nodes. The calculation formula is:

[0062] in, This represents the calculated signal strength fill value for missing grid points, expressed in units of signal strength measurement. This represents the signal strength value at the forward sampling point; This represents the signal strength value at the backward sampling point; Indicates the timestamp of the missing grid point; Indicates the timestamp of the forward sampling point; Represents the timestamp of the backward sampling point. Denominator Indicates the time interval between forward and backward sampling points, molecule This represents the proportional distribution of change in signal strength over a time interval. Edge nodes will calculate this. Fill in the matrix positions corresponding to the missing grid points to complete the missing value filling.

[0063] For packet round-trip delay sequences and connection hold-up sequences, when there are missing values ​​in the sampling time grid, the edge nodes are also filled using the above linear interpolation method, using the forward and backward sampled values ​​at the corresponding time in their respective sequences to complete the interpolation calculation.

[0064] In some embodiments, when the cloud center performs the operation of arranging all principal eigenvalue vectors in edge node number order to generate a node-eigenvalue matrix, it first obtains the number of edge nodes currently participating in the global feature tensor synthesis. The cloud center will With preset threshold Comparison, preset threshold This is the minimum number of matrix rows required for global feature tensor synthesis at the cloud center. (Preset threshold) The value of is determined based on the numerical stability requirements when the cloud center performs subsequent covariance matrix calculations. When the number of edge nodes is too small, the covariance matrix may exhibit singularity or near-singularity, affecting the accuracy of Mahalanobis distance calculations. According to statistical estimation theory, calculating a dimension of When calculating the covariance matrix, the sample size is usually not less than [amount missing]. Several times. In this embodiment, the dimension of the principal eigenvalue vector is 3. To ensure the invertibility of the covariance matrix and the robustness of the estimation, a preset threshold is used. Set to 3 times 3, that is The basis for this value is that when the number of samples is more than three times the dimension of the variable, the condition number of the covariance matrix tends to be stable numerically, and the quality of the pseudo-inverse solution can be guaranteed.

[0065] When the number of edge nodes Less than the preset threshold At this time, the cloud center appends zero-padding rows to the end of the generated node-eigenvalue matrix. The number of zero-padding rows is... Each row contains 3 elements, all of which are set to 0. The zero-padded node-eigenvalue matrix has 10 rows. The number of columns remains at 3. The edge nodes corresponding to the zero-padding rows do not actually exist. The introduction of zero-padding rows only expands the matrix dimension at the numerical level and does not affect the calculation of the actual edge node's communication load contribution. This is because the principal eigenvalue vector corresponding to the zero-padding rows is a zero vector, which will generate a fixed baseline offset in the Mahalanobis distance calculation. This baseline offset is consistent for all actual edge nodes and will not change the relative ranking of the communication load contributions of each edge node.

[0066] If the number of edge nodes Greater than or equal to the preset threshold The cloud center does not perform zero-padding; it directly constructs a node-eigenvalue matrix using the principal eigenvalue vectors of all current edge nodes. The node-eigenvalue matrix has the following number of rows. The number of columns is 3.

[0067] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A cloud-edge collaborative intelligent management and control method for IoT devices, characterized in that, The method includes: Edge nodes collect communication status vectors of the IoT devices under their jurisdiction. The communication status vectors include signal strength sequences, data packet round-trip time sequences, and connection hold duration sequences. The edge node aligns the communication state vector with time labels and maps it to a two-dimensional state matrix, and performs a diagonalization transformation on the two-dimensional state matrix to generate the local communication feature spectrum of the edge node. The edge nodes upload the local communication feature spectrum to the cloud center. After receiving multiple local communication feature spectra uploaded by multiple edge nodes, the cloud center performs global feature tensor synthesis processing to generate a global feature tensor of the communication network. The cloud center calculates the communication load contribution of each edge node based on the global feature tensor and the local communication feature spectrum of each edge node, and feeds back the communication load contribution to the corresponding edge node. The edge node adjusts the heartbeat packet sending interval and data retransmission timeout between the edge node and the IoT device based on the communication load contribution.

2. The intelligent management and control method for IoT devices based on cloud-edge collaboration according to claim 1, characterized in that, The edge node aligns the communication state vector with time labels and maps it to a two-dimensional state matrix, then performs a diagonalization transformation on the two-dimensional state matrix to generate the local communication feature spectrum of the edge node, including: Extract the timestamps corresponding to each signal intensity value in the signal intensity sequence, and align all timestamps to the same sampling time grid. Using the row index of the sampling time grid to represent the time dimension and the column index to represent the device dimension, the signal strength sequence, the data packet round-trip delay sequence, and the connection hold duration sequence are respectively filled into the three channel layers of the two-dimensional state matrix to form a three-dimensional channel state matrix; Singular value decomposition is performed on each channel layer of the three-dimensional channel state matrix to extract the maximum singular value and the corresponding left and right singular vectors of each channel layer, and the maximum singular values ​​of the three channel layers are combined into the main eigenvalue vector of the local communication feature spectrum. The outer product matrix of the left singular vector and the right singular vector is reconstructed into a channel feature matrix, and the channel feature matrices of the three channel layers are concatenated along the depth direction to generate the spectral component matrix of the local communication feature spectrum.

3. The intelligent management and control method for IoT devices based on cloud-edge collaboration according to claim 2, characterized in that, The edge node performs singular value decomposition on each channel layer of the three-dimensional channel state matrix, extracting the maximum singular value and the corresponding left and right singular vectors of each channel layer, and combines the maximum singular values ​​of the three channel layers into the principal eigenvalue vector of the local communication feature spectrum, including: Singular value decomposition is performed on the signal strength channel layer to obtain the first singular value sequence and the corresponding first left singular vector and first right singular vector. The maximum singular value is extracted from the first singular value sequence as the main feature value of the signal strength. Singular value decomposition is performed on the round-trip delay channel layer of the data packets to obtain the second singular value sequence and the corresponding second left singular vector and second right singular vector. The maximum singular value is extracted from the second singular value sequence as the main feature value of the delay. Singular value decomposition is performed on the connection-keeping duration channel layer to obtain the third singular value sequence and the corresponding third left singular vector and third right singular vector. The maximum singular value is extracted from the third singular value sequence as the connection-keeping principal feature value. The main feature value vector is generated by sequentially arranging the main feature value of the signal strength, the main feature value of the time delay, and the main feature value of the connection preservation.

4. The intelligent management and control method for IoT devices based on cloud-edge collaboration according to claim 3, characterized in that, The edge node reconstructs the outer product matrix of the left singular vector and the right singular vector into a channel feature matrix, and concatenates the channel feature matrices of the three channel layers along the depth direction to generate the spectral component matrix of the local communication feature spectrum, including: Obtain the first left singular vector and the first right singular vector corresponding to the signal strength channel layer, calculate the outer product of the first left singular vector and the first right singular vector to generate a first outer product matrix, and multiply each element in the first outer product matrix by the main eigenvalue of the signal strength to obtain the signal strength channel feature matrix; Obtain the second left singular vector and the second right singular vector corresponding to the time delay channel layer, calculate the outer product of the second left singular vector and the second right singular vector to generate the second outer product matrix, and multiply each element in the second outer product matrix by the time delay principal eigenvalue to obtain the time delay channel feature matrix; Obtain the third left singular vector and the third right singular vector corresponding to the connection-keeping duration channel layer, calculate the outer product of the third left singular vector and the third right singular vector to generate the third outer product matrix, and multiply each element in the third outer product matrix by the connection-keeping principal eigenvalue to obtain the connection-keeping channel feature matrix; The signal strength channel feature matrix, the time delay channel feature matrix, and the connection-keeping channel feature matrix are stacked along the depth direction to generate the spectral component matrix.

5. The intelligent management and control method for IoT devices based on cloud-edge collaboration according to claim 1, characterized in that, The edge nodes upload the local communication feature spectrum to the cloud center. After receiving multiple local communication feature spectra uploaded by multiple edge nodes, the cloud center performs global feature tensor synthesis processing to generate a global feature tensor of the communication network, including: The cloud center receives the first principal eigenvalue vector and the first spectral component matrix uploaded by the first edge node, and receives the second principal eigenvalue vector and the second spectral component matrix uploaded by the second edge node, until it receives the Nth principal eigenvalue vector and the Nth spectral component matrix uploaded by the Nth edge node. The cloud center arranges all the principal feature value vectors in the order of edge node numbers to generate a node-feature value matrix. The row index of the node-feature value matrix corresponds to the edge node number, and the column index corresponds to the position of each component of the principal feature value vector. The cloud center splices all spectral component matrices along the first dimension to generate a global spectral tensor. The first dimension of the global spectral tensor corresponds to the edge node number, the second dimension corresponds to the matrix row index, the third dimension corresponds to the matrix column index, and the fourth dimension corresponds to the channel depth index. The cloud center uses the node-eigenvalue matrix as a weight coefficient matrix to perform weighted compression processing on the first dimension of the global spectral tensor, generating a three-dimensional global feature tensor.

6. The intelligent management and control method for IoT devices based on cloud-edge collaboration according to claim 5, characterized in that, The cloud center uses the node-eigenvalue matrix as a weight coefficient matrix to perform weighted compression processing on the first dimension of the global spectral tensor, generating a three-dimensional global feature tensor, including: The cloud center extracts the element in the i-th row and j-th column of the node-feature value matrix as the j-th feature weight coefficient of the i-th edge node, where the value of i ranges from 1 to N and the value of j ranges from 1 to 3. The cloud center extracts the first spectral component matrix, the second spectral component matrix, and the third spectral component matrix corresponding to the i-th edge node from the global spectral tensor, and multiplies the first spectral component matrix, the second spectral component matrix, and the third spectral component matrix by the first feature weight coefficient, the second feature weight coefficient, and the third feature weight coefficient of the i-th edge node, respectively, to obtain the three weighted spectral component matrices of the i-th edge node. The cloud center adds the three weighted spectral component matrices of the i-th edge node along the depth direction to generate the compressed spectral component matrix of the i-th edge node; The cloud center stacks the compressed spectral component matrices of all N edge nodes along the first dimension to generate the three-dimensional global feature tensor.

7. The intelligent management and control method for IoT devices based on cloud-edge collaboration according to claim 1, characterized in that, The cloud center calculates the communication load contribution of each edge node based on the global feature tensor and the local communication feature spectrum of each edge node, and feeds back the communication load contribution to the corresponding edge node, including: The cloud center calculates the mean vector and covariance matrix of the global feature tensor along the first dimension. The cloud center obtains the principal feature value vector from the local communication feature spectrum of the target edge node, and calculates the difference vector between the principal feature value vector and the mean vector; The cloud center multiplies the transpose of the difference vector by the inverse of the covariance matrix, and then multiplies it by the difference vector again to obtain the Mahalanobis distance value of the target edge node. The cloud center uses the Mahalanobis distance value as the communication load contribution of the target edge node and sends the communication load contribution to the target edge node.

8. The intelligent management and control method for IoT devices based on cloud-edge collaboration according to claim 1, characterized in that, The edge node adjusts the heartbeat packet sending interval and data retransmission timeout between the edge node and the IoT device based on the communication load contribution, including: The edge node receives the communication load contribution from the cloud center and obtains the current number of local device connections and the current data throughput of the edge node. Multiply the communication load contribution by the number of local device connections to obtain a first adjustment factor, and divide the communication load contribution by the current data throughput to obtain a second adjustment factor; Get the preset basic heartbeat sending interval and basic data retransmission timeout; The updated heartbeat packet transmission interval is obtained by subtracting the product of the first adjustment factor and the first preset coefficient from the basic heartbeat transmission interval. The updated data retransmission timeout is obtained by adding the product of the second adjustment factor and the second preset coefficient to the basic data retransmission timeout.

9. The intelligent management and control method for IoT devices based on cloud-edge collaboration according to claim 2, characterized in that, When the edge node aligns the communication state vector to the same sampling time grid according to the time label, if the sampling time grid is missing a signal strength value, then the missing value is filled by linear interpolation.

10. The intelligent management and control method for IoT devices based on cloud-edge collaboration according to claim 5, characterized in that, When the cloud center arranges all the main feature value vectors in the order of edge node numbers, if the number of edge nodes N is less than a preset threshold, zero rows are added to the end of the node-feature value matrix so that the number of matrix rows reaches the preset threshold.

Images