An edge-computing-based internet-of-things multi-node data fusion method

Abstract

This invention discloses a multi-node data fusion method for the Internet of Things (IoT) based on edge computing, comprising the following steps: Step 1: collecting raw monitoring data from the IoT system; Step 2: performing preprocessing to generate standardized feature tensors; Step 3: deploying edge computing nodes, dividing and allocating the standardized feature tensors to the corresponding edge computing nodes; Step 4: extracting time and periodic enhancement features through an improved TimesNet network to generate frequency domain enhanced feature tensors; Step 5: performing node set fusion through an improved Set Transformer network to generate node fusion feature vectors; Step 6: performing edge intelligent inference to generate sensor node state prediction vectors, edge risk scores, and control execution command vectors; Step 7: collecting control execution feedback data and updating the edge intelligent inference process. This invention improves the efficiency of multi-node fusion in the IoT and the accuracy of edge decision-making.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IoT) and edge computing data processing technology, and in particular to a multi-node data fusion method for IoT based on edge computing. Background Technology

[0002] With the widespread application of Industrial Internet of Things (IIoT) and edge intelligent control technologies, multi-node monitoring and distributed data fusion play a core role in smart factories, smart buildings, and energy transmission and distribution. Existing multi-node data acquisition and fusion systems generally rely on centralized cloud-based modeling strategies. While they possess global processing capabilities, they still face the following challenges in high-frequency, large-scale, and spatially distributed sensor network scenarios: The sheer number of sensor nodes and inconsistent data sampling frequencies lead to significant gaps, misalignments, and fluctuations in the original monitoring data over time, making it difficult to generate high-quality input features. Long communication links and limited bandwidth at central nodes cause cloud congestion due to frequent data uploads, severely restricting the response speed and real-time control capabilities of IoT systems. Existing methods often ignore the spatial topology and geographical structure characteristics between nodes, and fusion strategies frequently rely on static rules or fixed weights, failing to effectively model the dynamic changes in spatial correlation and risk contribution of data from different nodes. Iterative updates of cloud-based intelligent models depend on offline training, making it difficult to incorporate feedback data from actual operation for online edge-side optimization, thus limiting the adaptability and robustness of IoT systems.

[0003] Therefore, how to provide a multi-node data fusion method for the Internet of Things based on edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a multi-node data fusion method for the Internet of Things (IoT) based on edge computing. This invention combines an improved TimesNet network and an improved Set Transformer network to construct a collaborative structure for multi-scale temporal modeling and node set fusion, and deploys an end-edge-cloud collaborative data processing architecture. This achieves the entire process from collecting raw monitoring data, standardization processing, time-series-frequency domain enhanced modeling, node fusion, edge intelligent inference, to feedback updates. The method of this invention possesses advantages such as strong data processing capability, high fusion accuracy, good real-time edge response, and strong model adaptability.

[0005] According to an embodiment of the present invention, a multi-node data fusion method for the Internet of Things based on edge computing includes the following steps: Step 1: Collect raw monitoring data from the IoT system to form the raw monitoring dataset; Step 2: Perform preprocessing on the original monitoring dataset to generate a standardized feature tensor; Step 3: Deploy edge computing nodes. Based on the geographical location and communication latency of the sensor nodes, divide and allocate the standardized feature tensor to the corresponding edge computing nodes. Step 4: Extract time and period enhancement features from the standardized feature tensor through an improved TimesNet network to generate a frequency domain enhancement feature tensor. The improved TimesNet network introduces multi-scale dynamic convolution, frequency domain segmented fusion, and main frequency period residual. Step 5: Input the frequency domain enhanced feature tensor into the improved Set Transformer network to perform node set fusion. The improved Set Transformer network introduces dynamic induced point sampling based on node importance and node energy gating to generate node fusion feature vectors. Step 6: Perform edge intelligent reasoning based on node fusion feature vectors to generate sensor node state prediction vectors, edge risk scores, and control execution command vectors, and upload them to the cloud of the IoT system; Step 7: Collect control execution feedback data, calculate feedback confidence and generate joint confidence vector, and update the edge intelligent inference process.

[0006] Optionally, step one specifically includes: Several sensor nodes are deployed in the Internet of Things system. Each sensor node is bound to a node identifier and configured with a data collection timestamp. The types of sensor nodes include temperature sensors, pressure sensors, current sensors, vibration sensors, and displacement sensors. The raw monitoring data of each sensor node is collected and the node identifier, sampling frequency, node geographical location and node type are recorded. The raw monitoring data includes the temperature value, pressure value, current value, vibration amplitude value and displacement value of each sensor node. The raw monitoring data is organized according to time steps to construct the raw monitoring dataset.

[0007] Optionally, step two specifically includes: Outlier detection and removal are performed on the physical quantity values ​​in the original monitoring dataset. Specifically, if the difference between the physical quantity value of the sensing node at the current time step and the mean of the physical quantity exceeds three times the standard deviation of the physical quantity, the current physical quantity value is determined to be an outlier and removed. The physical quantity values ​​are temperature values, pressure values, current values, vibration amplitude values, or displacement values. Linear interpolation was used to fill in the missing physical quantity values ​​in the original monitoring dataset; The physical quantity values ​​of all sensor nodes are synchronously rearranged according to the time step. The Z-Score normalization method is used to normalize the physical quantity values ​​of each sensing node to obtain normalized physical quantity values; The standardized physical quantity values ​​of all sensing nodes at each time step are combined according to the sensing node dimension, time step dimension, and physical quantity value dimension to generate a standardized feature tensor.

[0008] Optionally, step four specifically includes: The standardized feature tensor is input into the improved TimesNet network for temporal feature extraction. The improved TimesNet network includes a node spatiotemporal rearrangement module, a multi-scale dynamic convolution module, a frequency domain segmentation and fusion module, and a periodic residual injection module. The node spatiotemporal rearrangement module expands the standardized feature tensor along the sensor node dimension to obtain a sequence of physical quantity feature vectors arranged by time steps, which constitutes a physical quantity feature matrix. Construct a node mapping matrix, which records the attribution relationship between feature dimensions and sensing nodes; The multi-scale dynamic convolution module introduces node attribute vectors to construct convolution kernel weights and outputs a scale-fused feature matrix, specifically: Three parallel one-dimensional convolution channels are set up, and the kernel sizes of the one-dimensional convolution channels are 3, 5 and 7 respectively; The node identifier, sampling frequency, node geographic location, and node type are concatenated into a node attribute vector, and the node attribute vector is input into three parallel one-layer MLP structures to generate small-scale convolution weight matrices, medium-scale convolution weight matrices, and large-scale convolution weight matrices. The standardized feature tensor is convolved with the small-scale, medium-scale, and large-scale convolution weight matrices in one dimension to obtain the small-scale convolution feature matrix, the medium-scale convolution feature matrix, and the large-scale convolution feature matrix, respectively. The small-scale convolutional feature matrix, the medium-scale convolutional feature matrix, and the large-scale convolutional feature matrix are concatenated along the feature dimension to generate a scale fusion feature matrix. The frequency domain segmentation and fusion module maps the physical quantity feature matrix to the frequency domain, extracts and models the period, trend, and disturbance information of different frequency bands, and generates a frequency domain enhanced feature matrix, specifically: Perform a fast Fourier transform on the physical quantity feature matrix to obtain the spectral feature matrix; Two segmented frequency thresholds are set, and the spectral feature matrix is ​​divided into a low-frequency band feature matrix, a mid-frequency band feature matrix, and a high-frequency band feature matrix along the frequency axis. Feature extraction and channel fusion are performed on the low-frequency band feature matrix, mid-frequency band feature matrix and high-frequency band feature matrix respectively to obtain the frequency band fused feature matrix; The periodic residual injection module performs spectral energy analysis on the frequency band fusion feature matrix, identifies the dominant periodic frequency, and extracts the spectral segment corresponding to the dominant periodic frequency to obtain the dominant frequency feature matrix. The dominant frequency feature matrix is ​​transformed into the dominant frequency projection feature matrix through inverse Fourier transform; Set the residual weight coefficients, multiply the residual weight coefficients element-wise with the main frequency projection feature matrix, and add them to the scale fusion feature matrix to obtain the frequency domain enhancement feature matrix; Based on the node mapping matrix, the frequency domain enhanced feature matrix is ​​reconstructed into a frequency domain enhanced feature tensor, which is a three-dimensional tensor structure.

[0009] Optionally, the step of performing feature extraction and channel fusion on the low-frequency band feature matrix, mid-frequency band feature matrix, and high-frequency band feature matrix respectively to obtain a frequency band fused feature matrix is ​​as follows: At each time step, average pooling is performed on the channel dimension of the low-frequency feature matrix to obtain a low-frequency trend vector sequence. The low-frequency trend vector sequence is then input into a single-layer MLP structure to perform nonlinear mapping, generating a low-frequency fused feature matrix. The MLP structure uses the GELU activation function. The mid-frequency feature matrix is ​​subjected to a one-dimensional convolution operation with a kernel of 3 along the time dimension to obtain a preliminary convolution feature matrix. The preliminary convolution feature matrix is ​​then added element-wise to the mid-frequency feature matrix to generate a residual feature matrix. The residual feature matrix is ​​then activated by the ReLU function to generate a mid-frequency fusion feature matrix. Max pooling is performed on the high-frequency feature matrix in the time dimension to obtain the maximum activation value of each channel, and the values ​​are combined into a channel description vector. The channel description vector is then converted into a channel attention vector through a single MLP structure, which uses the Sigmoid activation function. The channel attention vector is multiplied channel by channel with the high-frequency band feature matrix to obtain the high-frequency band fusion feature matrix; The low-frequency band fusion feature matrix, mid-frequency band fusion feature matrix, and high-frequency band fusion feature matrix are concatenated by channels and then dimensionality is reduced by one-dimensional convolution to obtain the frequency band fusion feature matrix.

[0010] Optionally, step five specifically includes: Calculate the node importance score of all sensing nodes. Specifically, based on the frequency domain enhanced feature tensor, obtain the frequency domain enhanced feature vector of each sensing node at each time step, and calculate the ratio of the L2 norm of each frequency domain enhanced feature vector to the L2 norm of all frequency domain enhanced feature vectors. Based on the node importance scores, Top-M sampling is performed on the frequency domain enhanced feature vectors of all sensing nodes to obtain the induced point matrix; Combine the induced point matrices of all time steps to form the induced point tensor; At each time step, the frequency domain enhanced feature vector of each sensing node is linearly reduced and activated by tanh to generate node energy factors, and the node energy factors are then used to generate node gating weights through affine transformation and sigmoid activation. The node gating weights are multiplied element-wise with the corresponding frequency domain enhanced feature vectors to obtain the gated node feature vectors. The gated node feature vectors of all time steps and sensor nodes are combined to generate the gated node feature tensor. The inducing point tensor is mapped to the inducing query tensor through a trainable linear mapping matrix, and the gated node feature tensor is mapped to the key tensor and value tensor through two sets of trainable linear mapping matrices respectively. Perform a tensor dot product operation between the induced query tensor and the key tensor, divide by the square root of the key tensor dimension, normalize by softmax to generate an attention score tensor, and perform a tensor multiplication operation between the attention score tensor and the value tensor to obtain the node attention feature tensor. The node attention feature tensor is subjected to average pooling along the time dimension and the induction point dimension to obtain the node fusion feature vector.

[0011] Optionally, step six specifically includes: The node fusion feature vector is transformed into a mapped feature vector through linear mapping. The mapped feature vector is then passed through a fully connected layer and a Softmax activation function to generate a sensor node state prediction vector. Finally, the mapped feature vector is passed through a fully connected layer and a Sigmoid activation function to generate an edge risk score. The action trigger flag value is generated based on the edge risk score, specifically as follows: Set a risk threshold. If the risk score is greater than the risk threshold, the action trigger flag is set to 1, indicating that the edge control command generation stage has begun. If the edge risk score is less than or equal to the risk threshold, the action trigger flag is set to 0, indicating that the current sensor node operation status is maintained. When the action trigger flag value is equal to 1, the mapped feature vector is passed through a fully connected layer and a tanh activation function to generate a control execution instruction vector; each dimension value of the control execution instruction vector represents the output instruction value of the control execution unit; The sensor node state prediction vector, edge risk score, control execution command vector, and action trigger flag value are uploaded to the cloud of the Internet of Things system.

[0012] Optionally, step seven specifically includes: The control execution feedback data includes the actual feedback state vector of the sensor node, the feedback risk level value, and the control response feedback vector. Calculate the L2 norm of the vector difference between the actual feedback state vector and the state prediction vector of the sensing node to obtain the state prediction error, and calculate the state prediction confidence using a negative exponential function. Calculate the difference between the marginal risk score and the feedback risk level value to obtain the risk score deviation. Map the risk score deviation to the [0,1] interval through the Sigmoid function to obtain the normalized score deviation. Subtract the normalized score deviation from 1 to obtain the risk score confidence. Calculate the L1 norm of the vector difference between the control execution command vector and the control response feedback vector to obtain the control execution deviation. Add the control execution deviation to 1 and take the reciprocal to obtain the control command confidence. The confidence scores of state prediction, risk score, and control command are combined to form a joint confidence vector, which is then concatenated with the node fusion feature vector to continue edge intelligent inference.

[0013] The beneficial effects of this invention are: This invention provides a multi-node data fusion method for the Internet of Things (IoT) based on edge computing, which can effectively improve the fusion efficiency, prediction accuracy, and control response capability of multi-node data in IoT systems. First, this invention deploys several sensor nodes in the IoT system to collect various types of raw monitoring data, such as temperature, pressure, current, vibration amplitude, and displacement. By configuring the collection timestamp, node identifier, sampling frequency, and geographical location, a well-structured and time-aligned raw monitoring dataset is ensured. Preprocessing operations are performed to remove outliers, impute missing values, and standardize the data, effectively improving the data quality for modeling and generating a high-dimensional, structurally consistent standardized feature tensor.

[0014] Secondly, this invention constructs an edge computing task partitioning strategy based on geographical location and communication latency. It divides standardized feature tensors into multiple sub-tensors and assigns them to corresponding edge computing nodes, achieving the localization of data processing tasks and spatial distribution of computational load. Through a collaborative topology of edge-side data acquisition, edge-side feature caching and preliminary processing, and cloud-based model training and global control, it effectively reduces the communication pressure and response latency of the central node, providing support for real-time fusion and inference in complex distributed scenarios.

[0015] Furthermore, this invention constructs a time-series modeling structure based on an improved TimesNet network, integrating multi-scale dynamic convolution, frequency domain segmented fusion, and dominant frequency periodic residuals. This structure can accurately extract periodic, trend, and perturbation features at different time scales and frequency levels, significantly enhancing the expressive power and robustness of time-series modeling. Combined with an improved SetTransformer network, a dynamic induced point sampling mechanism based on node importance and a node energy gating mechanism are introduced to achieve global perception and cross-fusion of node set features based on frequency domain enhanced feature tensors, effectively improving the integrated expressive power of spatiotemporal features among different nodes. In edge intelligent inference, based on the node fusion feature vector, the system outputs sensor node state prediction vectors, edge risk scores, and control execution command vectors, exhibiting good interpretability and real-time performance. By collecting control execution feedback data, the system calculates the confidence scores of state prediction, risk scores, and control commands, and constructs a joint confidence vector injection inference process to achieve continuous optimization and online update mechanisms at the edge.

[0016] In summary, the method of this invention constructs a fusion process and edge computing closed-loop mechanism for multi-node IoT scenarios, improves data fusion accuracy and inference efficiency, strengthens the adaptive processing capability for dynamic environmental changes, and has the advantages of high processing efficiency, low response latency, adaptive model updates, and high fusion accuracy. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a multi-node data fusion method for the Internet of Things based on edge computing proposed in this invention; Figure 2 This is a flowchart of the improved TimesNet network structure in an IoT multi-node data fusion method based on edge computing proposed in this invention. Figure 3 This is a flowchart of the improved Set Transformer network structure in an edge computing-based multi-node data fusion method for the Internet of Things proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A multi-node data fusion method for the Internet of Things based on edge computing includes the following steps: Step 1: Collect raw monitoring data from the IoT system to form the raw monitoring dataset; Step 2: Perform preprocessing on the original monitoring dataset to generate a standardized feature tensor; Step 3: Deploy edge computing nodes. Based on the geographical location and communication latency of the sensor nodes, divide and allocate the standardized feature tensor to the corresponding edge computing nodes. In this invention, several edge computing nodes are deployed at the edge computing layer of the Internet of Things (IoT) system. Standardized feature tensors are divided into several sub-tensors according to the geographic location and communication latency of the sensing nodes, and each sub-tensor is assigned to a corresponding edge computing node. This enables the localization of data processing tasks and the spatial distribution of computational load, effectively reducing the communication pressure and response latency of the central node, and improving the processing efficiency of multi-node data fusion and the real-time response capability of the IoT system. The edge computing nodes are used to perform data caching and preliminary fusion calculations of the sub-tensors. Simultaneously, an edge-cloud collaborative data processing topology is constructed, where the edge side is responsible for the collection and reporting of raw monitoring data, the edge side is responsible for the caching and preprocessing calculations of the sub-tensors, and the cloud side is responsible for global fusion and model training, realizing a multi-layered, collaborative, efficient, and stable data fusion processing mechanism.

[0020] Step 4: Extract time and period enhancement features from the standardized feature tensor through an improved TimesNet network to generate a frequency domain enhancement feature tensor. The improved TimesNet network introduces multi-scale dynamic convolution, frequency domain segmented fusion, and main frequency period residual. Step 5: Input the frequency domain enhanced feature tensor into the improved Set Transformer network to perform node set fusion. The improved Set Transformer network introduces dynamic induced point sampling based on node importance and node energy gating to generate node fusion feature vectors. Step 6: Perform edge intelligent reasoning based on node fusion feature vectors to generate sensor node state prediction vectors, edge risk scores, and control execution command vectors, and upload them to the cloud of the IoT system; Step 7: Collect control execution feedback data, calculate feedback confidence and generate joint confidence vector, and update the edge intelligent inference process.

[0021] In this embodiment, step one specifically includes: Several sensor nodes are deployed in the Internet of Things system. Each sensor node is bound to a node identifier and configured with a data collection timestamp. The types of sensor nodes include temperature sensors, pressure sensors, current sensors, vibration sensors, and displacement sensors. The raw monitoring data of each sensor node is collected and the node identifier, sampling frequency, node geographical location and node type are recorded. The raw monitoring data includes the temperature value, pressure value, current value, vibration amplitude value and displacement value of each sensor node. The raw monitoring data is organized according to time steps to construct the raw monitoring dataset.

[0022] In this embodiment, step two specifically includes: Outlier detection and removal are performed on the physical quantity values ​​in the original monitoring dataset. Specifically, if the difference between the physical quantity value of the sensing node at the current time step and the mean of the physical quantity exceeds three times the standard deviation of the physical quantity, the current physical quantity value is determined to be an outlier and removed. The physical quantity values ​​are temperature values, pressure values, current values, vibration amplitude values, or displacement values. Linear interpolation was used to fill in the missing physical quantity values ​​in the original monitoring dataset; The physical quantity values ​​of all sensor nodes are synchronously rearranged according to the time step. The Z-Score normalization method is used to normalize the physical quantity values ​​of each sensing node to obtain normalized physical quantity values; The standardized physical quantity values ​​of all sensing nodes at each time step are combined according to the sensing node dimension, time step dimension, and physical quantity value dimension to generate a standardized feature tensor.

[0023] In this embodiment, step four specifically includes: The standardized feature tensor is input into the improved TimesNet network for temporal feature extraction. The improved TimesNet network includes a node spatiotemporal rearrangement module, a multi-scale dynamic convolution module, a frequency domain segmentation and fusion module, and a periodic residual injection module. The node spatiotemporal rearrangement module expands the standardized feature tensor along the sensor node dimension to obtain a sequence of physical quantity feature vectors arranged by time steps, which constitutes a physical quantity feature matrix. Construct a node mapping matrix, which records the attribution relationship between feature dimensions and sensing nodes, and retains node structure information while rearranging the standardized feature tensors into physical quantity feature matrices. In this invention, the three-dimensional normalized feature tensor is unfolded over time steps and reorganized into a two-dimensional physical quantity feature matrix with time as the main axis. This unifies the modeling of multi-node physical quantity features at each time step, facilitating the mining of time series features through multi-scale convolution operations. Simultaneously, to preserve the structural correspondence between each feature dimension and its corresponding sensing node during the rearrangement process, a node mapping matrix is ​​constructed to record the node identifiers to which the feature dimensions belong. This enables information tracking and enhancement based on node structure in subsequent modeling, helping to maintain spatial consistency and semantic isolation between nodes during the time modeling process and preventing the loss of structural information caused by feature rearrangement.

[0024] The multi-scale dynamic convolution module introduces node attribute vectors to construct convolution kernel weights and outputs a scale-fused feature matrix, specifically: Three parallel one-dimensional convolution channels are set up, and the kernel sizes of the one-dimensional convolution channels are 3, 5 and 7 respectively; The node identifier, sampling frequency, node geographic location, and node type are concatenated into a node attribute vector, and the node attribute vector is input into three parallel one-layer MLP structures to generate small-scale convolution weight matrices, medium-scale convolution weight matrices, and large-scale convolution weight matrices. In this invention, dynamic convolution weight matrices of small, medium and large scales are generated by node attribute vectors, which can improve the expressive power and modeling accuracy of the improved TimesNet network when processing heterogeneous node data, and effectively enhance the multi-scale perception capability and node structure sensitivity of time feature extraction.

[0025] The standardized feature tensor is convolved with the small-scale, medium-scale, and large-scale convolution weight matrices in one dimension to obtain the small-scale convolution feature matrix, the medium-scale convolution feature matrix, and the large-scale convolution feature matrix, respectively. The small-scale convolutional feature matrix, the medium-scale convolutional feature matrix, and the large-scale convolutional feature matrix are concatenated along the feature dimension to generate a scale fusion feature matrix. The frequency domain segmentation and fusion module maps the physical quantity feature matrix to the frequency domain, extracts and models the period, trend, and disturbance information of different frequency bands, and generates a frequency domain enhanced feature matrix, specifically: Perform a fast Fourier transform on the physical quantity feature matrix to obtain the spectral feature matrix; Two segmented frequency thresholds are set, and the spectral feature matrix is ​​divided into a low-frequency band feature matrix, a mid-frequency band feature matrix, and a high-frequency band feature matrix along the frequency axis. Feature extraction and channel fusion are performed on the low-frequency band feature matrix, mid-frequency band feature matrix and high-frequency band feature matrix respectively to obtain the frequency band fused feature matrix; In this invention, by mapping the physical quantity feature matrix to the frequency domain, and dividing the spectral feature matrix into low-frequency, mid-frequency, and high-frequency bands based on two set segmented frequency thresholds, important information such as period, trend, and disturbance is extracted from different frequency bands respectively. Feature extraction and channel fusion are then performed to generate a frequency band fusion feature matrix, thereby achieving accurate modeling and fusion of multi-scale frequency features. This significantly enhances the sensitivity of the improved TimesNet network to periodic changes and abnormal disturbances, and effectively improves the support capability of frequency domain information for node state modeling and inference tasks.

[0026] The periodic residual injection module performs spectral energy analysis on the frequency band fusion feature matrix, identifies the dominant periodic frequency, and extracts the spectral segment corresponding to the dominant periodic frequency to obtain the dominant frequency feature matrix. The dominant frequency feature matrix is ​​transformed into the dominant frequency projection feature matrix through inverse Fourier transform; In this invention, by performing spectral energy analysis on the frequency band fusion feature matrix, the dominant periodic frequency is identified and its corresponding spectral segment is extracted. Then, the dominant frequency projection feature matrix is ​​obtained through inverse Fourier transform. This process effectively strengthens the explicit expression of the dominant period in the time domain features, which helps the improved TimesNet network capture the time domain response features of key periodic behavior patterns and abnormal disturbances, and improves the accuracy of periodic change modeling and the ability to predict periodic-driven state evolution.

[0027] Set the residual weight coefficients, multiply the residual weight coefficients element-wise with the main frequency projection feature matrix, and add them to the scale fusion feature matrix to obtain the frequency domain enhancement feature matrix; Based on the node mapping matrix, the frequency domain enhanced feature matrix is ​​reconstructed into a frequency domain enhanced feature tensor, which is a three-dimensional tensor structure.

[0028] In this embodiment, the step of performing feature extraction and channel fusion on the low-frequency band feature matrix, mid-frequency band feature matrix, and high-frequency band feature matrix respectively to obtain the frequency band fused feature matrix is ​​as follows: At each time step, average pooling is performed on the channel dimension of the low-frequency feature matrix to obtain a low-frequency trend vector sequence. The low-frequency trend vector sequence is then input into a single-layer MLP structure to perform nonlinear mapping, generating a low-frequency fused feature matrix. The MLP structure uses the GELU activation function. The mid-frequency feature matrix is ​​subjected to a one-dimensional convolution operation with a kernel of 3 along the time dimension to obtain a preliminary convolution feature matrix. The preliminary convolution feature matrix is ​​then added element-wise to the mid-frequency feature matrix to generate a residual feature matrix. The residual feature matrix is ​​then activated by the ReLU function to generate a mid-frequency fusion feature matrix. Max pooling is performed on the high-frequency feature matrix in the time dimension to obtain the maximum activation value of each channel, and the values ​​are combined into a channel description vector. The channel description vector is then converted into a channel attention vector through a single MLP structure, which uses the Sigmoid activation function. The channel attention vector is multiplied channel by channel with the high-frequency band feature matrix to obtain the high-frequency band fusion feature matrix; The low-frequency band fusion feature matrix, mid-frequency band fusion feature matrix, and high-frequency band fusion feature matrix are concatenated by channels and then dimensionality is reduced by one-dimensional convolution to obtain the frequency band fusion feature matrix.

[0029] In this embodiment, step five specifically includes: Calculate the node importance score of all sensing nodes. Specifically, based on the frequency domain enhanced feature tensor, obtain the frequency domain enhanced feature vector of each sensing node at each time step, and calculate the ratio of the L2 norm of each frequency domain enhanced feature vector to the L2 norm of all frequency domain enhanced feature vectors. Based on the node importance scores, Top-M sampling is performed on the frequency domain enhanced feature vectors of all sensing nodes to obtain the induced point matrix; Combine the induced point matrices of all time steps to form the induced point tensor; At each time step, the frequency domain enhanced feature vector of each sensing node is linearly reduced and activated by tanh to generate node energy factors, and the node energy factors are then used to generate node gating weights through affine transformation and sigmoid activation. The node gating weights are multiplied element-wise with the corresponding frequency domain enhanced feature vectors to obtain the gated node feature vectors. The gated node feature vectors of all time steps and sensor nodes are combined to generate the gated node feature tensor. The inducing point tensor is mapped to the inducing query tensor through a trainable linear mapping matrix, and the gated node feature tensor is mapped to the key tensor and value tensor through two sets of trainable linear mapping matrices respectively. Perform a tensor dot product operation between the induced query tensor and the key tensor, divide by the square root of the key tensor dimension, normalize by softmax to generate an attention score tensor, and perform a tensor multiplication operation between the attention score tensor and the value tensor to obtain the node attention feature tensor. The node attention feature tensor is subjected to average pooling along the time dimension and the induction point dimension to obtain the node fusion feature vector.

[0030] In this invention, node importance scores are obtained by calculating the L2 norm proportion of the frequency domain enhanced feature vector. Based on these scores, Top-M sampling is performed to construct an induced point tensor. Simultaneously, a gated node feature tensor is generated by combining node energy factors. Furthermore, an attention mechanism between the induced query tensor and the key-value tensor of the gated node features is utilized to extract global key features. Finally, a node fusion feature vector is generated through average pooling across time and node dimensions. This method, while fusing key node information, suppresses redundant and noisy features, significantly enhancing the model's selective attention to heterogeneous information from multiple nodes and its global modeling ability, thereby improving the prediction accuracy and decision stability in the edge intelligent inference stage.

[0031] In this embodiment, step six specifically includes: The node fusion feature vector is transformed into a mapped feature vector through linear mapping. The mapped feature vector is then passed through a fully connected layer and a Softmax activation function to generate a sensor node state prediction vector. Finally, the mapped feature vector is passed through a fully connected layer and a Sigmoid activation function to generate an edge risk score. The action trigger flag value is generated based on the edge risk score, specifically as follows: Set a risk threshold. If the risk score is greater than the risk threshold, the action trigger flag is set to 1, indicating that the edge control command generation stage has begun. If the edge risk score is less than or equal to the risk threshold, the action trigger flag is set to 0, indicating that the current sensor node operation status is maintained. When the action trigger flag value is equal to 1, the mapped feature vector is passed through a fully connected layer and a tanh activation function to generate a control execution instruction vector; each dimension value of the control execution instruction vector represents the output instruction value of the control execution unit; The sensor node state prediction vector, edge risk score, control execution command vector, and action trigger flag value are uploaded to the cloud of the Internet of Things system.

[0032] In this embodiment, step seven specifically includes: The control execution feedback data includes the actual feedback state vector of the sensor node, the feedback risk level value, and the control response feedback vector. Calculate the L2 norm of the vector difference between the actual feedback state vector and the state prediction vector of the sensing node to obtain the state prediction error, and calculate the state prediction confidence using a negative exponential function. Calculate the difference between the marginal risk score and the feedback risk level value to obtain the risk score deviation. Map the risk score deviation to the [0,1] interval through the Sigmoid function to obtain the normalized score deviation. Subtract the normalized score deviation from 1 to obtain the risk score confidence. Calculate the L1 norm of the vector difference between the control execution command vector and the control response feedback vector to obtain the control execution deviation. Add the control execution deviation to 1 and take the reciprocal to obtain the control command confidence. The confidence scores of state prediction, risk score, and control command are combined to form a joint confidence vector, which is then concatenated with the node fusion feature vector to continue edge intelligent inference.

[0033] Example 1 To verify the feasibility of this invention in practice, it was applied to the multi-node data fusion and edge intelligent risk control task of an environmental monitoring system in a certain industrial park. The park deploys multiple sensor substations to collect environmental and equipment operating status data such as gas concentration, temperature and humidity, electrical equipment operating voltage, current, vibration frequency, and partial discharge. The deployment of these sensor nodes exhibits strong spatial differences, and communication links between some sensor nodes are unstable, resulting in typical IoT deployment problems such as data latency and limited local computing resources.

[0034] In the actual scenario, a total of 56 sensor substations were deployed within the park, each collecting approximately eight indicators at a sampling frequency of 1 Hz per second. The raw data volume was enormous and exhibited obvious periodic fluctuations, such as high operating intensity during the day and lower intensity at night, and frequent occurrences of increased humidity and decreased electrical load during rainy days. Traditional data aggregation architectures experienced significant latency during peak transmission periods, making it difficult to provide efficient early warnings and responses at the edge, severely impacting the real-time performance and reliability of monitoring.

[0035] During implementation, raw monitoring data is first reported by each sensor substation to the edge computing node. An improved TimesNet network performs multi-scale temporal modeling and periodic enhancement modeling on the standardized feature tensor. Through frequency domain segmentation fusion and periodic residual injection mechanisms, the feature structure of the sensor nodes under multiple time scales and dominant periods is extracted, forming a frequency domain enhanced feature tensor. Subsequently, the frequency domain enhanced feature tensor is input into an improved Set Transformer network that incorporates induced point sampling and energy gating mechanisms to extract key node set features and output the fused node vector. In the edge intelligent inference module, a lightweight neural network performs state prediction, risk scoring, and control command generation on the fused node feature vector. Once the risk score exceeds a threshold, an alarm is immediately triggered, and the control command is simultaneously fed back to the field equipment through the execution module.

[0036] To evaluate the performance of the method of this invention, real park operation data was selected for comparative experiments. The comparative methods included: Method A: traditional centralized modeling, where all data was uploaded to the cloud before modeling; and Method B: a lightweight model based solely on single-point prediction, without node fusion. The experimental results are shown in Table 1.

[0037] Table 1. Performance Comparison of the Invention Method and Comparative Methods in Industrial Park Monitoring Scenarios

[0038] As shown in Table 1, the method of this invention outperforms the comparative methods in all indicators. In terms of prediction mean square error, the method of this invention is 0.047, significantly lower than the 0.124 of comparative method A and 0.098 of comparative method B, indicating that the method of this invention has a stronger ability to model multi-period signals. The anomaly detection accuracy of the method of this invention reaches 96.8%, significantly improving the accuracy of risk perception. The false alarm rate of the system alarm is controlled below 1.5%, demonstrating good discrimination ability of node fusion and edge inference. The method of this invention also performs outstandingly in terms of response speed, with peak response latency and control command latency reduced to 247ms and 311ms respectively, meeting the basic real-time requirements of industrial sites. The network data transmission volume is significantly reduced from 217MB / day of comparative method A to 38MB / day, greatly alleviating network bandwidth pressure and making it suitable for data-intensive distributed deployment scenarios. The method of this invention not only effectively solves the problems of difficulty in multi-node data fusion, untimely response, and high central computing load, but also demonstrates superior prediction accuracy, system robustness, and resource efficiency in real-world deployments.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-node data fusion method for the Internet of Things based on edge computing, characterized in that, The steps include the following: Step 1: Collect raw monitoring data from the IoT system to form the raw monitoring dataset; Step 2: Perform preprocessing on the original monitoring dataset to generate a standardized feature tensor; Step 3: Deploy edge computing nodes. Based on the geographical location and communication latency of the sensor nodes, divide and allocate the standardized feature tensor to the corresponding edge computing nodes. Step 4: Extract time and period enhancement features from the standardized feature tensor through an improved TimesNet network to generate a frequency domain enhancement feature tensor. The improved TimesNet network introduces multi-scale dynamic convolution, frequency domain segmented fusion, and main frequency period residual. Step 5: Input the frequency domain enhanced feature tensor into the improved Set Transformer network to perform node set fusion. The improved Set Transformer network introduces dynamic induced point sampling based on node importance and node energy gating to generate node fusion feature vectors. Step 6: Perform edge intelligent reasoning based on node fusion feature vectors to generate sensor node state prediction vectors, edge risk scores, and control execution command vectors, and upload them to the cloud of the IoT system; Step 7: Collect control execution feedback data, calculate feedback confidence and generate joint confidence vector, and update the edge intelligent inference process.

2. The IoT multi-node data fusion method based on edge computing according to claim 1, characterized in that, Step one specifically includes: Several sensor nodes are deployed in the Internet of Things system. Each sensor node is bound to a node identifier and configured with a data collection timestamp. The types of sensor nodes include temperature sensors, pressure sensors, current sensors, vibration sensors, and displacement sensors. The raw monitoring data of each sensor node is collected and the node identifier, sampling frequency, node geographical location and node type are recorded. The raw monitoring data includes the temperature value, pressure value, current value, vibration amplitude value and displacement value of each sensor node. The raw monitoring data is organized according to time steps to construct the raw monitoring dataset.

3. The IoT multi-node data fusion method based on edge computing according to claim 1, characterized in that, Step two specifically includes: Outlier detection and removal are performed on the physical quantity values ​​in the original monitoring dataset. Specifically, if the difference between the physical quantity value of the sensing node at the current time step and the mean of the physical quantity exceeds three times the standard deviation of the physical quantity, the current physical quantity value is determined to be an outlier and removed. The physical quantity values ​​are temperature values, pressure values, current values, vibration amplitude values, or displacement values. Linear interpolation was used to fill in the missing physical quantity values ​​in the original monitoring dataset; The physical quantity values ​​of all sensor nodes are synchronously rearranged according to the time step. The Z-Score normalization method is used to normalize the physical quantity values ​​of each sensing node to obtain normalized physical quantity values; The standardized physical quantity values ​​of all sensing nodes at each time step are combined according to the sensing node dimension, time step dimension, and physical quantity value dimension to generate a standardized feature tensor.

4. The IoT multi-node data fusion method based on edge computing according to claim 1, characterized in that, Step four specifically includes: The standardized feature tensor is input into the improved TimesNet network for temporal feature extraction. The improved TimesNet network includes a node spatiotemporal rearrangement module, a multi-scale dynamic convolution module, a frequency domain segmentation and fusion module, and a periodic residual injection module. The node spatiotemporal rearrangement module expands the standardized feature tensor along the sensor node dimension to obtain a sequence of physical quantity feature vectors arranged by time steps, which constitutes a physical quantity feature matrix. Construct a node mapping matrix, which records the attribution relationship between feature dimensions and sensing nodes; The multi-scale dynamic convolution module introduces node attribute vectors to construct convolution kernel weights and outputs a scale-fused feature matrix, specifically: Three parallel one-dimensional convolution channels are set up, and the kernel sizes of the one-dimensional convolution channels are 3, 5 and 7 respectively; The node identifier, sampling frequency, node geographic location, and node type are concatenated into a node attribute vector, and the node attribute vector is input into three parallel one-layer MLP structures to generate small-scale convolution weight matrices, medium-scale convolution weight matrices, and large-scale convolution weight matrices. The standardized feature tensor is convolved with the small-scale, medium-scale, and large-scale convolution weight matrices in one dimension to obtain the small-scale convolution feature matrix, the medium-scale convolution feature matrix, and the large-scale convolution feature matrix, respectively. The small-scale convolutional feature matrix, the medium-scale convolutional feature matrix, and the large-scale convolutional feature matrix are concatenated along the feature dimension to generate a scale fusion feature matrix. The frequency domain segmentation and fusion module maps the physical quantity feature matrix to the frequency domain, extracts and models the period, trend, and disturbance information of different frequency bands, and generates a frequency domain enhanced feature matrix, specifically: Perform a fast Fourier transform on the physical quantity feature matrix to obtain the spectral feature matrix; Two segmented frequency thresholds are set, and the spectral feature matrix is ​​divided into a low-frequency band feature matrix, a mid-frequency band feature matrix, and a high-frequency band feature matrix along the frequency axis. Feature extraction and channel fusion are performed on the low-frequency band feature matrix, the mid-frequency band feature matrix and the high-frequency band feature matrix respectively to obtain the frequency band fused feature matrix; The periodic residual injection module performs spectral energy analysis on the frequency band fusion feature matrix, identifies the dominant periodic frequency, and extracts the spectral segment corresponding to the dominant periodic frequency to obtain the dominant frequency feature matrix. The dominant frequency feature matrix is ​​transformed into the dominant frequency projection feature matrix through inverse Fourier transform; Set the residual weight coefficients, multiply the residual weight coefficients element-wise with the main frequency projection feature matrix, and add them to the scale fusion feature matrix to obtain the frequency domain enhancement feature matrix; Based on the node mapping matrix, the frequency domain enhanced feature matrix is ​​reconstructed into a frequency domain enhanced feature tensor, which is a three-dimensional tensor structure.

5. The IoT multi-node data fusion method based on edge computing according to claim 4, characterized in that, The process involves performing feature extraction and channel fusion on the low-frequency, mid-frequency, and high-frequency feature matrices respectively to obtain a frequency band fused feature matrix. Specifically: At each time step, average pooling is performed on the channel dimension of the low-frequency feature matrix to obtain a low-frequency trend vector sequence. The low-frequency trend vector sequence is then input into a single-layer MLP structure to perform nonlinear mapping, generating a low-frequency fused feature matrix. The MLP structure uses the GELU activation function. The mid-frequency feature matrix is ​​subjected to a one-dimensional convolution operation with a kernel of 3 along the time dimension to obtain a preliminary convolution feature matrix. The preliminary convolution feature matrix is ​​then added element-wise to the mid-frequency feature matrix to generate a residual feature matrix. The residual feature matrix is ​​then activated by the ReLU function to generate a mid-frequency fusion feature matrix. Max pooling is performed on the high-frequency feature matrix in the time dimension to obtain the maximum activation value of each channel, and the values ​​are combined into a channel description vector. The channel description vector is then converted into a channel attention vector through a single MLP structure, which uses the Sigmoid activation function. The channel attention vector is multiplied channel by channel with the high-frequency band feature matrix to obtain the high-frequency band fusion feature matrix; The low-frequency band fusion feature matrix, mid-frequency band fusion feature matrix, and high-frequency band fusion feature matrix are concatenated by channels and then dimensionality is reduced by one-dimensional convolution to obtain the frequency band fusion feature matrix.

6. The IoT multi-node data fusion method based on edge computing according to claim 1, characterized in that, Step five specifically includes: Calculate the node importance score of all sensing nodes. Specifically, based on the frequency domain enhanced feature tensor, obtain the frequency domain enhanced feature vector of each sensing node at each time step, and calculate the ratio of the L2 norm of each frequency domain enhanced feature vector to the L2 norm of all frequency domain enhanced feature vectors. Based on the node importance scores, Top-M sampling is performed on the frequency domain enhanced feature vectors of all sensing nodes to obtain the induced point matrix; Combine the induced point matrices of all time steps to form the induced point tensor; At each time step, the frequency domain enhanced feature vector of each sensing node is linearly reduced and activated by tanh to generate node energy factors, and the node energy factors are then used to generate node gating weights through affine transformation and sigmoid activation. The node gating weights are multiplied element-wise with the corresponding frequency domain enhanced feature vectors to obtain the gated node feature vectors. The gated node feature vectors of all time steps and sensor nodes are combined to generate the gated node feature tensor. The inducing point tensor is mapped to the inducing query tensor through a trainable linear mapping matrix, and the gated node feature tensor is mapped to the key tensor and value tensor through two sets of trainable linear mapping matrices respectively. Perform a tensor dot product operation between the induced query tensor and the key tensor, divide by the square root of the key tensor dimension, normalize by softmax to generate an attention score tensor, and perform a tensor multiplication operation between the attention score tensor and the value tensor to obtain the node attention feature tensor. The node attention feature tensor is subjected to average pooling along the time dimension and the induction point dimension to obtain the node fusion feature vector.

7. The IoT multi-node data fusion method based on edge computing according to claim 1, characterized in that, Step six specifically includes: The node fusion feature vector is transformed into a mapped feature vector through linear mapping. The mapped feature vector is then passed through a fully connected layer and a Softmax activation function to generate a sensor node state prediction vector. Finally, the mapped feature vector is passed through a fully connected layer and a Sigmoid activation function to generate an edge risk score. The action trigger flag value is generated based on the edge risk score, specifically: Set a risk threshold. If the risk score is greater than the risk threshold, the action trigger flag is set to 1, indicating that the edge control command generation stage has begun. If the edge risk score is less than or equal to the risk threshold, the action trigger flag is set to 0, indicating that the current sensor node operation status is maintained. When the action trigger flag value is equal to 1, the mapped feature vector is passed through a fully connected layer and a tanh activation function to generate a control execution instruction vector; each dimension value of the control execution instruction vector represents the output instruction value of the control execution unit; The sensor node state prediction vector, edge risk score, control execution command vector, and action trigger flag value are uploaded to the cloud of the Internet of Things system.

8. The IoT multi-node data fusion method based on edge computing according to claim 1, characterized in that, Step seven specifically includes: The control execution feedback data includes the actual feedback state vector of the sensor node, the feedback risk level value, and the control response feedback vector. Calculate the L2 norm of the vector difference between the actual feedback state vector and the state prediction vector of the sensing node to obtain the state prediction error, and calculate the state prediction confidence using a negative exponential function. Calculate the difference between the marginal risk score and the feedback risk level value to obtain the risk score deviation. Map the risk score deviation to the [0,1] interval through the Sigmoid function to obtain the normalized score deviation. Subtract the normalized score deviation from 1 to obtain the risk score confidence. Calculate the L1 norm of the vector difference between the control execution command vector and the control response feedback vector to obtain the control execution deviation. Add the control execution deviation to 1 and take the reciprocal to obtain the control command confidence. The confidence scores of state prediction, risk score, and control command are combined to form a joint confidence vector, which is then concatenated with the node fusion feature vector to continue edge intelligent inference.

Images