A control method and system for internet of things remote monitoring

Abstract

The application discloses a kind of control method and system of internet of things remote monitoring, it is related to internet of things intelligent monitoring technical field, including, internet of things equipment operation data set is input into space-time diagram convolution model, local space-time feature extraction is carried out in edge layer, cross-device collaborative mode analysis is carried out in cloud layer, generates abnormal propagation path, three-dimensional space coordinate parameters in internet of things equipment operation data set are mapped as the node of topological structure, simultaneously, the interactive data between equipment is mapped as the edge of topological structure, constructs dynamic knowledge graph, abnormal propagation path is injected into dynamic knowledge graph, the fault influence weight between node is updated using improved graph convolution fusion algorithm, obtains propagation risk node, and the dynamic sorting and community clustering analysis are carried out to propagation risk node.The application enhances the identification ability to fault behavior in internet of things equipment by improved graph convolution fusion algorithm and space-time diagram convolution model, and simultaneously improves the efficiency of edge and cloud collaborative analysis.

Description

A control method and system for remote monitoring via the Internet of Things Technical Field

[0001] This invention relates to the field of IoT intelligent monitoring technology, and in particular to a control method and system for IoT remote monitoring. Background Technology

[0002] With the rapid development of Internet of Things (IoT) technology, remote monitoring has been widely applied in various fields such as industrial automation, smart cities, and energy management. As an important application of IoT, remote monitoring collects equipment operating status through sensor networks and combines data communication and intelligent analysis technologies to achieve real-time monitoring and control of widely distributed target devices. Traditional remote monitoring systems are mainly based on a centralized architecture, relying on uploading data from a large number of terminal devices to a cloud center for unified processing and decision-making. This approach has certain advantages in terms of scalability, centralized data management, and historical data analysis.

[0003] However, existing technologies still have limitations: First, traditional spatiotemporal graph convolutional models use fixed topologies and cannot dynamically adapt to changes in the interaction relationships between devices, resulting in limited accuracy in predicting anomaly propagation paths; Second, data collaboration between the edge and the cloud lacks an efficient gradient aggregation mechanism, making it difficult to effectively integrate local features and global patterns, thus affecting the real-time decision-making accuracy and response speed of IoT remote monitoring. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a control method for remote monitoring of the Internet of Things to solve the problems of inaccurate prediction of abnormal propagation paths and low efficiency of edge-cloud collaboration.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a control method for remote monitoring of the Internet of Things, which includes inputting the Internet of Things device operation dataset into a spatiotemporal graph convolutional model, extracting local spatiotemporal features at the edge layer, performing cross-device collaborative mode analysis at the cloud layer, and generating anomaly propagation paths.

[0008] The three-dimensional spatial coordinate parameters in the IoT device operation dataset are mapped to nodes in the topology, and the interaction data between devices is mapped to edges in the topology to construct a dynamic knowledge graph.

[0009] Anomaly propagation paths are injected into a dynamic knowledge graph. An improved graph convolutional fusion algorithm is applied to update the fault impact weights between nodes, obtain propagation risk nodes, and perform dynamic sorting and community clustering analysis on propagation risk nodes to form a multi-level remote monitoring scheme.

[0010] The multi-level remote monitoring scheme is converted into control commands to drive the edge execution unit of the Internet of Things to adjust device parameters, synchronously collect device operation feedback data, and incrementally update the spatiotemporal graph convolution model.

[0011] As a preferred embodiment of the control method for remote monitoring of the Internet of Things (IoT) according to the present invention, the IoT device operation dataset includes three-dimensional spatial coordinate parameters, inter-device interaction data, device operation status data, and spatiotemporal reference data.

[0012] In a preferred embodiment of the IoT remote monitoring control method of the present invention, the generation of the abnormal propagation path specifically includes the following steps.

[0013] The edge layer and cloud layer are built and initialized, and the cross-layer gradient aggregation engine is used to parameterize and stack the edge layer and cloud layer to construct a spatiotemporal graph convolution model.

[0014] The dataset of IoT device operation is input into the spatiotemporal graph convolution model, and the edge layer uses lightweight graph convolution kernels to extract local spatiotemporal features to form a device anomaly feature map;

[0015] The cloud layer uses deep residual networks to perform cross-device collaborative pattern analysis and generate a global correlation feature map.

[0016] The device anomaly feature map and the global correlation feature map are concatenated in the feature channel to obtain a comprehensive anomaly feature vector;

[0017] A multi-hop attention mechanism is used to aggregate multi-hop neighborhood information of the comprehensive anomaly feature vector to generate anomaly propagation paths.

[0018] As a preferred embodiment of the IoT remote monitoring control method of the present invention, the construction of the dynamic knowledge graph specifically includes the following steps.

[0019] A spatial embedding algorithm is used to perform associative mapping on 3D spatial coordinate parameters to form nodes of the topological structure; the HTTP protocol is used to perform attribute mapping on data exchanged between devices to obtain the edges of the topological structure.

[0020] Dynamic weights are assigned and weighted to the edges and nodes of the topology to generate a dynamic knowledge graph.

[0021] As a preferred embodiment of the IoT remote monitoring control method of the present invention, the step of acquiring the propagation risk node specifically includes the following steps.

[0022] Anomaly propagation paths are injected into the dynamic knowledge graph through an Apache Flink pipeline, and an improved graph convolutional fusion algorithm is applied to perform neighborhood aggregation on the dynamic knowledge graph to form a fault association enhancement graph.

[0023] Based on the fault association enhancement graph, the fault impact weights between nodes are updated to generate nodes with propagation risk.

[0024] As a preferred embodiment of the control method for IoT remote monitoring according to the present invention, the formation of a multi-level remote monitoring scheme specifically includes the following steps.

[0025] The LeaderRank algorithm is used to dynamically sort the nodes that are at risk of propagation, generating a priority list for risk propagation.

[0026] Spectral clustering is used to perform community clustering analysis on the risk propagation priority list to obtain high-risk sets; hierarchical spatial mapping is then performed on the high-risk sets to form a multi-level remote monitoring scheme.

[0027] As a preferred embodiment of the control method for IoT remote monitoring described in this invention, the incremental update of the spatiotemporal graph convolutional model specifically includes the following steps.

[0028] The multi-level remote monitoring solution is converted into control commands via the Modbus protocol, and the PID controller is driven by the Internet of Things to perform equipment parameter regulation.

[0029] Simultaneously, the IoT gateway is used to collect device operation feedback data and input it into the spatiotemporal graph convolution model;

[0030] The spatiotemporal graph convolutional model uses gated causal convolution to calculate the anomaly probability of the equipment operation feedback data, obtain the anomaly propagation probability distribution, and perform gradient backpropagation on the anomaly propagation probability distribution to generate updated gradient parameters.

[0031] Based on the updated gradient parameters, the convolution kernel parameters of the spatiotemporal graph convolution model are incrementally updated, and the updated spatiotemporal graph convolution model is output.

[0032] Secondly, the present invention provides a control system for remote monitoring via the Internet of Things, comprising an anomaly analysis module, a map construction module, a scheme generation module, and a closed-loop optimization module.

[0033] The anomaly analysis module is used to input the device spatiotemporal operation dataset into the spatiotemporal graph convolutional model, extract local spatiotemporal features at the edge layer, perform cross-device collaborative mode analysis at the cloud layer, and generate anomaly propagation paths.

[0034] The graph construction module is used to map the three-dimensional spatial coordinate parameters in the spatiotemporal operation dataset of IoT devices into nodes of the topological structure, and at the same time map the interaction data between devices into edges of the topological structure to build a dynamic knowledge graph.

[0035] The scheme generation module is used to inject abnormal propagation paths into the dynamic knowledge graph, apply an improved graph convolutional fusion algorithm to update the fault impact weights between nodes, obtain propagation risk nodes, and perform dynamic sorting and community clustering analysis on propagation risk nodes to form a multi-level remote monitoring scheme.

[0036] The closed-loop optimization module is used to convert multi-level remote monitoring schemes into control commands, drive the edge execution unit of the Internet of Things to adjust device parameters, synchronously collect device operation feedback data, and incrementally update the spatiotemporal graph convolution model.

[0037] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the control method for Internet of Things remote monitoring as described in the first aspect of the present invention.

[0038] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the control method for remote monitoring of the Internet of Things as described in the first aspect of the present invention.

[0039] The beneficial effects of this invention are as follows: by using a dynamic knowledge graph and an improved graph convolution fusion algorithm, the accuracy of anomaly propagation path prediction is improved, and the ability to identify fault evolution behavior in IoT device networks is enhanced; by constructing a spatiotemporal graph convolution model and performing incremental updates, the efficiency of edge and cloud collaborative analysis is effectively improved, control response latency is reduced, and the real-time performance and stability of remote monitoring are improved. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 is a flowchart of the control method for remote monitoring of the Internet of Things.

[0042] Figure 2 is a schematic diagram of the control system for IoT remote monitoring.

[0043] Figure 3 is a flowchart of the abnormal propagation path generation process.

[0044] Figure 4 is a flowchart of the dynamic sorting of nodes at risk of transmission. Detailed Implementation

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0048] Referring to Figures 1-4, an embodiment of the present invention is provided, which offers a control method for remote monitoring of an Internet of Things (IoT) including the following steps:

[0049] S1. Input the IoT device operation dataset into the spatiotemporal graph convolutional model, extract local spatiotemporal features at the edge layer, and perform cross-device collaborative mode analysis at the cloud layer to generate anomaly propagation paths.

[0050] Specifically, the operations include the following:

[0051] S1.1 Collect IoT device operation datasets, which include three-dimensional spatial coordinate parameters, device interaction data, device operation status data, and spatiotemporal reference data.

[0052] The three-dimensional spatial coordinate parameters include the device's longitude, latitude, and altitude location information; the longitude and latitude are collected through built-in GPS and Beidou positioning devices, and the altitude is obtained using a barometer;

[0053] Inter-device interaction data includes communication traffic, packet transmission frequency, and connection status data; communication traffic is monitored using an edge gateway, packet transmission frequency is collected through the packet capture unit (such as libpcap) of the network interface, and protocol type is obtained using a communication protocol parser;

[0054] The equipment operating status data includes temperature and humidity, voltage, current, and vibration amplitude; temperature and humidity are collected using temperature and humidity sensors, voltage is collected using a voltage divider circuit, current is collected using a Hall current sensor, and vibration amplitude is collected using an accelerometer.

[0055] Spatiotemporal reference data includes timestamps, time zone information, and synchronization clock signals; timestamps are collected using an NTP server, time zone information is collected using the GeoNames geographic database, and synchronization clock signals are collected using a PTP time synchronization unit.

[0056] IoT device operation datasets can not only reflect the spatiotemporal operating status and interaction relationships of devices, but also enable collaborative analysis and real-time response in advanced control scenarios.

[0057] S1.2. The collected IoT device operation dataset is preprocessed. Specifically, the dataset is decomposed into five levels using the db8 wavelet basis to suppress high-frequency noise interference, and outlier identification and removal are performed using the 3σ criterion to ensure the purity of the IoT device operation data. Then, dynamic time warping is used to calibrate and align the timestamps of the dataset, and cubic spline interpolation is used to perform equally spaced resampling to eliminate differences in the collection frequency of different devices. Finally, Min-Max normalization is used to normalize the dataset to avoid the influence of different dimensions, and principal component analysis is used to compress the dimensions, outputting the preprocessed IoT device operation dataset.

[0058] S1.3. Construct and train a spatiotemporal graph convolutional model. Specifically, in the TensorFlow framework, a one-dimensional convolutional structure is called via the `filters` parameter, and the `depthwise_conv1d` function is used to perform depthwise separable convolution on the one-dimensional convolutional structure to obtain a lightweight graph convolution. The kernel size of the lightweight graph convolution is set to 3, and the output dimension is set to 64. A `mish` function is applied after the lightweight graph convolution to smooth gradient propagation, and batch normalization is used to perform standardization processing on the lightweight graph convolution, completing the construction of the edge layer. Residual blocks are called via the `num_blocks` class parameter, where each residual block contains two convolutional layers and a skip connection structure, and all convolutional layers use one-dimensional convolutional kernels. The three residual blocks are stacked to construct a deep residual network. The number of hidden units in the deep residual network is set to [128, 256, 512], and the random inactivation rate is set to 0.2. A multi-head attention mechanism is used to associate spatiotemporal features in the deep residual network, and a GELU function is applied to introduce non-linearity, completing the construction of the cloud layer.

[0059] In the gradient aggregation stage, the tf.GradientTape mechanism of the cross-layer gradient aggregation engine is used to capture the local gradients of the edge layer and the global gradients of the cloud layer respectively. The local gradients and global gradients are mixed by a weighted fusion algorithm to obtain cross-layer collaborative features. In the parameterization stacking stage, subclassing of tf.keras.Model is used to perform tensor concatenation and fully connected projection on the cross-layer collaborative features along the feature dimension to generate fusion weights. Based on the fusion weights, the edge layer and cloud layer are parameterized and stacked using backpropagation to complete the construction of the spatiotemporal graph convolutional model.

[0060] Next, the spatiotemporal graph convolutional model is trained. Further, the IoT device runtime dataset is divided into a sample set, a training set, and a validation set in a 6:2:2 ratio. On the sample set, MinMaxScaler is used for data normalization, and a sliding time window is used to segment the samples to obtain standardized time-series samples. On the training set, the SGD optimizer is used to update the gradients of the standardized time-series samples, and gradient clipping is used for training stabilization to obtain updated weights. The learning rate is dynamically adjusted on the updated weights to generate optimized spatiotemporal graph convolutional model parameters. On the validation set, the mean absolute error function is used to calculate the loss of the optimized spatiotemporal graph convolutional model parameters to form the validation loss. Training terminates when the validation loss reaches the convergence threshold for 10 consecutive rounds, and the trained spatiotemporal graph convolutional model is output synchronously.

[0061] It should be noted that the convergence threshold is defined based on the relative rate of change of the validation loss, and its value ranges from [0.001, 0.01].

[0062] The spatiotemporal graph convolution model lays the foundation for accurate prediction of anomaly propagation paths and also improves the dynamic response speed of advanced IoT control devices.

[0063] S1.4. Anomaly propagation paths are generated using a spatiotemporal graph convolution model. Specifically, the IoT device operation dataset is input into the spatiotemporal graph convolution model via the Input interface. The edge layer performs depthwise separable convolution on the IoT device operation dataset using lightweight graph convolution kernels, and then performs graph convolution using Chebyshev multinomials to form a spatial feature tensor. Local spatiotemporal features of the spatial feature tensor are extracted using one-dimensional convolution kernels, and max pooling is performed along the time axis to capture key temporal patterns. Local spatiotemporal features and key temporal patterns are integrated across modalities to generate spatiotemporal fusion features. These spatiotemporal fusion features are standardized using a batch normalization layer, and a nonlinear transformation is performed using the Mish activation function to generate an initial feature map. The initial feature map is then enhanced using residual connections to output a device anomaly feature map containing spatial correlations and temporal variation patterns.

[0064] It should be noted that Chebyshev polynomials refer to orthogonal polynomial bases defined based on recursive relations, which can be directly called through the chebyshev function in the scipy library;

[0065] The device anomaly feature map is input into the cloud layer. The deep residual network of the cloud layer uses a one-dimensional convolutional kernel to extract features from the device anomaly feature map and generate a preliminary feature representation. The preliminary feature representation is then aggregated between nodes using graph convolution to obtain a cross-device collaborative mode matrix. The cross-device collaborative mode matrix is ​​then weighted using the Softmax function to obtain the inter-device association weight matrix. The inter-device association weight matrix is ​​then non-linearly introduced through the GELU activation function and randomly discarded using Dropout with a random inactivation rate of 0.2 to prevent overfitting and form a regularized feature representation.

[0066] Regularized feature representations and device anomaly feature maps are added element-wise through three residual blocks. Max pooling is used to progressively downsample between residual blocks, ensuring enhanced feature abstraction while compressing feature size, thus acquiring high-level semantic features. A multi-head attention mechanism is used to perform feature interaction and fusion on the high-level semantic features, further capturing long-range dependencies across devices and forming global spatiotemporal correlation features. These global spatiotemporal correlation features are then linearly projected using a fully connected layer, outputting a global correlation feature map containing cross-device collaborative patterns.

[0067] One-dimensional convolution kernels are used to adjust the dimensions of the device anomaly feature map and the global correlation feature map to ensure alignment in the feature dimensions. In the feature channel, tensor concatenation is performed on the device anomaly feature map and the global correlation feature map along the aligned feature dimensions to obtain a fused feature tensor. The fused feature tensor is then standardized by batch normalization to eliminate the distribution differences between cross-layer features and output a comprehensive anomaly feature vector.

[0068] A multi-hop attention mechanism is used to map the comprehensive anomaly feature vector into the feature space to form an initial feature representation. A fully connected function is used to perform a linear transformation on the initial feature representation to generate three sets of projection vectors for query, key, and value. The neighborhood of the three sets of projection vectors is then expanded to obtain a multi-hop attention score. The Softmax function is used to aggregate the multi-hop neighborhood information of the multi-hop attention score to generate a multi-hop intermediate result. The state of the multi-hop intermediate result is updated through a gated recurrent unit to obtain the update hop count. The update hop count reflects the spatiotemporal pattern of anomaly propagation. The update hop count is weighted and fused to form an anomaly propagation feature representation. The anomaly propagation feature representation is then linearly projected into a low-dimensional space to generate an anomaly propagation path.

[0069] S2. Map the three-dimensional spatial coordinate parameters in the spatiotemporal operation dataset of IoT devices to nodes in the topology, and simultaneously map the interaction data between devices to edges in the topology to construct a dynamic knowledge graph.

[0070] Specifically, the operations include the following:

[0071] S2.1. A spatial embedding algorithm is used to perform an associative mapping on the 3D spatial coordinate parameters to form nodes in the topological structure. Specifically, the spatial embedding algorithm is used to perform Geohash encoding on the 3D spatial coordinate parameter set, converting the 3D spatial coordinate parameters into a string representation. Then, TF-IDF weighting is used to perform vectorization transformation on the string representation to obtain the geographic embedding vector. Matrix multiplication is used to transform the dimensionality of the geographic embedding vector to generate an intermediate feature matrix. K-nearest neighbor clustering is used to perform nearest neighbor search and similarity clustering on the intermediate feature matrix to obtain device similarity relationships. Based on the device similarity relationships, the geographic embedding vector is associatively mapped to form nodes in the topological structure.

[0072] S2.2 Utilize the HTTP protocol to perform attribute mapping on the data exchanged between devices and obtain the edges of the topology. In specific operations, the HTTP message parser is used to extract the request features of the data exchanged between devices, and the key fields of the request features are matched using the PCRE syntax of the HTTP protocol to obtain standardized interaction records. A weighted statistical algorithm is applied to perform multi-dimensional fusion and normalization transformation on the standardized interaction records to generate weighted device association relationships.

[0073] The validity of device associations is screened using an effective threshold. For example, a sliding window is used to count the request frequency of device associations and obtain the RPS (requests per second) value. When the RPS value exceeds the effective threshold, the validity of the device association is considered to have passed the verification. The verified device associations are then transformed into a topological structure to form a preliminary topological connection. According to the HTTP protocol, the preliminary topological connection is mapped to attributes to form the edges of the topological structure.

[0074] It should be noted that the effective threshold is based on specific business needs and is defined by the request frequency distribution of device association, with a value range of [3, 10].

[0075] S2.3. Dynamically assign and integrate the weights of edges and nodes to generate a dynamic knowledge graph. In the specific operation, the nodes are monitored in real time to obtain the node state data. The node state data is discretized and decomposed using the entropy weight method to obtain the information entropy of the node state. The information entropy of the node state is compressed in dimension and linearly weighted through principal component analysis to generate the node weight coefficient.

[0076] The t-SNE algorithm is used to extract features and construct vectors for edges, forming edge feature vectors. Based on the edge feature vectors, the power iteration method is used to perform eigenvalue decomposition and numerical extraction to generate edge weight coefficients. Based on the node weight coefficients and edge weight coefficients, the nodes and edges are integrated in a low-dimensional vector space through a graph embedding algorithm to obtain a weighted graph structure. The weighted graph structure is dynamically rendered using a visualization engine to generate a dynamic knowledge graph.

[0077] S3. Inject abnormal propagation paths into the dynamic knowledge graph, apply an improved graph convolutional fusion algorithm to update the fault impact weights between nodes, obtain propagation risk nodes, and perform dynamic sorting and community clustering analysis on propagation risk nodes to form a multi-level remote monitoring scheme.

[0078] Specifically, the operations include the following:

[0079] S3.1. Anomaly propagation paths are injected into the dynamic knowledge graph through an Apache Flink pipeline, and an improved graph convolutional fusion algorithm is applied to perform neighborhood aggregation on the dynamic knowledge graph to form a fault association enhancement graph. Specifically, firstly, the anomaly propagation paths are injected into the dynamic knowledge graph through the Apache Flink DataStream API; secondly, the anomaly propagation paths and the dynamic knowledge graph are combined using a Cypher query from the graph database to generate a fault labeling graph; thirdly, the improved graph convolutional fusion algorithm is used to perform path tracing and neighborhood aggregation on the fault labeling graph to obtain the association strength between nodes; finally, the sliding window averaging method is used to perform exponential decay weighted calculations on the association strength between nodes to generate fault propagation coefficients. The specific mathematical formula is as follows:

[0080]

[0081] Where C represents the fault propagation coefficient; τ represents the time span of the sliding window; k represents the time offset within the sliding window; s represents the correlation strength between nodes; and t represents the current time point.

[0082] Gaussian filtering is used to smooth the fault propagation coefficients over time to obtain stable propagation coefficients; based on the stable propagation coefficients, feature enhancement is performed on the fault labeling map using a weighted average method to generate a fault association enhancement map;

[0083] The improved graph convolution fusion algorithm makes the following improvements based on the existing graph convolution fusion method: on the one hand, by parsing the anomaly propagation path injected by Apache Flink in real time, it enhances the ability of dynamic knowledge graph to perform time-series analysis of fault propagation; on the other hand, it uses exponential decay weighted operation based on sliding window to optimize the dynamic fusion of the association strength between nodes, and realizes the collaborative representation of anomaly propagation path and dynamic knowledge graph.

[0084] S3.2. Based on the fault association enhancement graph, update the fault impact weights between nodes to generate propagation risk nodes. In specific operations, extract the edge weights in the fault association enhancement graph and use a sliding window to divide the edge weights into multiple time segments; use an exponentially weighted moving average to fuse and normalize the edge weights of each time segment to form fault propagation weights.

[0085] The initial node representation is formed by aggregating nodes in the fault association enhancement map using a weighted average method. Multi-scale feature decomposition of the fault association enhancement map is performed using wavelet transform to extract temporal features. The initial node representation and temporal features are then weighted and fused to obtain a dynamic risk representation. The dynamic risk representation is weighted and nonlinearly transformed using the Softmax function to obtain the fault influence weights between nodes. Based on the fault propagation weights, a dynamic programming algorithm is used to update the fault influence weights between nodes in real time, obtaining optimized node weights. Finally, based on the optimized node weights, importance sampling is performed on the nodes to form propagation risk nodes.

[0086] S3.3. The LeaderRank algorithm is used to dynamically sort the nodes at risk of propagation, generating a priority list for risk propagation. Specifically, the adjacency matrix method is used to map the nodes to a graph structure and standardize edge weights, constructing a weighted directed graph. The nodes at risk of propagation are used as the initial active nodes in the weighted directed graph, and a background node is added to each initial active node for global association. In the iterative calculation phase, the LeaderRank algorithm is used to update the weights and iterate the vectors of the initial active nodes in parallel, generating a scalar weight vector. A damping factor is used to proportionally adjust the scalar weight vector to obtain the control weight transfer ratio. Based on the control weight transfer ratio, the scalar weight vector is sparsely compressed using matrix multiplication to form the updated node weight values. The specific mathematical formula is as follows.

[0087]

[0088] Among them, R (f+1) This represents the update of node weights after (f+1) iterations, where f is the iteration number, α is the damping factor, x is a column vector of all 1s, N is the total number of nodes in the weighted directed graph after removing background nodes, and W is the transition probability matrix. R represents the transpose operation. (f) This indicates that the node weight value is updated after f iterations;

[0089] It should be noted that the damping factor is based on the definition of state transition probability in random walk theory, and its value range is (0,1); the all-1 column vector refers to an N-dimensional column vector with all elements having a value of 1, which is obtained by assigning all 1s to the unit vector; the transition probability matrix is ​​obtained by normalizing the out-degree of the adjacency relationship of the weighted directed graph.

[0090] Based on the updated node weight values, the weight of the background nodes is recycled using the weight averaging method, and the weight of the background nodes is evenly distributed to other nodes. The background nodes are then removed using the node selection operator to form the final nodes to be sorted. The final nodes to be sorted are then sorted in descending order using the quicksort algorithm to generate a risk propagation priority list.

[0091] It should be noted that the node selection operator is defined based on the index identifier of the background node, and the value range is {0,1}^N, where N is the total number of nodes.

[0092] S3.4. Spectral clustering is used to perform community clustering analysis on the risk propagation priority list to obtain a high-risk set. This high-risk set is then subjected to hierarchical spatial mapping to form a multi-level remote monitoring scheme. Specifically, a similarity matrix is ​​constructed and symmetrically normalized on the risk propagation priority list using a Gaussian kernel function to obtain the Laplace matrix. Spectral clustering is then used to perform eigenvalue decomposition on the Laplace matrix, and the first D non-zero eigenvalues ​​are selected for vector reconstruction to obtain low-dimensional feature vectors. MinMax standardization is applied to normalize the dimensions of the low-dimensional feature vectors, forming a low-dimensional embedding space. K-means clustering is then used to perform community clustering in the low-dimensional embedding space to obtain high-risk communities. These high-risk communities are then weighted and aggregated to output the high-risk set.

[0093] Density clustering is performed on high-risk sets to form potential risk regions. Kernel density estimation is used to fit the risk density distribution of these regions, generating risk heatmap values. Potential risk regions are then hierarchically divided using three threshold levels: for example, the first-level threshold is set to [0.8-1.0], the second-level threshold to [0.5-0.8], and the third-level threshold to [0.3-0.5]. When the risk heatmap value falls within the first-level threshold range, the potential risk region is defined as a core monitoring point, requiring real-time monitoring. When the risk heatmap value falls within the second-level threshold range, the potential risk region is defined as a secondary monitoring point, requiring periodic polling. When the risk heatmap value falls within the third-level threshold range, the potential risk region is defined as an edge monitoring point, requiring daily inspection.

[0094] The core monitoring points, secondary monitoring points, and edge monitoring points are hierarchically integrated in the spatial topology to obtain a three-level monitoring architecture. The three-level monitoring architecture is hierarchically spatially mapped by Kriging interpolation to generate a heat map of the monitoring area. The heat map of the monitoring area is then uniformly sampled to obtain monitoring intensity parameters. These monitoring intensity parameters are then weighted and fused to generate a multi-level remote monitoring scheme.

[0095] S4. Convert the multi-level remote monitoring scheme into control commands to drive the edge execution unit of the Internet of Things to adjust device parameters, synchronously collect device operation feedback data, and incrementally update the spatiotemporal graph convolution model.

[0096] Specifically, the operations include the following:

[0097] S4.1 The multi-level remote monitoring scheme is converted into control commands via the Modbus protocol, driving the IoT edge execution unit to perform device parameter control through the PID controller. In specific operation, the multi-level remote monitoring scheme is encoded into commands via the Modbus protocol, mapping the multi-level remote monitoring scheme into control commands; the IoT edge execution unit performs protocol parsing and digital-to-analog conversion on the control commands to obtain analog signals; the PID controller performs error calculation on the analog signals to obtain deviation signals, and performs signal conditioning and numerical conversion through a digital signal processor to generate control deviation; the proportional coefficient amplifies the control deviation in real time, the integral coefficient accumulates historical deviations, and the derivative coefficient predicts the rate of change of deviation to obtain the parameter control quantity;

[0098]

[0099] Where u represents the parameter control quantity, P represents the proportional coefficient, Z represents the current time, g(Z) represents the control deviation at the current time Z, I represents the integral coefficient, a represents the integral variable, and D represents the derivative coefficient;

[0100] It should be noted that the proportional coefficient is defined based on the actual response speed requirement through step response testing, and its value range is [0.1, 10.0]; the integral coefficient is defined based on the actual steady-state accuracy requirement through the error accumulation time constant, and its value range is [0.01, 1.0]; the derivative coefficient is defined based on the actual anti-interference requirement through the oscillation period and damping ratio of the PID controller, and its value range is [0.001, 0.1]; the integral variable is defined based on the time accumulation effect.

[0101] Based on the parameter control values, the H-bridge power amplifier circuit is driven to adjust the gain of the equipment parameters in order to regulate the response speed of the equipment and improve its performance.

[0102] S4.2. Synchronously collect device operation feedback data using the IoT gateway and input it into the spatiotemporal graph convolution model. In specific operations, during device parameter adjustment, deploy EdgeX Foundry's acquisition software on the IoT gateway. Monitor device operation feedback data at fixed intervals (e.g., 100ms) via the Modbus protocol, and use dynamic time warping to perform timestamp calibration and interpolation compensation on the device operation feedback data to correct time differences and obtain warped device operation feedback data. Input the warped device operation feedback data into the spatiotemporal graph convolution model through the gRPC streaming data interface.

[0103] S4.3 The spatiotemporal graph convolutional model utilizes gated causal convolution to calculate the anomaly probability of equipment operation feedback data, obtaining the anomaly propagation probability distribution. Specifically, a sliding time window is used to extract multidimensional temporal features from the equipment operation feedback data. In the spatial dimension, gated causal convolution is used to aggregate neighborhood features of the multidimensional temporal features. In the temporal dimension, dilated causal convolution is used to expand the temporal features to capture long-term dependencies, outputting a comprehensive spatiotemporal tensor. High probability density estimation is used to calculate the anomaly probability of the comprehensive spatiotemporal tensor, obtaining the Mahalanobis distance score at each time step, and normalizing it using the softmax function to obtain the anomaly propagation probability value. The specific mathematical formula is as follows.

[0104]

[0105] Where Y represents the anomaly propagation probability value, γ represents the anomaly sensitivity coefficient, d represents the time step index, M1(t) represents the Mahalanobis distance score of node 1 at time step t, M2(t) represents the Mahalanobis distance score of node 2 at time step t, j represents the node index, and O represents the total number of nodes.

[0106] It should be noted that the anomaly sensitivity coefficient is defined based on the probability mapping slope of the Mahalanobis distance score, and its value range is [0.5, 2.0].

[0107] Outliers were removed from the anomaly propagation probability values ​​using the 3σ criterion to obtain corrected probability values. The anomaly propagation probability distribution was generated by fitting the probability density surface to the corrected probability values ​​using bicubic spline interpolation.

[0108] S4.4. Perform gradient backpropagation on the anomaly propagation probability distribution to generate updated gradient parameters. Specifically, the `sobel_edges` function of the TensorFlow framework is used to perform two-dimensional convolution embedding on the anomaly propagation probability distribution to obtain the spatial-temporal gradient distribution field. The L2 norm is used to normalize the spatial-temporal gradient distribution field element by element to form a differentiable penalty term gradient. Then, the automatic differentiation engine is used to perform backpropagation on the differentiable penalty term gradient, and bicubic spline interpolation is used for resampling to obtain a smooth backpropagation gradient. Gradient clipping is used to limit the magnitude of the smooth backpropagation gradient to prevent gradient explosion, and stochastic gradient descent (SGD) is applied for iterative updating to generate updated gradient parameters.

[0109] S4.5. Based on the updated gradient parameters, incrementally update the kernel parameters of the spatiotemporal graph convolutional model and output the updated spatiotemporal graph convolutional model. Specifically, the updated gradient parameters are input into the Adam optimizer in the spatiotemporal graph convolutional model; the Adam optimizer applies momentum-accelerated gradient descent to the updated gradient parameters and performs parameter constraints through L2 regularization to generate regularized parameter update amounts; the weights of the convolution kernels are decayed based on the regularized parameter update amounts to generate the final parameter increments; based on the final parameter increments, the kernel parameters are updated in-situ using the tf.Variable class, and the weights of the spatiotemporal graph convolutional model are reloaded using the model.set_weights function to output the updated spatiotemporal graph convolutional model.

[0110] This embodiment also provides a control system for IoT remote monitoring, including: an anomaly analysis module, a map construction module, a scheme generation module, and a closed-loop optimization module.

[0111] The anomaly analysis module is used to input the device spatiotemporal operation dataset into the spatiotemporal graph convolutional model, extract local spatiotemporal features at the edge layer, perform cross-device collaborative mode analysis at the cloud layer, and generate anomaly propagation paths.

[0112] The graph construction module is used to map the three-dimensional spatial coordinate parameters in the IoT device operation dataset to nodes of the topological structure, and at the same time map the interaction data between devices to edges of the topological structure, thereby constructing a dynamic knowledge graph.

[0113] The scheme generation module is used to inject abnormal propagation paths into the dynamic knowledge graph, apply an improved graph convolutional fusion algorithm to update the fault impact weights between nodes, obtain propagation risk nodes, and perform dynamic sorting and community clustering analysis on propagation risk nodes to form a multi-level remote monitoring scheme.

[0114] The closed-loop optimization module is used to convert multi-level remote monitoring schemes into control commands, drive the edge execution unit of the Internet of Things to adjust device parameters, synchronously collect device operation feedback data, and incrementally update the spatiotemporal graph convolution model.

[0115] This embodiment also provides a computer device suitable for the control method of IoT remote monitoring, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the IoT remote monitoring control method proposed in the above embodiment.

[0116] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0117] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the control method for realizing remote monitoring of the Internet of Things as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0118] In summary, this invention improves the accuracy of anomaly propagation path prediction and enhances the ability to identify fault evolution behavior in IoT device networks by using a dynamic knowledge graph and an improved graph convolution fusion algorithm. Furthermore, by constructing a spatiotemporal graph convolution model and performing incremental updates, it effectively improves the efficiency of edge-cloud collaborative analysis, reduces control response latency, and enhances the real-time performance and stability of remote monitoring.

[0119] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A control method for remote monitoring via the Internet of Things, characterized in that: include, The process involves inputting the IoT device operation dataset into a spatiotemporal graph convolutional model. The edge layer extracts local spatiotemporal features, while the cloud layer performs cross-device collaborative pattern analysis to generate anomaly propagation paths. Specifically, the steps include: building and initializing the edge and cloud layers, and using a cross-layer gradient aggregation engine to parameterize and stack the edge and cloud layers to construct the spatiotemporal graph convolutional model; inputting the IoT device operation dataset into the spatiotemporal graph convolutional model, with the edge layer using lightweight graph convolution kernels to extract local spatiotemporal features, forming a device anomaly feature map; using a deep residual network in the cloud layer to perform cross-device collaborative pattern analysis, generating a global correlation feature map; concatenating the device anomaly feature map and the global correlation feature map in the feature channels to obtain a comprehensive anomaly feature vector; using a multi-hop attention mechanism to aggregate multi-hop neighborhood information from the comprehensive anomaly feature vector to generate anomaly propagation paths; and mapping the 3D spatial coordinate parameters in the IoT device operation dataset to nodes in the topology, while mapping inter-device interaction data to edges in the topology, constructing a dynamic knowledge graph. Anomaly propagation paths are injected into a dynamic knowledge graph. An improved graph convolutional fusion algorithm is applied to update the fault impact weights between nodes, obtain propagation risk nodes, and perform dynamic sorting and community clustering analysis on propagation risk nodes to form a multi-level remote monitoring scheme. Furthermore, the improved graph convolutional fusion algorithm, on the one hand, parses the anomaly propagation paths injected by Apache Flink in real time, and on the other hand, uses exponential decay weighted operations based on sliding windows. The multi-level remote monitoring scheme is converted into control commands to drive the edge execution unit of the Internet of Things to adjust device parameters, synchronously collect device operation feedback data, and incrementally update the spatiotemporal graph convolution model.

2. The control method for remote monitoring of the Internet of Things as described in claim 1, characterized in that: The IoT device operation dataset includes three-dimensional spatial coordinate parameters, inter-device interaction data, device operation status data, and spatiotemporal reference data.

3. The control method for remote monitoring of the Internet of Things as described in claim 1, characterized in that: The construction of the dynamic knowledge graph specifically includes the following steps: using a spatial embedding algorithm to perform correlation mapping on the three-dimensional spatial coordinate parameters to form nodes of the topological structure; using the HTTP protocol to perform attribute mapping on the data exchanged between devices to obtain the edges of the topological structure; and dynamically assigning weights and integrating the edges and nodes of the topological structure to generate a dynamic knowledge graph.

4. The control method for remote monitoring of the Internet of Things as described in claim 1, characterized in that: The acquisition of risk nodes specifically includes the following steps: injecting the abnormal propagation path into the dynamic knowledge graph through the Apache Flink pipeline, and applying an improved graph convolutional fusion algorithm to perform neighborhood aggregation on the dynamic knowledge graph to form a fault association enhancement graph. Based on the fault association enhancement graph, the fault impact weights between nodes are updated to generate nodes with propagation risk.

5. The control method for remote monitoring of the Internet of Things as described in claim 1, characterized in that: The formation of the multi-level remote monitoring scheme specifically includes the following steps: using the LeaderRank algorithm to dynamically sort the risk propagation nodes and generate a risk propagation priority list; and using spectral clustering to perform community clustering analysis on the risk propagation priority list to obtain the high-risk set. A hierarchical spatial mapping is performed on high-risk sets to form a multi-level remote monitoring scheme.

6. The control method for remote monitoring of the Internet of Things as described in claim 5, characterized in that: The incremental update of the spatiotemporal graph convolutional model specifically includes the following steps: converting the multi-level remote monitoring scheme into control commands via the Modbus protocol, and driving the PID controller through the Internet of Things to execute equipment parameter regulation; simultaneously collecting equipment operation feedback data using the Internet of Things gateway and inputting it into the spatiotemporal graph convolutional model; using gated causal convolution to calculate the anomaly probability of the equipment operation feedback data, obtaining the anomaly propagation probability distribution, and performing gradient backpropagation on the anomaly propagation probability distribution to generate updated gradient parameters; based on the updated gradient parameters, incrementally updating the convolution kernel parameters of the spatiotemporal graph convolutional model, and outputting the updated spatiotemporal graph convolutional model.

7. A control system for IoT remote monitoring, based on the IoT remote monitoring control method according to any one of claims 1 to 6, characterized in that: It includes an anomaly analysis module, a graph construction module, a solution generation module, and a closed-loop optimization module. The anomaly analysis module is used to input the device spatiotemporal operation dataset into the spatiotemporal graph convolutional model, extract local spatiotemporal features at the edge layer, and perform cross-device collaborative mode analysis at the cloud layer to generate anomaly propagation paths. The graph construction module is used to map the three-dimensional spatial coordinate parameters in the IoT device operation dataset into nodes of the topological structure, and at the same time map the interaction data between devices into edges of the topological structure to build a dynamic knowledge graph. The scheme generation module injects the abnormal propagation path into the dynamic knowledge graph, applies an improved graph convolutional fusion algorithm to update the fault impact weights between nodes, obtains propagation risk nodes, and performs dynamic sorting and community clustering analysis on the propagation risk nodes to form a multi-level remote monitoring scheme. The closed-loop optimization module converts the multi-level remote monitoring scheme into control commands, drives the edge execution unit of the Internet of Things to adjust device parameters, synchronously collects device operation feedback data, and incrementally updates the spatiotemporal graph convolutional model.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the control method for IoT remote monitoring according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the control method for IoT remote monitoring as described in any one of claims 1 to 6.

Images