GNN fusion neural network embedded intelligent gateway algorithm based on long and short time sequences

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a GNN fusion neural network embedded intelligent gateway algorithm based on long and short time sequences, which comprises the following steps of: 1, acquiring network performance time sequence data of a gateway; 2, preprocessing the network performance time sequence data to form standardized time sequence data; 3, inputting the standardized time series data into a pre-trained GNN model for spatial-temporal feature extraction to obtain fusion features; and 4, performing state recognition or abnormality diagnosis based on the fusion features. In the invention, the processing object of the algorithm is converted from the traditional environmental parameter time sequence data to the network performance time sequence data, and the algorithm can more accurately and directly judge whether the network is congested, whether the link is poor in quality or whether the network is attacked by directly analyzing essential indexes such as network flow, delay and packet loss rate. Therefore, misjudgment possibly caused by indirectly inferring the network problem only through the environment parameters is avoided, and the algorithm can predict the network performance trend.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to an embedded intelligent gateway algorithm based on a GNN fusion neural network with long and short time series. Background Technology

[0002] The GNN fusion neural network embedded intelligent gateway algorithm is a composite algorithm that combines graph neural networks, traditional neural networks, and other AI technologies. It is a high-efficiency data processing and decision-making solution designed specifically for embedded intelligent gateways.

[0003] A search revealed that invention patent CN202411159693.2 discloses an intelligent fusion gateway and its method. In its edge computing analysis module, it introduces artificial intelligence-based data processing and analysis algorithms to perform time-series dynamic analysis and collaborative correlation coding of environmental parameter data. This captures the time-series correlation features between temperature, humidity, smoke concentration, water accumulation status, and liquid level depth, thereby obtaining a time-series representation of the equipment status and effectively identifying and predicting the equipment status. In this way, the intelligent fusion gateway can effectively identify the environmental conditions of the equipment by utilizing multimodal data of environmental parameters, providing strong support for equipment fault early warning and response.

[0004] The aforementioned disclosed technologies place excessive emphasis on the time-series monitoring of external environmental parameters such as temperature, humidity, and smoke concentration, while completely neglecting the collection and analysis of network performance time-series data, such as traffic, latency, and packet loss rate, which are crucial to the smart gateway itself.

[0005] As stated in its background technology, "Traditional gateways...often ignore the environmental conditions in which the device operates," but their solutions go to the other extreme. This analytical paradigm, which relies solely on environmental data, has a fundamental limitation: it cannot directly perceive and diagnose the health status of the network itself, such as congestion, link failures, or security attacks. This leads to a serious disconnect between its decision-making and the gateway's core function of ensuring reliable and efficient network transmission, thus making it impossible to provide accurate early warnings and effective interventions for real network performance problems. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides an embedded intelligent gateway algorithm based on a GNN fusion neural network with both long and short time series, aiming to solve the technical problem that existing analysis paradigms relying on environmental data cannot accurately predict and effectively intervene in real network performance issues.

[0007] To achieve the above objectives, the present invention provides the following technical solution, comprising the following steps: Step 1: Collecting network performance time-series data of the gateway; Step 2: Preprocessing the network performance time-series data to form standardized time-series data; Step 3: Inputting the standardized time-series data into a pre-trained graph neural network (GNN) model for spatiotemporal feature extraction to obtain fused features; Step 4: Performing state recognition or anomaly diagnosis based on the fused features.

[0008] Furthermore, in the step of collecting network performance time-series data of the gateway, environmental parameter time-series data and device operating parameter time-series data are also collected simultaneously; the standardized time-series data is multimodal time-series data formed by uniformly standardizing network performance time-series data, environmental parameter time-series data and device operating parameter time-series data.

[0009] Furthermore, the multimodal time series data is constructed as a graph structure, with gateways or connecting devices as nodes and connection relationships or communication traffic as edges, and the node features include the multimodal time series data.

[0010] Furthermore, the pre-trained GNN model incorporates a Long Short-Term Memory (LSTM) unit to simultaneously extract both short-term and long-term temporal dependencies of node features.

[0011] Furthermore, the network performance time-series data includes one or more of the following: network traffic, transmission latency, packet loss rate, protocol type distribution, CPU and memory usage.

[0012] Further steps for preprocessing time-series data include: segmenting the time-series data using a sliding window and standardizing or normalizing the data within each window.

[0013] Furthermore, the GNN model employs one or more combinations of Graph Convolutional Network (GCN), Graph Attention Network (GAT), or Spatiotemporal Graph Convolutional Network (ST-GCN).

[0014] Furthermore, the fused features output by the GNN model are passed through an attention mechanism layer to evaluate the importance weights of features at different time steps and at different nodes for state recognition.

[0015] Furthermore, the algorithm is deployed in the hardware acceleration module of the embedded gateway, and the pre-trained GNN model is a lightweight model that has been pruned and quantized.

[0016] Furthermore, the gateway includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the algorithm as described in any one of claims 1 to 9.

[0017] Compared with the prior art, the beneficial effects of the present invention are:

[0018] 1. In this invention, the algorithm shifts its processing object from traditional environmental parameter time-series data to network performance time-series data. By directly analyzing essential indicators such as network traffic, latency, and packet loss rate, the algorithm can more accurately and directly determine whether the network is congested, whether the link quality is poor, or whether it is under attack. This avoids misjudgments that may result from inferring network problems indirectly through environmental parameters. The algorithm can predict network performance trends. For example, by analyzing long-term time-series patterns in traffic data, it can predict potential future bandwidth bottlenecks; by analyzing short-term spikes in latency, it can issue early warnings before a complete network outage. This transforms the gateway from a passive device that records events after they occur to an active maintenance node that provides early warnings before they happen. After sensing the network performance status in real time, the gateway can dynamically adjust its resource allocation strategy. For example, when network congestion is detected, it automatically allocates higher priority to critical services, ensuring the smooth operation of core services, thereby improving overall network efficiency and user experience.

[0019] 2. This invention standardizes and unifies network performance data, environmental parameter data, and device operating parameter data to form multimodal time-series data. Furthermore, it innovatively constructs the entire network topology as a graph structure, where devices or gateways are nodes, connections are edges, and multimodal data serves as node features, laying the foundation for subsequent deep relationship mining. Traditional methods analyze individual devices or parameters in isolation. This innovation treats the entire network as an interconnected graph. When the latency of a node abnormally increases, the algorithm can simultaneously analyze its neighboring nodes, uplink traffic, and data center ambient temperature, thereby accurately pinpointing the root cause: is it a local device failure, a network link problem, or poor heat dissipation causing device frequency throttling? This global perspective greatly improves the diagnostic accuracy of complex faults. The graph structure can reveal the potential impact between non-directly connected devices. For example, the algorithm may learn that whenever a server in rack A operates under high load, the heat it generates causes a performance degradation in another network device in the same rack. This deep, cross-modal association rule is difficult to discover using traditional analysis methods, providing crucial insights for preventing correlated faults.

[0020] 3. In this invention, GNN is responsible for extracting spatial correlation features from the network topology; LSTM is responsible for extracting temporal dependency features from the temporal data of each node; and the final attention mechanism dynamically evaluates the importance weight of features at different time points and different network nodes for the current diagnostic task. This architecture can understand both "network traffic periodically increases at the end of each workday" and "a switch failure will cause all its downstream devices to simultaneously lose network access." This spatiotemporal joint modeling capability makes the algorithm's characterization of network status more comprehensive and profound. The attention mechanism acts like an "AI detective," not only providing diagnostic conclusions but also "marking" the key evidence for making that judgment. For example, when the model warns of a potential fault, it can show which nodes, at which time point, and which indicator experienced abnormal fluctuations. This greatly enhances the trust of operations and maintenance personnel in AI decision-making and helps them quickly identify problems. Due to its powerful representation learning capabilities, this hybrid model architecture can adapt to various network topologies and business scenarios, effectively analyzing and diagnosing whether it's industrial IoT, data centers, or smart home networks, reducing the cost of redeveloping models for specific scenarios. Attached Figure Description

[0021] 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:

[0022] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1:

[0025] Please see Figure 1 This embodiment provides the following technical solution, including the following steps: Step 1: Collect network performance time-series data of the gateway; Step 2: Preprocess the network performance time-series data to form standardized time-series data; Step 3: Input the standardized time-series data into a pre-trained graph neural network (GNN) model for spatiotemporal feature extraction to obtain fused features; Step 4: Perform state recognition or anomaly diagnosis based on the fused features.

[0026] Specifically, network performance time-series data is continuously collected and polled at fixed time intervals using the embedded gateway's built-in network card driver, protocol stack interface, and sensor interface. The time interval can be configured between 1 millisecond and 10 minutes. The collected raw network performance time-series data undergoes data cleaning, noise reduction, and standardization. This includes segmenting the continuous data stream using a sliding window method, and applying Z-Score standardization to the data within each window to eliminate the influence of dimensions, forming a standardized time-series data matrix with uniform dimensions and numerical norms. The preprocessed standardized time-series data matrix is ​​input into a pre-trained graph neural network model. This model extracts the implicit spatiotemporal correlation features in the data by aggregating the historical states of nodes and the states of neighboring nodes, ultimately outputting a low-dimensional, dense fusion feature vector. The fusion feature vector is then input into a fully connected classifier or regressor for state recognition or anomaly diagnosis, outputting specific state labels or anomaly scores.

[0027] By collecting and analyzing network performance time-series data and extracting spatiotemporal features using a GNN model, a direct and accurate perception and diagnosis of the intelligent gateway's operational status is achieved. This solution overcomes the limitations of existing technologies that infer network status indirectly through environmental parameters, and can directly identify core issues such as network congestion and link failures, significantly improving the gateway's proactive operation and maintenance capabilities.

[0028] In the step of collecting network performance time-series data of the gateway, environmental parameter time-series data and device operation parameter time-series data are also collected simultaneously; the standardized time-series data is multimodal time-series data formed by uniformly standardizing the network performance time-series data, environmental parameter time-series data and device operation parameter time-series data.

[0029] Specifically, in the data acquisition step: multimodal time-series data from different physical interfaces are acquired synchronously, including: environmental parameter time-series data: readings from temperature, humidity, and vibration sensors; device operating parameter time-series data: system status data from the gateway itself, such as CPU utilization and memory usage; and network performance time-series data: kernel metrics from the network layer, including throughput, round-trip latency, and packet loss rate. In the preprocessing step, one-hot encoding is used to encode categorical data such as protocol types, and these are then combined with numerical data using min-max normalization, ultimately concatenating them to form a unified multimodal time-series data tensor.

[0030] By synchronously collecting and standardizing multimodal time-series data, a unified data tensor is constructed, providing a comprehensive and multidimensional data foundation for this algorithm. This enables the fusion processing of heterogeneous information from the environment, devices, and networks, providing data support for subsequent global and systematic analysis and decision-making in the model, and avoiding the one-sidedness of decisions caused by a single data dimension.

[0031] Multimodal time series data is constructed as a graph structure, with gateways or connected devices as nodes and connection relationships or communication traffic as edges. The node features contain multimodal time series data.

[0032] Specifically, the method for constructing multimodal time series data into a graph structure is as follows: Define the graph structure: Define the gateway and each terminal device, sensor, or upstream network node connected to it as a node in the graph. Define edge relationships: If there is a physical connection between two nodes or data communication occurs between them over a period of time, an edge is established between these two nodes; this edge can be undirected or directed and can be assigned a weight, the weight value of which is determined by communication traffic, number of data packets, or connection strength. Define node features: The feature vector of each node consists of the statistics of its associated multimodal time series data within the most recent time window;

[0033] By abstracting network entities into nodes and constructing edges based on connectivity, thus forming graph-structured data, this algorithm can explicitly model and utilize the inherent spatial relationships in network topology. This method lays the foundation for subsequent applications of graph neural networks to process structured data in non-Euclidean spaces, enabling the algorithm to understand the interrelationships between devices.

[0034] The pre-trained GNN model incorporates Long Short-Term Memory (LSTM) units to simultaneously extract both short-term and long-term temporal dependencies of node features.

[0035] Specifically, the pre-trained GNN model is a spatiotemporal graph convolutional network that incorporates long short-term memory units. Its operations include: Temporal feature extraction: The feature sequence of each node is first processed by an LSTM unit to capture its dynamic long-term dependencies, outputting a condensed temporal feature vector. Spatial feature extraction: The temporal feature vectors of each node output by the LSTM are used as input features to the current graph convolutional layer. Graph convolution operations are used to aggregate information from each node's first-order neighbors, updating the node representation and thus capturing spatial dependencies in the network topology. These two steps can be stacked in multiple layers to achieve deeper spatiotemporal feature fusion.

[0036] By employing a spatiotemporal graph convolutional network structure that integrates LSTM units, this algorithm achieves the collaborative extraction and fusion of long-term dependencies in temporal data and spatial dependencies in topological data. This hybrid model architecture can simultaneously capture the spatiotemporal dynamic evolution of network states, greatly improving the accuracy of state identification and prediction.

[0037] Network performance time-series data includes one or more of the following: network traffic, transmission latency, packet loss rate, protocol type distribution, and CPU and memory usage.

[0038] Specifically, network performance time-series data includes, but is not limited to: Network traffic: bits per second or packets per second, differentiated by physical port or logical protocol. Transmission delay: end-to-end delay or intra-bridge forwarding delay measured based on the ICMP protocol. Packet loss rate: the ratio of the number of packets sent to the number of packets received calculated on a specific flow or port. Protocol type distribution: statistics on the number or percentage of packets from various protocols such as HTTP, MQTT, TCP, and UDP over a period of time. CPU and memory utilization: the percentage of processor and memory resources used by the gateway's operating system kernel.

[0039] By clearly defining the specific composition of network performance time-series data, including key indicators such as traffic, latency, and packet loss rate, the completeness and relevance of the data collected by the algorithm are ensured. This design enables the algorithm to comprehensively characterize the health of network performance, providing detailed data support for accurate diagnosis.

[0040] The steps for preprocessing time series data include: dividing the time series data into segments using a sliding window, and standardizing or normalizing the data within each window.

[0041] Specifically, the preprocessing steps for time-series data include: Sliding window segmentation: An infinitely long time-series data is segmented using a sliding window with configurable length and step size, with each window forming a data sample. Standardization: For the data within each window, the mean and standard deviation are calculated for each feature sequence dimension, and Z-score standardization is performed to ensure it conforms to a standard normal distribution with a mean of 0 and a standard deviation of 1. This step effectively eliminates the influence of differences in the dimensions and numerical ranges of different sensors, accelerating model convergence.

[0042] By employing sliding window segmentation and Z-Score normalization to preprocess the original time-series data, noise interference and the influence of inconsistencies in the units of measurement between different feature dimensions are effectively eliminated. This step ensures the data quality of the input model and significantly improves the model's training efficiency, convergence speed, and generalization performance.

[0043] GNN models employ one or more combinations of Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), or Spatiotemporal Graph Convolutional Networks (ST-GCN).

[0044] Specifically, GNN models can choose any of the following architectures or combinations thereof: Graph Convolutional Networks: Define convolution operations in the frequency domain using spectral graph theory to achieve smoothing and aggregation of node features. Graph Attention Networks: Introduce an attention mechanism, allowing the model to adaptively learn and assign different importance weights to different neighbors of a node, followed by weighted aggregation. Spatiotemporal Graph Convolutional Networks: Specifically designed for temporal graph data, using parallel 1D-CNN and GCN to extract temporal and spatial features respectively.

[0045] By specifying that GNN models can use advanced architectures such as GCN, GAT, or ST-GCN, the algorithm framework is given a high degree of flexibility and scalability. Depending on different application scenarios and performance requirements, the most suitable network structure can be selected to achieve the best balance between performance and efficiency.

[0046] The fused features output by the GNN model are passed through an attention mechanism layer to evaluate the importance weights of features at different time steps and at different nodes for state recognition.

[0047] Specifically, after the GNN model outputs fused features, a multi-head self-attention mechanism layer is added: this layer takes the feature vector sequence of all nodes output by the GNN model as input. By calculating the correlation between elements within the sequence, a dynamic importance weight is assigned to each time step and each node feature. Finally, the weighted feature vectors of all nodes are pooled and aggregated to generate a global, context-aware feature representation for the final decision.

[0048] By introducing a multi-head self-attention mechanism at the end of the model, dynamic weight allocation for features at different time steps and nodes is achieved. This mechanism enhances the interpretability of the model and can automatically focus on the key features and key nodes most relevant to the current diagnostic task, thereby improving the accuracy and reliability of decision-making.

[0049] The algorithm is deployed in the hardware acceleration module of the embedded gateway, and the pre-trained GNN model is a lightweight model that has been pruned and quantized.

[0050] Specifically, the algorithm is deployed on the embedded gateway in the following ways: Hardware deployment: The pre-trained GNN model is deployed on the gateway's dedicated hardware acceleration module, which can be an NPU, TPU, or FPGA. Model lightweighting: To adapt to resource-constrained embedded environments, the model undergoes structured pruning to reduce the number of parameters before deployment, and INT8 quantization is performed to reduce computational accuracy and storage overhead, thereby optimizing inference speed and power consumption.

[0051] By deploying a pruned and quantized lightweight model on a dedicated hardware acceleration module, the deployment challenge of complex neural network models in resource-constrained embedded gateway environments has been successfully solved. This solution significantly reduces computational latency and power consumption while maintaining algorithm performance, meeting the real-time requirements of embedded devices.

[0052] The gateway includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the algorithm as described in any one of claims 1 to 9.

[0053] Specifically, it includes: one or more processors; a memory storing one or more computer programs; multiple network interfaces for connecting to internal and external networks; sensor interfaces for connecting to external environmental sensors; and when one or more computer programs are executed by one or more processors, the gateway performs the method steps as described in any one of claims 1 to 9.

[0054] By providing an embedded smart gateway hardware entity that includes a processor, memory, and multiple interfaces, the aforementioned algorithm method is realized as a physical carrier and productized. This claim protects the hardware device executing the algorithm, providing solid legal protection for the practical application and commercial promotion of the technology.

[0055] Example 2:

[0056] Please see Figure 1 In this embodiment, the staff uses the method disclosed in this invention in an edge computing scenario of an industrial Internet of Things park;

[0057] This invention applies the embedded intelligent gateway algorithm based on a long-short time series GNN fusion neural network. The gateway uses an Advantech ARK-3533 industrial PC as its hardware platform, equipped with an Intel Atom x6425E processor and an Intel Movidius Myriad X VPU acceleration module. Through its built-in Intel I210 Gigabit Ethernet card and RS485 sensor interface, it continuously collects network performance time-series data (including port traffic, TCP retransmission rate, end-to-end latency), environmental parameters (room temperature and humidity, rack vibration amplitude), and equipment operating parameters (CPU load, memory usage) at 1-second intervals. The raw data is segmented into segments with a sliding window length of 60 seconds and a step size of 10 seconds, and then standardized using the Z-Score method to form a multimodal time-series data tensor. Subsequently, the campus network topology is constructed as a graph structure, with core switches and 12 edge gateways as nodes, and physical connections and average communication traffic as edges. Node features include the standardized multimodal data. The graph data was input into a lightweight ST-GCN model (integrated with LSTM units) that had undergone pre-pruning and INT8 quantization, achieving real-time inference at 35 frames per second on the VPU acceleration module. The model extracted the topological relationships between nodes through spatiotemporal convolutional layers, captured the periodic regularity of traffic data using LSTM units, and ultimately identified periodic packet loss anomalies in edge gateway 3 due to poor heat dissipation (the packet loss rate increased to 12% when the ambient temperature exceeded 35 degrees Celsius) through a multi-head self-attention mechanism. It also predicted that the core switch would experience a performance bottleneck due to memory overflow in 15 minutes. The gateway proactively triggered a dynamic traffic scheduling strategy, migrating the critical tasks of gateway 3 to gateway 5 and clearing the core switch cache in advance, successfully avoiding network outages and increasing the campus network availability from 99.5% to 99.98%.

[0058] The working principle of this invention is as follows: The embedded intelligent gateway first synchronously collects network performance time-series data (including traffic, latency, packet loss rate, protocol distribution, and CPU / memory usage), environmental parameter time-series data (such as temperature and humidity), and device operating parameter time-series data. This data is then processed through sliding window segmentation and Z-score normalization to form a unified multimodal time-series data tensor. Subsequently, this data is constructed into a graph structure, where the gateway and its connected devices serve as nodes, physical connections or communication relationships are used as edges, and node features contain multimodal time-series data. This graph data is then input into a pre-trained lightweight GNN model (such as one fused with LSTM). The unit uses a spatiotemporal graph convolutional network (SPCFN) to extract long-term and short-term temporal dependencies of node features through LSTM, and aggregates neighbor node information through graph convolution to capture spatial topological relationships, thereby outputting a vector of fused spatiotemporal features. This fused feature is then dynamically evaluated by an attention mechanism layer to assess the importance weights of node features at different time steps. Finally, a classifier or regressor uses the weighted features to achieve accurate network state identification or anomaly diagnosis. After pruning and quantization optimization, the entire algorithm is deployed on the hardware acceleration module (such as NPU / FPGA) of the gateway to achieve real-time perception and proactive early warning of network congestion, link failure, and performance degradation.

[0059] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An embedded intelligent gateway algorithm based on a GNN fusion neural network with long and short time series, characterized in that: Includes the following steps: Step 1: Collect network performance time-series data from the gateway; Step 2: Preprocess the network performance time-series data to form standardized time-series data; Step 3: Input the standardized time-series data into a pre-trained graph neural network (GNN) model to extract spatiotemporal features and obtain fused features; Step 4: Perform state recognition or anomaly diagnosis based on the fusion features.

2. The embedded intelligent gateway algorithm based on long and short time series GNN fusion neural network according to claim 1, characterized in that: In the step of collecting network performance time-series data of the gateway, environmental parameter time-series data and device operating parameter time-series data are also collected simultaneously. The standardized time-series data refers to multimodal time-series data formed by uniformly standardizing network performance time-series data, environmental parameter time-series data, and equipment operating parameter time-series data.

3. The embedded intelligent gateway algorithm based on long and short time series GNN fusion neural network according to claim 2, characterized in that: The multimodal time series data is constructed into a graph structure, with gateways or connected devices as nodes and connection relationships or communication traffic as edges, and the node features include the multimodal time series data.

4. The embedded intelligent gateway algorithm based on long and short time series GNN fusion neural network according to claim 3, characterized in that: The pre-trained GNN model incorporates Long Short-Term Memory (LSTM) units to simultaneously extract both short-term and long-term temporal dependencies of node features.

5. The embedded intelligent gateway algorithm based on long and short time series GNN fusion neural network according to claim 1, characterized in that: The network performance time-series data includes one or more of the following: network traffic, transmission latency, packet loss rate, protocol type distribution, and CPU and memory usage.

6. The embedded intelligent gateway algorithm based on long and short time series GNN fusion neural network according to claim 1, characterized in that: The steps for preprocessing time series data include: dividing the time series data into segments using a sliding window, and standardizing or normalizing the data within each window.

7. The embedded intelligent gateway algorithm based on long and short time series GNN fusion neural network according to claim 1, characterized in that: The GNN model employs one or more combinations of Graph Convolutional Network (GCN), Graph Attention Network (GAT), or Spatiotemporal Graph Convolutional Network (ST-GCN).

8. The embedded intelligent gateway algorithm based on long and short time series GNN fusion neural network according to claim 4, characterized in that: The fused features output by the GNN model are passed through an attention mechanism layer to evaluate the importance weights of features at different time steps and at different nodes for state recognition.

9. The embedded intelligent gateway algorithm based on long and short time series GNN fusion neural network according to claim 1, characterized in that: The algorithm is deployed in the hardware acceleration module of an embedded gateway, and the pre-trained GNN model is a lightweight model that has been pruned and quantized.

10. An embedded smart gateway, characterized in that: The gateway includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the algorithm as described in any one of claims 1 to 9.

Images