Big model based internet of things terminal cooperative perception and intelligent scheduling system

Abstract

This invention provides a large-scale model-based collaborative sensing and intelligent scheduling system for IoT terminals, relating to the field of network topology management technology. It includes: a construction module for acquiring multi-dimensional data traffic characteristics and service flow metadata of each terminal node within the IoT cluster; converting the multi-dimensional data traffic characteristics into data flow polarity vectors through a feature mapping network of the large-scale model; and constructing a dynamic data attribute representation matrix. A generation module is used to calculate the data link congestion degree and data encapsulation protocol matching degree between each node and the edge scheduling center based on the dynamic data attribute representation matrix, and to construct a dynamic data flow coupling topology diagram simulating the potential field distribution. This invention reduces the deviation between actual transmission latency and expected latency, improving the real-time performance and scheduling adaptability of the IoT cluster under dynamic load.

Description

Technical Field

[0001] This invention relates to the field of network topology management technology, and in particular to an IoT terminal collaborative sensing and intelligent scheduling system based on a large model. Background Technology

[0002] In IoT data communication scenarios, such as an environmental monitoring IoT cluster deployed in a large factory area, there are usually multiple sensor terminal nodes that work with different sampling periods. Each node sends business flow data to the edge scheduling center through a wireless link. In actual operation, some nodes will generate instantaneous data traffic exceeding the average level during specific periods, such as when a sudden equipment alarm triggers synchronous reporting, while other nodes may be in an idle state. Most existing data scheduling methods rely on relatively static routing configurations or congestion avoidance strategies based on a single threshold judgment.

[0003] Taking this factory area as an example, when multiple gas concentration sensors in a certain area simultaneously start reporting at high frequency due to local anomalies, the shared link from that area to the edge dispatch center may experience short-term congestion. In this situation, the existing scheduling mechanism may find it difficult to dynamically perceive changes in the polarity of data traffic between different nodes in a short period of time, and it is also difficult to accurately assess the adaptability of heterogeneous data encapsulation protocols under the current link state. At the same time, the edge dispatch center has relatively limited ability to predict the congestion propagation path of service flow metadata between nodes, and may not be able to know in a timely manner the possible spread trend of congestion in the topology, resulting in a certain deviation between the rate of change of bandwidth allocation and the actual demand. Furthermore, when the scheduling system triggers data diversion operations based on experience or fixed rules, there may sometimes be a non-constant difference between the actual transmission delay and the expected delay of each link. If these differences are not effectively addressed, they may gradually accumulate with fluctuations in service load, thereby affecting the timeliness of collaborative sensing by IoT terminals and the consistency of scheduling command execution to a certain extent. Summary of the Invention

[0004] This invention provides a collaborative sensing and intelligent scheduling system for IoT terminals based on a large model, which reduces the deviation between actual transmission latency and expected latency, and improves the real-time collaborative sensing and scheduling adaptability of IoT clusters under dynamic load.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] The first aspect is an IoT terminal collaborative sensing and intelligent scheduling system based on a large-scale model, including:

[0007] The module is used to acquire multidimensional data traffic characteristics and business flow metadata of each terminal node in the IoT cluster. It converts the multidimensional data traffic characteristics into data flow polarity vectors through the feature mapping network of the large model and constructs a dynamic data attribute representation matrix.

[0008] The generation module is used to calculate the data link congestion degree and data encapsulation protocol matching degree between each node and the edge scheduling center based on the dynamic data attribute representation matrix, and to construct a dynamic data flow coupling topology diagram simulating the potential field distribution.

[0009] The analysis module is used to extract the time-varying characteristics of data link congestion and the distribution of protocol matching degree between topology nodes based on the dynamic data flow coupling topology graph, and obtain the logical space state offset and service flow diffusion gradient; based on the logical space state offset and service flow diffusion gradient, it calculates the gradient diffusion path and bandwidth allocation change rate vector of service flow metadata in the dynamic data flow coupling topology graph, and obtains the predicted scheduling trajectory sequence.

[0010] The correction module is used to determine the state transition threshold and control the scheduling inertia compensation for each trajectory point based on the predicted scheduling trajectory sequence, and dynamically trigger the diversion command of the target data transmission channel; calculate the expected transmission delay of each link based on the gradient diffusion path and the bandwidth allocation change rate vector, and collect the actual transmission delay of each node after executing the current diversion command in real time, and use the difference between the actual transmission delay and the expected transmission delay as the delay deviation feedback quantity; use the delay deviation feedback quantity to perform gradient iterative correction of the scheduling inertia compensation control, determine the collaborative sensing scheduling execution strategy and send it to each IoT terminal.

[0011] In a second aspect, a computer-readable storage medium storing a program that, when executed by a processor, implements the system.

[0012] The above-described solution of the present invention has at least the following beneficial effects:

[0013] A large-model feature mapping network is used to efficiently convert traffic features into data flow polarity vectors and construct a dynamic data attribute representation matrix, achieving standardized representation of the underlying characteristics of IoT terminal data traffic and service flows. Simultaneously, the congestion degree of data links and the matching degree of data encapsulation protocols between each terminal node and the edge scheduling center are calculated. By constructing a dynamic data flow coupling topology diagram simulating the potential field distribution, the coupling relationship of data flows within the cluster and the distribution of link states are presented. The logical space state offset and service flow diffusion gradient are solved to deduce the gradient diffusion path, bandwidth allocation change rate vector, and predicted scheduling trajectory sequence, enabling prediction of service flow transmission status and scheduling trends. By determining state transition thresholds and controlling scheduling inertia compensation for the scheduling trajectory, data transmission channel diversion commands can be dynamically and intelligently triggered. Gradient iterative correction is performed based on the feedback of the deviation between expected and actual transmission delays, continuously optimizing the scheduling inertia compensation control logic and determining the final collaborative sensing scheduling execution strategy to be issued to each terminal, enhancing the collaborative sensing capability and real-time intelligent scheduling of IoT terminal cluster service flow transmission. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the IoT terminal collaborative sensing and intelligent scheduling system based on a large model, provided by an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of the dynamic data flow coupling topology provided in an embodiment of the present invention.

[0016] Figure 3 This is a schematic diagram of the logical space state offset and the service flow diffusion gradient provided in an embodiment of the present invention. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0018] like Figure 1 As shown, embodiments of the present invention propose an IoT terminal collaborative sensing and intelligent scheduling system based on a large model, comprising:

[0019] The module is used to acquire multidimensional data traffic characteristics and business flow metadata of each terminal node in the IoT cluster. It converts the multidimensional data traffic characteristics into data flow polarity vectors through the feature mapping network of the large model and constructs a dynamic data attribute representation matrix.

[0020] The generation module is used to calculate the data link congestion degree and data encapsulation protocol matching degree between each node and the edge scheduling center based on the dynamic data attribute representation matrix, and to construct a dynamic data flow coupling topology diagram simulating the potential field distribution.

[0021] The analysis module is used to extract the time-varying characteristics of data link congestion and the distribution of protocol matching degree between topology nodes based on the dynamic data flow coupling topology graph, and obtain the logical space state offset and service flow diffusion gradient; based on the logical space state offset and service flow diffusion gradient, it calculates the gradient diffusion path and bandwidth allocation change rate vector of service flow metadata in the dynamic data flow coupling topology graph, and obtains the predicted scheduling trajectory sequence.

[0022] The correction module is used to determine the state transition threshold and control the scheduling inertia compensation for each trajectory point based on the predicted scheduling trajectory sequence, and dynamically trigger the diversion command of the target data transmission channel; calculate the expected transmission delay of each link based on the gradient diffusion path and the bandwidth allocation change rate vector, and collect the actual transmission delay of each node after executing the current diversion command in real time, and use the difference between the actual transmission delay and the expected transmission delay as the delay deviation feedback quantity; use the delay deviation feedback quantity to perform gradient iterative correction of the scheduling inertia compensation control, determine the collaborative sensing scheduling execution strategy and send it to each IoT terminal.

[0023] In this embodiment of the invention, a large-model feature mapping network is used to efficiently convert traffic features into data flow polarity vectors and construct a dynamic data attribute representation matrix, thereby achieving standardized representation of the underlying characteristics of IoT terminal data traffic and service flows. Simultaneously, the data link congestion degree and data encapsulation protocol matching degree between each terminal node and the edge scheduling center are calculated. By constructing a dynamic data flow coupling topology diagram simulating potential field distribution, the coupling relationship of data flows within the cluster and the distribution of link states are presented. The logical space state offset and service flow diffusion gradient are solved to deduce the gradient diffusion path, bandwidth allocation change rate vector, and predicted scheduling trajectory sequence, enabling prediction of service flow transmission status and scheduling trends. By determining the state transition threshold and controlling scheduling inertia compensation for the scheduling trajectory, data transmission channel diversion commands can be dynamically and intelligently triggered. Gradient iterative correction is performed based on the feedback of the deviation between expected and actual transmission delays, continuously optimizing the scheduling inertia compensation control logic and determining the final collaborative perception scheduling execution strategy to be issued to each terminal, enhancing the collaborative perception capability and real-time intelligent scheduling of IoT terminal cluster service flow transmission.

[0024] In a preferred embodiment of the present invention, the process of acquiring multidimensional data traffic features and service flow metadata of each terminal node within an IoT cluster, converting the multidimensional data traffic features into data flow polarity vectors through a feature mapping network of a large model, and constructing a dynamic data attribute representation matrix may include:

[0025] In this embodiment of the invention, raw data packets sent within a continuous time window are collected from each terminal node within the IoT cluster. Based on the transmission rate sequence, data packet length sequence, and time interval sequence of adjacent data packets in the raw data packets, multidimensional data traffic characteristics of each terminal node are extracted. Simultaneously, based on the service type code, source terminal identifier, destination edge dispatch center address, and data encapsulation format label in the raw data packets, service flow metadata of each terminal node is extracted. Specifically, in a large-scale factory environmental monitoring IoT cluster scenario, a data acquisition agent is deployed at the edge dispatch center. This agent establishes a stable data interaction connection with all sensor terminal nodes within the cluster via the IoT cluster's wireless communication link, and is used to collect all raw data packets sent by each terminal node within a continuous time window in real time. The continuous time window setting needs to adapt to the sampling cycle differences of the factory terminal nodes and the fluctuation characteristics of high-frequency sensor reporting and idle states. The time window length is set to T, with a value range of 5s-30s. It can be dynamically adjusted according to the overall load of the terminal nodes. When a sudden data traffic surge occurs within the cluster, T is automatically adjusted to 5s to improve the real-time performance of data acquisition; in idle states, it is adjusted to 30s to reduce the acquisition load on the edge dispatch center. After collecting the raw data packets, each terminal node's data packets are parsed frame by frame to extract two types of core information: multidimensional data traffic characteristics and service flow metadata. For multidimensional data traffic characteristics, three key time sequences are extracted from the header fields and transmission records of the raw data packets: transmission rate sequence, data packet length sequence, and time interval sequence between the arrival of adjacent data packets. Among these, the transmission rate sequence... This refers to the instantaneous transmission rate of the i-th terminal node, collected every 100ms within a continuous time window. It is extracted by dividing the total number of bytes in the original data packet by the duration of the corresponding time segment. The calculation formula is as follows: ,in Let be the total number of bytes of data packets transmitted by the i-th terminal node within a 100ms time slice at time t (t∈[0,T]). The sampling time interval is fixed at 100ms. Data packet length sequence. This refers to the i-th terminal node within a continuous time window, the i-th terminal node... The actual number of bytes in a transmitted data packet. The sequence number is extracted directly from the header length field of the original data packet. The time interval sequence between the arrival of adjacent data packets. It refers to the i-th terminal node. The data packet and the first -1 is the time difference between the arrival of a data packet at the edge dispatch center, in milliseconds, calculated using the following formula: ,in For the i-th terminal node The timestamp of each data packet arriving at the edge scheduling center The start time of the time window, when When =1, .

[0026] Simultaneously, service flow metadata for each terminal node is extracted from the packet header identifier field and data payload header of the original data packets. This metadata includes four core pieces of information: service type code, source terminal identifier, destination edge dispatch center address, and data encapsulation format tag. The service type code is set according to the service requirements of the plant's environmental monitoring; for example, code 001 corresponds to gas concentration monitoring, code 002 to temperature and humidity monitoring, and code 003 to equipment status monitoring. This code is directly extracted from the service identifier field of the original data packets. The source terminal identifier is a unique 16-bit hardware code for each terminal node, fixed at the factory, used to distinguish different terminal nodes, and is extracted from the source address field of the data packets. The destination edge dispatch center address is the combination of the IP address and port number of the edge dispatch center deployed in the plant, used to determine the target node for data transmission, and is extracted from the destination address field of the data packets. The data encapsulation format tag identifies the encapsulation protocol type of the data packets; for example, the MQTT protocol corresponds to the tag MQTT, the HTTP protocol corresponds to the tag HTTP, and the CoAP protocol corresponds to the tag CoAP. This tag is extracted from the protocol identifier field of the data packets, providing a basis for subsequent protocol matching calculations. After extraction, the multidimensional data traffic characteristics and business flow metadata of each terminal node are associated and stored. During storage, the source terminal identifier is used as the index, and the corresponding transmission rate sequence, data packet length sequence, time interval sequence of adjacent data packet arrival are bound with the business type code, destination edge scheduling center address, and data encapsulation format label to form a unique data set for each terminal node.

[0027] The multidimensional data traffic features of each terminal node are input into a feature mapping network. A nonlinear polarity transformation is performed on the transmission rate, packet length, and time interval, mapping the multidimensional data traffic features into a data flow polarity vector with directional and magnitude attributes. Specifically, this involves constructing a large-scale feature mapping network based on a deep learning architecture. This network converts the extracted multidimensional data traffic features (transmission rate sequence, packet length sequence, and time interval sequence of adjacent packet arrivals) into a data flow polarity vector with directional and magnitude attributes. The network adopts a three-layer structure: an input layer, a hidden layer, and an output layer. The structure and parameter settings of each layer are as follows: The input layer receives the multidimensional data traffic features of each terminal node, with an input dimension of 3×m. Each neuron in the input layer corresponds to a single traffic feature (transmission rate, packet length, time interval) at a sampling point, with a total of 3×m neurons. The input data is the extracted... , , Three sequences are input, and the data needs to be normalized before input to eliminate the influence of different units. The normalized data ranges from [0, 1] to ensure the stability of network training and feature mapping. The hidden layer uses the ReLU activation function to perform a non-linear transformation on the normalized features of the input layer, and to explore the intrinsic correlation between multi-dimensional flow features. The number of neurons in the hidden layer is set to 2×3×m, and the parameters are optimized using stochastic gradient descent (SGD). The learning rate is set to 0.001, the number of iterations is set to 1000, and the regularization coefficient is set to 0.0001 to avoid network overfitting. The calculation process of the hidden layer is h = max Where h is the hidden layer output vector, The weight matrix from the input layer to the hidden layer has dimensions (2×3×m)×(3×m). Here is the bias vector of the hidden layer, with dimensions (2×3×m)×1. This is the normalized feature vector of the input layer. The output layer outputs the data stream polarity vector, with a dimension of 2, corresponding to the direction and magnitude components of the data stream polarity vector, respectively. A linear activation function is used to ensure the continuity and interpretability of the output results. The calculation process of the output layer is as follows: ,in The output data stream polarity vector. = , Let be the direction component of the data flow polarity vector of the i-th terminal node, with a value range of [0, 2π), corresponding to the angle in polar coordinates. Let be the magnitude component of the data flow polarity vector of the i-th terminal node, with a value range of [0, 1], corresponding to the magnitude in polar coordinates. This is the weight matrix from the hidden layer to the output layer, with dimensions 2×(2×3×m). This is the bias vector for the output layer, with a dimension of 2×1.

[0028] After the network was built, historical traffic data from the factory's IoT cluster was used for training. The training dataset contained multi-dimensional data traffic characteristics and corresponding label values ​​for each terminal node under different load states (idle, normal, burst) over the past three months. The label values ​​were composed of manually labeled data flow polarity vectors (direction component and magnitude component). During training, the network parameters were optimized by minimizing the loss function. The loss function used was the mean squared error loss function, calculated using the formula L= Where N is the number of training samples, The polarity vector of the data stream output by the network. The data stream polarity vectors are manually labeled, and the training continues until the loss function converges (the convergence threshold is set to 10). -4After this, the trained feature mapping network is obtained. The extracted and normalized multi-dimensional data flow features of each terminal node are input into the trained feature mapping network. The network receives the feature data through the input layer, performs a non-linear transformation through the hidden layer, and then outputs the corresponding data flow polarity vector through the output layer. , where the directional component Used to characterize the dynamic changing trend of data traffic at the terminal node, for example when When the value is in the range [0, π / 2), it indicates that the traffic of the terminal node is increasing. When the value is in the range [π / 2, π], it indicates that the flow rate is decreasing. When the value is within the range [π, 2π), it indicates that the traffic is in a stable state, corresponding to the idle or regular reporting scenarios of the terminal node, and the magnitude component. Used to characterize the intensity of data traffic at this terminal node. The closer to 1, the greater the flow intensity; the closer to 0, the smaller the flow intensity. This can distinguish between burst flow and flow in idle conditions.

[0029] Alignment is performed point-by-point with the terminal nodes on the timestamp field carried in the data flow polarity vector and business flow metadata. The aligned timestamp is used as the row index of the matrix, and the terminal node number and business attribute are used as the column index. The direction component and magnitude component of the aligned data flow polarity vector are filled into the corresponding cells of the matrix to construct a dynamic data attribute representation matrix. Specifically, this includes: aligning the timestamp field carried in the data flow polarity vector and business flow metadata with the terminal nodes point-by-point to ensure the spatiotemporal consistency of the data and provide a unified index benchmark for matrix construction. The timestamp alignment adopts a time synchronization mechanism accurate to milliseconds, using the system time of the edge scheduling center as the benchmark. The timestamps in the business flow metadata of each terminal node are calibrated to eliminate the time difference caused by clock deviations between different terminal nodes. The calibration formula is as follows: ,in For the i-th terminal node The timestamp after data packet calibration The clock deviation between the terminal node and the edge scheduling center is obtained in real time via the NTP time synchronization protocol to ensure that the calibrated timestamp error does not exceed 1ms. After calibration, the aligned timestamps are used as row indices of the dynamic data attribute representation matrix. The row index is taken from all calibrated timestamps within a continuous time window, with each timestamp corresponding to one row of the matrix. The number of rows is the same as the number of sampling points within the time window, i.e., the row index is... The column indexes of the matrix are set up using the terminal node number and business attributes. The column indexes are set up as follows: each terminal node corresponds to two columns, which correspond to the directional component and the magnitude component of the data flow polarity vector, respectively. The naming format of the column indexes is terminal node number-directional component and terminal node number-magnitude component. For example, the terminal node with the number 001 corresponds to two columns: 001-directional component and 001-magnitude component. The number of columns is 2×N, which ensures that each terminal node has a unique column index corresponding to its two polarity vector components.

[0030] After the index is set, the direction component and magnitude component of the aligned data stream polarity vector are filled into the corresponding cells of the matrix to construct a dynamic data attribute representation matrix. The specific filling rule is as follows: for the i-th terminal node, at the timestamp... For rows corresponding to (j=1,2,…,m), the i-direction component column should be filled with the direction component of the terminal node at that timestamp. The i-modulus component column is filled with the modulus component corresponding to the terminal node at that timestamp. If a terminal node does not transmit data packets within a certain timestamp, the two columns of cells corresponding to that timestamp for that terminal node are filled with 0, indicating that the terminal node did not transmit data at that moment. The constructed dynamic data attribute representation matrix is ​​as follows: A matrix with 2N rows and 2N columns has the matrix expression M= Where M is the dynamic data attribute representation matrix, Let be the directional component corresponding to the i-th terminal node at the j-th timestamp. The matrix represents the modulus component corresponding to the i-th terminal node at the j-th timestamp. This matrix can be updated in real time. Every time a sampling time interval passes, a new row corresponding to the latest timestamp is added, while the row corresponding to the earliest timestamp is deleted, ensuring that the matrix always reflects the data flow polarity characteristics of each terminal node within a continuous time window.

[0031] By collecting multidimensional traffic characteristics and business flow metadata from terminals, and combining them with a constructed large-scale model feature mapping network, the multidimensional traffic characteristics are converted into data flow polarity vectors with both directional and intensity attributes, capturing the dynamic differences between burst traffic and idle states of IoT terminals in the factory area.

[0032] like Figure 2 As shown, in a preferred embodiment of the present invention, the data link congestion degree and data encapsulation protocol matching degree between each node and the edge scheduling center are calculated based on the dynamic data attribute characterization matrix, and a dynamic data flow coupling topology diagram simulating the potential field distribution is constructed, which may include:

[0033] In this embodiment of the invention, based on a dynamic data attribute representation matrix, the magnitude component sequence and direction component deflection trajectory of each terminal node within a continuous time window, as well as the data encapsulation format label sequence corresponding to each terminal node, are extracted. Specifically, this includes: based on the constructed dynamic data attribute representation matrix M, where matrix elements contain the data stream polarity vector direction components of each terminal node at different timestamps. With modulus component Where i is the terminal node number (i=1, 2, ..., N) and j is the timestamp (j=1, 2, ..., m). First, perform column-wise extraction on this matrix to obtain the complete feature sequence of each terminal node within a continuous time window. Simultaneously, associate the business flow metadata to extract the corresponding data encapsulation format label sequence. For the extraction of the modulus component sequence, for the i-th terminal node, extract all elements of the i-modulus component column in matrix M to constitute the modulus component sequence of that terminal node. This sequence represents the change in traffic intensity at each timestamp of the i-th terminal node within a continuous time window, and its relationship with the modulus component. The physical meaning is consistent, and the value range is [0, 1]. The closer it is to 1, the greater the traffic intensity of the terminal node at that timestamp, corresponding to the traffic characteristics of sudden sensor reports in the factory area scenario.

[0034] For the extraction of the directional component deflection trajectory, for the i-th terminal node, all elements of the i-directional component column in matrix M are extracted to form the directional component sequence of that terminal node. Then, the sequence is subjected to temporal difference processing to obtain the directional component deflection trajectory. ,in (j=2, 3, ..., m), with values ​​ranging from (-2π, 2π), is used to characterize the instantaneous deflection of the dynamic change trend of the traffic at the terminal node, for example... A value greater than 0 indicates that the upward trend in traffic flow is intensifying or the downward trend is slowing down. A value less than 0 indicates a slowdown in the upward trend of traffic or an intensification in the downward trend, which can capture the instantaneous change characteristics of sudden traffic surges at the plant terminals. For the extraction of the data encapsulation format tag sequence, the service flow metadata bound to the i-th terminal node is associated, and the data encapsulation format tags corresponding to all transmitted data packets of that terminal node within a continuous time window are extracted to form the data encapsulation format tag sequence. ,in The data encapsulation format label for the data packet transmitted by the i-th terminal node at the j-th timestamp. If the terminal node did not transmit a data packet within that timestamp (the corresponding cell in the matrix is ​​filled with 0), then... After filling in the blanks and extracting the data, the magnitude component sequence, direction component deflection trajectory, and data encapsulation format label sequence of each terminal node are associated and stored.

[0035] The cumulative amplitude of the modulus component sequence is compared with the preset throughput threshold of the edge scheduling center. This ratio is then weighted and aggregated using the dispersion index of the directional component deflection trajectory to calculate the data link congestion and data encapsulation protocol compatibility between each node and the edge scheduling center. Specifically, this includes calculating the data link congestion, an index used to quantify the link load saturation between the i-th terminal node and the edge scheduling center. The value ranges from [0, 1]. A value closer to 1 indicates more severe link congestion, while a value closer to 0 indicates less congestion. The calculation process consists of two steps: first, calculating the cumulative amplitude of the modulus component sequence; and second, weighted aggregation using the dispersion index of the directional component deflection trajectory. The specific calculation formula is as follows: ,in Let represent the data link congestion level between the i-th terminal node and the edge scheduling center. This is a weighting coefficient with a value range of [0.6, 0.8]. Considering the scenario of sudden flow in the factory area, the impact of flow intensity is taken first, therefore, it is set to... =0.7, The cumulative amplitude of the modulus component sequence of the i-th terminal node is calculated as follows: This characterizes the total traffic intensity of the terminal node within a continuous time window. This is the preset throughput threshold for the edge scheduling center, in dimensionless form, with a value range of [m×0.3, m×0.6]. It can be dynamically adjusted according to the overall load of the factory terminal cluster. Let be the dispersion index of the directional component deflection trajectory of the i-th terminal node. This index characterizes the instability of traffic change trends. A higher dispersion indicates more drastic traffic fluctuations and a greater likelihood of link congestion. The calculation formula is: ,in The average value of the directional component deflection trajectory of the i-th terminal node is calculated as follows: , The value of is in the range [0, 2π). After calculation, it needs to be normalized to ensure that... After normalization, the value range is [0, 1], and... To maintain a consistent range of values ​​and avoid issues related to units, It is the directional component The first-order difference is used to capture the instantaneous velocity of traffic change trends.

[0036] Next, the data encapsulation protocol matching degree is calculated. This indicator quantifies the degree of adaptation between the data encapsulation protocol of the i-th terminal node and the current link state. The value range is [0, 1]. The closer the value is to 1, the better the protocol adaptation and the higher the data transmission efficiency. The closer the value is to 0, the worse the protocol adaptation and the more prone it is to transmission stuttering or data loss. The calculation process is based on the correlation analysis between the data encapsulation format label sequence and the link state. The specific calculation formula is as follows: ,in Let be the data encapsulation protocol matching degree between the i-th terminal node and the edge scheduling center. Let i be the instantaneous link congestion level of the i-th terminal node at the j-th timestamp. The data encapsulation format label for the i-th terminal node at the j-th timestamp. The protocol adaptation weight function is dynamically assigned based on the protocol type and link status. When there is no data transmission, the protocol adaptation degree is assigned a value of 0.

[0037] Data link congestion and data encapsulation protocol matching degree are mapped to the potential field strength scalar of the topology vertex and the coupling weight coefficient of the edge connecting the vertices, respectively. Based on the potential field strength scalar, coupling weight coefficient and the identifier of each terminal node, the spatial topology relationship is initialized, and an initial node link mapping network is constructed. Specifically, the calculated data link congestion degree value and data encapsulation protocol matching degree value of each terminal node are converted into the potential field strength representation value of the vertex in the topology structure and the coupling weight representation value of the edge connecting the vertices, respectively. The link congestion degree value corresponding to a single terminal node is converted into the potential field strength value of the topology vertex corresponding to that terminal node according to a fixed amplification ratio. The range of the amplified values ​​is larger. The larger the potential field strength value, the more prominent the link congestion of the corresponding terminal node and the stronger the spatial potential field radiation effect. It can correspond to the high load node characteristics formed by the sudden traffic reporting of sensors in the factory scene. The edge scheduling center is set as the central core vertex of the entire topology network, and a fixed and larger potential field strength value is set separately to ensure that the potential field radiation effect of the central vertex is stronger than that of all terminal node vertices, guiding the network's business data to converge and be transmitted to the edge scheduling center.

[0038] The protocol compatibility value between a single terminal node and the edge scheduling center is directly used as the coupling weight value of the connection edge between the terminal node vertex and the core vertex of the center. The larger the coupling weight value, the better the protocol compatibility between the two nodes, the higher the degree of data transmission coupling, and the stronger the link communication stability. Considering the actual situation of indirect data interaction and mutual influence of link resources between terminal nodes within the factory IoT cluster, a rule for calculating the coupling weight of the connection edge between any two different terminal node vertices is constructed. This rule incorporates the protocol compatibility level and link congestion level of the two nodes, reflecting both protocol compatibility and avoiding the problem of excessive coupling between highly congested nodes causing further deterioration of the overall network link load. After completing the parameter mapping and assignment, the coupling weight is calculated based on the potential field strength value of each vertex and the coupling weight value of each connection edge. The initial planar spatial topology is initialized and arranged using the values ​​and unique identifiers of each terminal node, forming an initial node link mapping network. The core vertex of the edge dispatch center is used as the origin of the two-dimensional plane coordinates. The initial planar coordinates of each terminal node vertex are assigned according to the actual physical distance between each terminal node and the edge dispatch center and the direction of traffic transmission. The establishment of topological connections between nodes is determined according to the magnitude of the coupling weight. Topological connections are established between the core vertex and all terminal node vertices. Topological connections are only added between terminal nodes when the coupling weight reaches a preset threshold value to avoid network structure redundancy caused by low-association connections. Each topological vertex is labeled with the corresponding terminal node number and potential field strength value, and each topological connection edge is labeled with the corresponding coupling weight value, thus completing the construction of the initial node link mapping network.

[0039] In the initial node link mapping network, using the potential field strength scalar as the reference radiation source, a virtual gravitational and repulsive field distribution pattern is constructed along the decay direction of the coupling weight coefficient. The resultant force direction and displacement step size of each vertex are calculated. Based on the resultant force direction and displacement step size, the spatial positions of the vertices in the initial node link mapping network are iteratively rearranged to obtain a dynamic data flow coupling topology. Specifically, this includes: constructing virtual gravitational and repulsive fields. The virtual gravitational field is generated by the core vertex and is used to attract all terminal node vertices towards the core vertex, ensuring efficient data transmission to the edge scheduling center. The virtual repulsive field is generated by the interaction between terminal node vertices to avoid excessive aggregation of congested node vertices, alleviating link congestion, and simultaneously preventing non-congested nodes from being excessively attracted, leading to wasted link resources. The specific field strength calculation is as follows: the virtual gravitational field strength of the core vertex relative to the i-th terminal node vertex... The calculation formula is: ,in This is the gravitational coefficient, with a value of 0.5 (dimensionless), used to adjust the strength of gravity. The potential field strength scalar at the core vertex. Let be the potential field strength scalar of the i-th terminal node vertex. The distance between the i-th terminal vertex and the core vertex is calculated from the coordinates of the two points. ,in( , Let be the current coordinates of the i-th terminal node vertex. Let be the unit vector pointing from the i-th terminal vertex to the core vertex. To ensure that the direction of gravity points towards the core apex, the first The virtual repulsive field strength of each terminal node vertex to the i-th terminal node vertex The calculation formula is: ,in This is the repulsion coefficient, with a value of 1.2 (dimensionless), used to adjust the repulsion strength. , They are respectively the i-th, The potential field strength scalar of each terminal node vertex. For the i-th, The distance between the vertices of each terminal node For from the first Each terminal node vertex points to a unit vector of the i-th terminal node vertex, ensuring that the repulsive forces are directed away from each other.

[0040] Calculate the resultant force direction and displacement step size of each terminal node vertex, and the total resultant force of the i-th terminal node vertex. The vector sum of all attractive and repulsive forces is calculated using the following formula: ,in The resultant force is the sum of the repulsive force vectors of all other terminal node vertices to the i-th terminal node vertex, and the direction of the resultant force is the total resultant force. The vector direction, i.e. ,in The magnitude of the total resultant force and the displacement step size. The formula used to characterize the distance a vertex moves in each iteration is as follows: ,in This is the step size coefficient, with a value of 0.1m, used to control the iteration speed and prevent excessive displacement from causing topological instability. The maximum magnitude of the total resultant force of all terminal node vertices is used to normalize the displacement step size, ensuring that the displacement step size of each vertex matches the magnitude of its own force. The greater the force, the greater the displacement, allowing for rapid adjustment to a reasonable position. Based on the resultant force direction and displacement step size, the spatial positions of vertices in the initial node link mapping network are iteratively rearranged. The iteration process is as follows: calculate the total resultant force, resultant force direction, and displacement step size of all current terminal node vertices; update the coordinates of each terminal node vertex according to the displacement step size and resultant force direction, using the following update formula: Calculate the updated edge coupling weight coefficients and potential field strength scalars between each vertex. Determine if the iteration converges; the convergence condition is the displacement step size of all terminal vertices. If the iteration distance is ≤0.01m and convergence is not achieved, repeat the iteration steps; if convergence is achieved, stop the iteration. After convergence, the final dynamic data flow coupling topology graph is obtained. In this topology graph, the vertex positions reflect the link congestion status and protocol adaptation degree of each terminal node in real time. Terminal nodes with high congestion levels move away from other congested nodes due to repulsion, while moving closer to the core vertex under the influence of attraction. Terminal nodes with good protocol adaptation have higher coupling between the core vertex and the edge. Only highly coupled edges are retained between terminal nodes to avoid network redundancy.

[0041] By extracting feature sequences and label sequences from the dynamic data attribute representation matrix, the link congestion degree and protocol matching degree are quantitatively calculated. Through iterative rearrangement, a dynamic data flow coupling topology graph is generated, which transforms the abstract link state and protocol adaptation relationship into an intuitive spatial topology structure, reflecting the distribution and diffusion trend of link congestion caused by sudden traffic from IoT terminals in the factory area, and improving the pertinence and timeliness of intelligent scheduling.

[0042] like Figure 3 As shown, in a preferred embodiment of the present invention, based on the dynamic data flow coupling topology graph, the time-varying characteristics of data link congestion between topology nodes and the distribution of protocol matching degree are extracted to obtain the logical space state offset and service flow diffusion gradient. Based on the logical space state offset and service flow diffusion gradient, the gradient diffusion path and bandwidth allocation change rate vector of service flow metadata in the dynamic data flow coupling topology graph are calculated to obtain the predicted scheduling trajectory sequence, which may include:

[0043] In this embodiment of the invention, based on a dynamic data flow coupled topology graph, joint time series analysis is performed on the coupling weight coefficients between topology nodes and the potential field strength scalars of corresponding vertices to extract the time-varying feature sequence of data link congestion. Specifically, this includes: setting a rolling time window based on a system-preset continuous time window (with a value range of 5s-30s to adapt to the burst characteristics of IoT terminal traffic in the factory area), with the rolling step size kept completely synchronized with the data sampling time interval to ensure the continuity and real-time performance of the time series and avoid data loss or time deviation. For each topology vertex (corresponding to an IoT terminal or edge scheduling center node) in the dynamic data flow coupled topology graph, the potential field strength scalar of the vertex is continuously collected in chronological order within each rolling time window. At the same time, the coupling weight coefficients of the edges connecting the vertex with all adjacent topology nodes and the core node of the edge scheduling center are collected simultaneously. Two independent original time series are constructed: the original time series of potential field strength scalars and the original time series of coupling weight coefficients. Due to sudden traffic fluctuations in the factory's IoT terminals, instantaneous noise may be present in the original time series, affecting the accuracy of subsequent feature extraction. Therefore, a sliding window smoothing algorithm is used to filter the two sets of original time series to eliminate instantaneous noise interference. The smoothing window length is set to 5 sampling intervals (i.e., 500ms), which effectively filters out noise while preserving the dynamic changes of the time series features. The smoothing calculation formula is as follows: ,in For the first The scalar (dimensionless) potential field strength of a topological vertex after smoothing within the t-th rolling time window. For the first The topological vertex in the th ... The original potential field strength scalar (dimensionless) within a rolling time window. For the first The topological vertex and the first The coupling weight coefficient (dimensionless, with a value range of [0, 1]) of the edges connecting the topological vertices after smoothing in the t-th rolling time window. For the edge in the th... The original coupling weight coefficients (dimensionless, with values ​​ranging from [0, 1]) within each rolling time window. In order to be with the first The topological node numbers adjacent to each vertex.

[0044] Based on the linear mapping relationship between the potential field strength scalar and data link congestion in artificial potential field theory (the potential field strength scalar and congestion are positively correlated), the smoothed potential field strength scalar time series data... Inversely convert to the real-time congestion level of the corresponding data link. The mapping formula is ,in is the potential field strength amplification factor (value is 10, dimensionless), used to map the potential field strength scalar to a value range ([0, 1]) that corresponds to the actual link congestion level. This is applied to the calculated real-time congestion level value. The core temporal features are extracted for each time window, including the congestion rise rate (the increase in congestion per unit time), the congestion fall rate (the decrease in congestion per unit time), the congestion peak (the maximum value of congestion within a certain time period), and the congestion duration (the number of consecutive time windows in which congestion exceeds a preset threshold). These features are arranged in chronological order to form a complete sequence set, and finally, the time-varying feature sequence of data link congestion is obtained.

[0045] Based on the time-varying feature sequence, and combined with the data encapsulation format label sequence corresponding to each terminal node, a logical space coordinate mapping is performed to obtain the protocol matching degree distribution density among the topology nodes. Specifically, this includes: constructing a two-dimensional abstract logical space to realize the spatial mapping between link congestion characteristics and protocol types, where the horizontal axis (x-axis) of the logical space is defined as the time-varying feature value of the data link congestion degree, taking values ​​in the range [0, 1], and correlated with the congestion degree numerical value. To maintain consistency, the vertical axis (y-axis) of the logical space is defined as the encoded value of the data encapsulation format label. To achieve quantitative analysis of protocol types, different data encapsulation protocols are uniformly encoded: MQTT is encoded as 1, HTTP as 2, CoAP as 3, and no data transmission (empty label) as 0. All encoded values ​​are dimensionless integers to ensure that protocol types can participate in subsequent spatial distribution calculations. Following the principle of one-to-one correspondence between time windows, the time-varying feature sequences are... Congestion level for each time window The data encapsulation format tag encoding value of the terminal within the corresponding time window. The matching is performed, mapping the points to discrete coordinates in a two-dimensional logical space. ), where t is the rolling time window number, and each discrete coordinate point corresponds to a congestion degree-protocol type combination, realizing the spatial binding of congestion time-series characteristics in the time dimension and label information in the protocol type dimension. To characterize the adaptation distribution law of various data encapsulation protocols under different congestion states, a two-dimensional Gaussian kernel density estimation algorithm is used to perform continuous spatial fitting on all discrete coordinate points, solving for the protocol matching degree distribution density corresponding to any coordinate point in the logical space. This algorithm can effectively fit the continuous distribution characteristics of discrete data and adapt to the dynamic characteristics of protocol matching degree changing with congestion degree. The calculation formula is f(x,y)= The Gaussian kernel function K(⋅) is used to achieve smooth weighting of discrete coordinate points, ensuring the continuity and rationality of the distribution density calculation. Its expression is K(z) = Where f(x,y) is the protocol matching degree distribution density (dimensionless, ranging from [0,1]) corresponding to the coordinate point (x,y) in the two-dimensional logical space. The larger the density value, the higher the frequency of the congestion degree-protocol type combination corresponding to the coordinate point, and the better the protocol matching degree. n is the total number of discrete coordinate points, consistent with the total number of rolling time windows. h1 is the kernel function bandwidth (dimensionless, with a value of 0.1), used to control the smoothness of density estimation and avoid overfitting or underfitting. Let be the dimensionless characteristic value of the link congestion level in the t-th time window. is the dimensionless data encapsulation format label encoding value for the t-th time window. z is the input parameter of the kernel function (dimensionless). By traversing point by point, the protocol matching degree distribution density is calculated for all coordinate points (x∈[0,1],y∈{0,1,2,3}) in the two-dimensional logical space, forming a continuous and complete protocol matching degree spatial distribution map, and finally obtaining the protocol matching degree distribution density f(x,y) between topology nodes.

[0046] Orthogonal feature coupling operations are performed on the time-varying feature sequence of data link congestion and the protocol matching degree distribution density to obtain the logical space state offset. Specifically, this includes: performing dimensionless normalization on the two input features to ensure that their value ranges are consistent and to eliminate the computational bias caused by dimensional differences. The extreme value normalization method is used to map all congestion values ​​in the sequence to the interval [0, 1]. For the protocol matching degree distribution density f(x, y), the same extreme value normalization method is used to map the global distribution density values ​​to the interval [0, 1], resulting in the normalized protocol matching degree distribution density. The normalized time-varying feature sequence and the normalized protocol matching degree distribution density are used as two orthogonal feature vectors. An orthogonal coupling coefficient and a nonlinear adjustment term are introduced, and orthogonal feature coupling operation is performed to solve for the logical space state offset of the topology node within each time window. This offset characterizes the degree to which the state change of the topology node in the logical space deviates from the baseline state, reflecting the collaborative change characteristics of link congestion and protocol adaptation. The calculation formula is as follows: ,in For the first The logical space state offset of a topological node in the t-th time window (dimensionless, with a value range of [0, 1.5]). For the first The time-varying characteristic value (dimensionless) of the normalized link congestion degree of each topological node within the t-th time window. For the first The normalized protocol matching degree distribution density (dimensionless) of the topological nodes in the t-th time window, where = (Normalized congestion eigenvalues) = (Data encapsulation format label encoding value) The orthogonal coupling coefficient (dimensionless, with a value of 0.5) is used to adjust the coupling weight between congestion time-varying characteristics and protocol matching degree distribution characteristics, highlighting the dominant role of congestion time-varying characteristics while also taking into account the influence of protocol matching degree distribution. This is a non-linear adjustment term (dimensionless, with a value range of [0, 1]), used to correct the offset deviation when the congestion is extremely low or extremely high, ensuring that the offset can accurately reflect the actual state changes of the topology node. The logical space state offset of the i-th topology node is calculated one time window at a time in chronological order, and all offsets are arranged by time sequence number to form the logical space state offset sequence of the topology node.

[0047] A gradient operator is applied to the logic space state offset along the scalar decay direction of the potential field strength in the dynamic data flow coupled topology graph to obtain the business flow diffusion gradient. Specifically, this includes: constructing a three-dimensional spatial surface of the potential field strength, which is built based on the spatial topology parameters and the potential field strength scalar of the dynamic data flow coupled topology graph, using the two-dimensional planar coordinates of each topology vertex in the dynamic data flow coupled topology graph. , (Unit: m) are used as the x-axis and y-axis coordinates of the three-dimensional potential surface, and the potential field strength scalar of each topological vertex after smoothing is used as the z-axis coordinate of the three-dimensional potential surface, forming the three-dimensional potential space surface Z= (x, y), where (x, y) is the spatial distribution function of the potential field strength scalar. This surface can intuitively reflect the distribution law of the potential field strength in the topological space. According to the artificial potential field theory, the potential field strength scalar gradually decays along the direction from the core node of the edge scheduling center to the surrounding IoT terminal nodes. The tilt direction of the surface is the potential direction of business flow diffusion. By incorporating the surface normal vector calculation algorithm, the normal vector of the potential field space surface at each topological vertex is calculated. The surface normal vector can characterize the tilt direction of the surface at that point, which is completely consistent with the gradient direction of business flow diffusion. The detailed calculation process of the surface normal vector is as follows: Calculate the potential field space surface Z= The first-order partial derivative of (x, y) at each topological vertex, i.e., the partial derivative in the x-direction. and y-direction partial derivative The partial derivatives are used to characterize the rate of inclination of the surface along the x-axis and y-axis, reflecting the rate of change of the potential field strength scalar in space. The calculation formula is as follows: ,in Let be the first partial derivative of the potential field surface in the x-direction at the coordinate point (x, y) (dimensionless, unit: 1 / m). y is the first-order partial derivative in the y-direction (dimensionless, unit is 1 / m). , The step size for calculating partial derivatives (in meters, with a value of 0.01m) is kept consistent with the units of the two-dimensional plane coordinates of the topological vertices to ensure that the calculation of partial derivatives has no unit deviation. , These represent the positive and negative offsets of the vertex (x, y) along the x-axis. The corresponding potential field strength scalar (dimensionless). , These represent the positive and negative offsets of the vertex (x, y) along the y-axis. The corresponding potential field strength scalar (dimensionless) after offset is obtained by linear interpolation. The linear interpolation formula is as follows: This ensures the accuracy of partial derivative calculations.

[0048] Based on the calculated first-order partial derivatives, the original normal vectors of the potential space surface at the topological vertex are constructed. The direction of the normal vector points towards the convexity of the surface, which is the opposite direction of the potential field intensity decay. To ensure that the direction of the normal vector is consistent with the direction of business flow diffusion (the direction of potential field intensity decay), the partial derivative components are negatively represented. The calculation formula is as follows: ,in This is the original three-dimensional normal vector (dimensionless), with three components corresponding to the normal vector components in the x, y, and z directions, respectively. , The x-axis and y-axis normal vector components (dimensionless) are negativeed, pointing towards the direction of potential field intensity decay, consistent with the direction of business flow diffusion. The z-axis component is 1 (dimensionless), used to maintain the three-dimensional spatial characteristics of the normal vector, ensuring the accuracy of the direction representation and avoiding direction deviation caused by two-dimensional projection. Normalization is performed to eliminate the influence of the normal vector magnitude, resulting in a unit normal vector that retains only the direction information. Normalization ensures that the normal vectors of vertices with different topologies can be directly used for gradient mapping, avoiding gradient magnitude deviations caused by differences in magnitude. The normalization formula is: ,in Original normal vector The modulus (dimensionless) is calculated using the following formula: The unit normal vector after normalization The modulus is 1 (dimensionless), retaining only directional information to represent the direction of business flow diffusion. After calculating the surface normal vector, the gradient operator is mapped to the logic space state offset. The logic space state offset is used as the gradient magnitude and vector-coupled with the unit normal vector to obtain the business flow diffusion gradient. This gradient contains both the direction of business flow diffusion (determined by the unit normal vector) and the driving force of business flow diffusion (determined by the state offset). The business flow diffusion gradient of each topology node is calculated sequentially in chronological order, and all gradients are arranged by time sequence number to form the business flow diffusion gradient sequence of each topology node.

[0049] Based on the logical space state offset and the business flow diffusion gradient, the logical space state offset is mapped to the starting reference point for topology optimization. Using the vector direction of the business flow diffusion gradient as a potential energy conduction guide, the link connectivity of adjacent topology nodes is traversed step-by-step along the conduction guide in the dynamic data flow coupling topology graph. The decay sequence of link coupling weights is extracted and an accumulation convergence judgment is performed. Continuous links falling into a preset convergence interval are combined into gradient diffusion paths. Specifically, this includes: mapping the logical space state offset to the starting reference point for topology optimization, with the mapping rule strictly following the principle of prioritizing highly congested nodes; and selecting the logical space state offset... The topology node with the largest state offset is used as the starting reference point for topology optimization. A larger state offset indicates higher link congestion and a more severe deviation from the baseline state, making it a core node for sudden traffic surges in the factory's IoT terminals and the starting source for outward business flow diffusion. This aligns with the actual need for business flow to diffuse from highly congested nodes to less congested nodes. If multiple topology nodes have the same and maximum state offset values, the node with the largest potential field strength scalar is selected as the starting reference point. This further ensures that the starting reference point is a highly congested core node, avoiding unreasonable diffusion paths caused by biased starting point selection. The vector direction of the business flow diffusion gradient is used as the sole guiding direction for potential energy transmission. That is, along the direction of the business flow diffusion gradient (unit normal vector direction), starting from the starting reference point in the dynamic data flow coupling topology graph, the surrounding adjacent topology nodes are traversed layer by layer. Simultaneously, the link connectivity between adjacent nodes is checked to ensure that the traversed links are valid connected links. The traversal rule prioritizes traversing the coupling weight coefficient of the edges connected to the current node. Neighboring nodes with a potential field strength ≥ 0.3 (dimensionless) are selected. This threshold has been verified through extensive experiments to ensure protocol compatibility and data transmission stability. Simultaneously, neighboring nodes with potential field strength scalars smaller than the current node are prioritized for traversal. These nodes have lower link congestion levels than the current node, representing the optimal direction for service flow diffusion and effectively alleviating pressure on highly congested nodes.

[0050] During the step-by-step traversal, the coupling weight coefficients of each effective link are continuously extracted in traversal order. These coupling weight coefficients are arranged in traversal order to form a decay sequence of link coupling weights. Since the traversal direction is the direction of potential field strength decay, the protocol matching degree of the link will decrease slightly. Therefore, the coupling weight coefficients show a gradual decay trend with the traversal direction. This sequence can reflect the dynamic change law of link matching degree. An accumulation convergence judgment is performed on the decay sequence of link coupling weights to determine the endpoint of service flow diffusion and avoid unlimited extension of the traversal process. The convergence judgment consists of two steps: first, the cumulative sum of the decay sequence is calculated; then, the cumulative sum is compared with a preset convergence threshold. If the cumulative sum falls within the preset convergence interval, the traversal stops. The calculation formula is as follows: ,in It is the cumulative sum (dimensionless) of the link coupling weight decay sequence, reflecting the total matching degree of the traversed links. A convergence threshold (dimensionless, ranging from [2, 3]) is added to the coupling weights. This threshold is dynamically adjusted based on the network size of the dynamic data flow coupling topology graph. The larger the network size (the more terminal nodes), the larger the threshold value, ensuring the rationality of the convergence determination. The convergence error (dimensionless, with a value of 0.1) is used to control the precision of convergence determination and avoid premature or late termination of traversal due to small deviations. When the sum of coupled weights falls into the preset convergence interval, the topology node traversal process is immediately terminated, and all continuous valid links traversed during the traversal are combined and spliced ​​to form a complete gradient diffusion path for business flow metadata. If the edge scheduling center core node (the node with the largest potential field strength scalar) has been reached during the traversal but the sum still does not fall into the convergence interval, the edge scheduling center core node is used as the endpoint to form a gradient diffusion path. If there are no adjacent nodes that meet the conditions (no adjacent nodes with coupled weight coefficients ≥ 0.3 and lower potential field strength) during the traversal, the current node is used as the endpoint to ensure that each starting reference point corresponds to a valid gradient diffusion path.

[0051] The physical topology spacing parameters of each link segment in the gradient diffusion path are extracted sequentially along the node extension sequence. These physical topology spacing parameters are then perturbationally coupled with the potential field strength scalars of the corresponding nodes on the gradient diffusion path to obtain an intermediate adjustment sequence characterizing the dynamic scaling of link bandwidth resources. This intermediate adjustment sequence is then subjected to temporal difference processing and vector normalization mapping to obtain the bandwidth allocation change rate vector. Specifically, this involves extracting the physical topology spacing parameters between adjacent topology nodes in each segment of the path sequentially along the node extension sequence of the gradient diffusion path (from the starting reference point to the ending point). ,in Let be the physical straight-line distance between the p-th node and the q-th node (adjacent nodes) in the gradient diffusion path, calculated using the two points' planar coordinates. The calculation formula is: ,in( , Let be the two-dimensional planar coordinates of the p-th node. , Let q be the two-dimensional planar coordinates (in meters) of the q-th node. The coordinate parameters are consistent with the node coordinates in the dynamic data flow coupled topology graph to ensure the accuracy of the spacing calculation. Simultaneously, the potential field strength scalar corresponding to each node on the gradient diffusion path is extracted, i.e., the potential field strength scalar of the p-th node. scalar of potential field strength at the q-th node Physical topology spacing parameters Scalar of potential field strength at corresponding nodes , Perturbation coupling calculation is performed to reflect the combined impact of physical spacing and link congestion status on bandwidth allocation. The dynamic scaling of bandwidth resources for each link segment is calculated, and the scaling of all link segments is arranged in path order to form an intermediate adjustment sequence characterizing the dynamic scaling of link bandwidth resources. The perturbation coupling calculation formula is as follows: ,in This represents the dynamic scaling range of the bandwidth of the link segment between the p-th node and the q-th node in the gradient diffusion path (dimensionless, ranging from [0, 1]). The larger the value, the greater the bandwidth adjustment required for this link segment. This is the bandwidth adjustment factor (dimensionless, with a value of 0.8), used to adjust the range of bandwidth scaling to ensure that the scaling range matches the actual network bandwidth adjustment capability. The difference in potential field strength between the p-th node and the q-th node (dimensionless) indicates that the larger the difference is, the greater the difference in the link congestion level between the two nodes, the greater the bandwidth scaling range, and the need to balance the link load through bandwidth adjustment. The physical topology spacing between two nodes is the larger the spacing, the smaller the bandwidth scaling range. This is because the longer the transmission distance, the smoother the bandwidth adjustment, avoiding data transmission instability caused by sudden bandwidth changes. This is the coupling weight coefficient (dimensionless) of the edge connecting the p-th node and the q-th node. The disturbance adjustment term (dimensionless, with a value range of [0, 1]) is a variable. The larger the coupling weight coefficient, the larger the value of the disturbance adjustment term, and the more stable the bandwidth scaling, reflecting the constraint effect of protocol matching degree on bandwidth adjustment.

[0052] The bandwidth of each link segment is dynamically scaled according to the link segment order of the gradient diffusion path. Arranged sequentially, an intermediate adjustment sequence is formed, reflecting the dynamic scaling distribution of bandwidth resources along the entire gradient diffusion path. Temporal differential processing is performed on the intermediate adjustment sequence to calculate the difference in bandwidth scaling amplitude between adjacent link segments, resulting in a temporal differential sequence. This sequence characterizes the rate of change of bandwidth scaling amplitude and reflects the dynamic trend of bandwidth adjustment. Performing vector normalization mapping eliminates numerical differences between elements in the sequence, yielding a bandwidth allocation change rate vector. This vector characterizes the relative rate of bandwidth change for each link segment. The normalization calculation formula is as follows: ,in Assign a rate of change vector (dimensionless) to the normalized bandwidth with a magnitude of 1, retaining only the relative rate and trend of bandwidth change. For time-series difference sequences The modulus (dimensionless) is calculated using the following formula: ,in It is the i-th element (dimensionless) in the time-difference sequence.

[0053] The temporal evolution nodes of the dynamic scaling amplitude of each bandwidth resource in the bandwidth allocation change rate vector are extracted. These nodes are then superimposed with the spatial topological coordinates of the gradient diffusion path using displacement compensation to construct a spatiotemporal coupled state matrix. This matrix is ​​then discretized, sliced, and reassembled according to a preset scheduling time window to obtain the predicted scheduling trajectory sequence. Specifically, this includes: extracting the bandwidth allocation change rate vector. The temporal evolution nodes corresponding to each component are defined as points in time when the dynamic scaling of bandwidth resources changes significantly. The criterion is that when the absolute value of a component of the bandwidth allocation change rate vector is greater than 0.5 (dimensionless), the time window corresponding to that component is a temporal evolution node, indicating that the bandwidth scaling has changed drastically at that moment, requiring close monitoring and adjustment of the scheduling strategy. All temporal evolution nodes are then organized into a temporal evolution node sequence in chronological order. The temporal evolution node sequence is superimposed with the spatial topological coordinates of the gradient diffusion path using displacement compensation. The purpose of displacement compensation is to correct the spatial position deviation corresponding to the temporal evolution nodes. The formula for displacement compensation superposition is as follows: ,in, For the first Spatial topological coordinates of each temporal evolution node after displacement compensation. For the first The original spatial coordinates on the gradient diffusion path corresponding to each temporal evolution node. This is the displacement compensation coefficient (valued at 0.05 m / ms), used to adjust the magnitude of displacement compensation. It is matched with the time unit and spatial coordinate unit to ensure that the dimensions of the displacement compensation amount are consistent with the coordinate unit. Assign a rate of change vector to the bandwidth The first in Each component (dimensionless) reflects the rate of bandwidth change corresponding to the time series evolution node.

[0054] Based on the displacement-compensated spatiotemporal coordinates, a spatiotemporal coupling state matrix is ​​constructed. This matrix integrates the temporal dimension features (temporal evolution nodes) with spatial dimension features (topological coordinates) and bandwidth dimension features (bandwidth allocation change rate) of the service flow, comprehensively representing the spatiotemporal coupling characteristics of service flow scheduling. The row index of the matrix represents the temporal evolution node, with each row corresponding to one temporal evolution node. The column indexes are divided into three categories: the displacement-compensated x-axis coordinate, the displacement-compensated y-axis coordinate, and the corresponding bandwidth allocation change rate. The matrix expression is S = The spatiotemporal coupled state matrix S is discretized and reassembled according to a preset scheduling time window. The length of the preset scheduling time window is consistent with the continuous time window T (5s-30s) to ensure the uniformity of the scheduling time scale. The slicing rule is to divide the spatiotemporal coupled state matrix into several time segments of length T in chronological order. Each time segment corresponds to an independent slice matrix, which contains the spatiotemporal state information of all time-series evolution nodes within that time segment. Then, all slice matrices are reassembled and arranged in chronological order to form a complete sequence set, ultimately obtaining the predicted scheduling trajectory sequence.

[0055] Improving the timeliness of collaborative sensing and the accuracy of intelligent scheduling of IoT terminals can effectively alleviate link congestion, optimize the efficiency of bandwidth resource allocation across the entire network, reduce data transmission latency, and ensure the stable operation of IoT terminal services in the factory area.

[0056] In a preferred embodiment of the present invention, based on the predicted scheduling trajectory sequence, state transition threshold determination and scheduling inertia compensation control are performed on each trajectory point to dynamically trigger the diversion command of the target data transmission channel. Based on the gradient diffusion path and bandwidth allocation change rate vector, the expected transmission delay of each link is calculated, and the actual transmission delay after each node executes the current diversion command is collected in real time. The difference between the actual transmission delay and the expected transmission delay is used as the delay deviation feedback. The delay deviation feedback is used to perform gradient iterative correction on the scheduling inertia compensation control to determine the collaborative sensing scheduling execution strategy and distribute it to each IoT terminal. This may include:

[0057] In this embodiment of the invention, based on the predicted scheduling trajectory sequence, the link load status parameters corresponding to each trajectory point in the sequence are extracted point by point. The link load status parameters are compared with a preset state transition threshold on a cycle-by-cycle basis to identify the critical trajectory points where the load state change rate crosses the state transition threshold. The critical trajectory points are arranged in chronological order and combined into a target transition trajectory point set. Specifically, this includes: extracting the link load status parameters corresponding to each trajectory point in the sequence point by point along the chronological order of the predicted scheduling trajectory sequence. The link load status parameters mainly include three core indicators: link utilization, bandwidth occupancy, and data queue length. All of these are dimensionless parameters with values ​​ranging from [0, 1]. Link utilization is obtained by the ratio of the actual amount of data transmitted by the current link to the maximum transmission capacity of the link. Bandwidth occupancy is obtained by the ratio of the bandwidth occupied by the current link to the total bandwidth of the link. Data queue length is obtained by the ratio of the number of data frames waiting to be transmitted in the link buffer queue to the maximum capacity of the buffer queue. The three parameters work together to characterize the real-time load status of the link, avoiding deviations caused by a single parameter.

[0058] Based on the business characteristics and link transmission threshold requirements of IoT terminals in the factory area, three sets of state transition thresholds are preset, corresponding to link utilization transition threshold, bandwidth occupancy transition threshold, and data queuing length transition threshold, respectively. The threshold values ​​have been verified through extensive experiments to adapt to the load change requirements of different business scenarios. Specifically, the link utilization transition threshold is set to 0.7, the bandwidth occupancy transition threshold to 0.65, and the data queuing length transition threshold to 0.6, all of which are dimensionless parameters. The extracted link load state parameters of each trajectory point are compared with the corresponding preset state transition thresholds on a cycle-by-cycle basis. The comparison cycle is consistent with the system scheduling cycle (set to 100ms) to ensure real-time judgment. The comparison rule is that when the rate of change of any link load state parameter of a trajectory point crosses the corresponding preset state transition threshold, the trajectory point is determined to be a critical trajectory point where the rate of change of load state change has abruptly changed. That is, the link load corresponding to the trajectory point changes from a stable state to an unstable state, or from an unstable state to a stable state, requiring subsequent scheduling inertia compensation and channel adjustment. All identified critical trajectory points are arranged in chronological order, and duplicate or invalid trajectory points (such as trajectory points misjudged due to instantaneous noise) are removed, forming a set of target transition trajectory points.

[0059] Based on the set of target transition trajectory points, the state offset residuals of each target transition trajectory point in the set are extracted between the current scheduling cycle and the previous scheduling cycle. Momentum attenuation filtering and damping smoothing calculations are then performed on the state offset residuals to obtain the scheduling inertia compensation control quantity. Specifically, this includes: extracting the state offset residuals of each target transition trajectory point in the set between the current scheduling cycle and the previous scheduling cycle. The state offset residual refers to the difference in link load state parameters of the same target transition trajectory point in two adjacent scheduling cycles, used to characterize the degree of state offset caused by inertia during scheduling. The extraction rule is to calculate the difference in link utilization, bandwidth occupancy, and data queuing length between the current and previous scheduling cycles for each target transition trajectory point. These three differences together constitute the state offset residual vector of that trajectory point. Since the state offset residuals may contain instantaneous noise (such as abnormal differences caused by instantaneous link fluctuations), momentum attenuation filtering is performed on the state offset residuals to filter out noise interference and retain the core trend of residual change. The momentum attenuation filtering calculation formula is as follows: ,in For the first The state offset residuals (dimensionless) of the target transition trajectory points after momentum decay filtering. The momentum decay coefficient (dimensionless, with a value of 0.2) is used to control the influence weight of historical residuals. The smaller the value, the greater the influence of the current residuals and the faster the filtering response. For the first The state offset residual (dimensionless) after filtering over a scheduling cycle at each target transition trajectory point. For the first The original state offset residuals (dimensionless) of each target transition trajectory point in the current scheduling cycle. Even after momentum decay filtering, the residuals may still exhibit slight fluctuations, requiring further damping smoothing calculations to reduce these fluctuations and ensure the stability of subsequent compensation control quantities. The damping smoothing calculation formula is as follows: ,in For the first The state offset residuals (dimensionless) of the target jump trajectory points after damped smoothing calculation. The damping coefficient (dimensionless, with a value of 0.15) is used to control the smoothing effect. The larger the value, the more obvious the smoothing effect and the smaller the residual fluctuation. The state offset residual (dimensionless) after momentum decay filtering. This is the filtered state offset residual (dimensionless) from the previous scheduling cycle.

[0060] The state offset residuals after momentum decay filtering and damping smoothing are mapped to scheduling inertia compensation control quantities. The mapping rule is to perform a dot product operation between the smoothed residual vector and a preset compensation coefficient matrix to obtain the scheduling inertia compensation control quantity corresponding to each target transition trajectory point. This control quantity is used to correct inertial deviations during the scheduling process, ensuring that channel adjustments can adapt to changes in link load in a timely manner. The mapping calculation formula is as follows: ,in For the first The scheduling inertia compensation control quantity (dimensionless, value range [0, 0.5]) corresponds to each target transition trajectory point. K is the compensation coefficient matrix (dimensionless), a 3×3 diagonal matrix. The diagonal elements correspond to the compensation weights of link utilization, bandwidth occupancy, and data queuing length, with values ​​of 0.4, 0.3, and 0.3, respectively, used to balance the influence of the three load parameters on the compensation control quantity. The state offset residual vector after smoothing calculation (dimensionless) is used to calculate the scheduling inertia compensation control quantity corresponding to all target transition trajectory points point by point, forming a set of scheduling inertia compensation control quantities.

[0061] Based on the scheduling inertia compensation control, the initial data transmission channels associated with the target transition trajectory points are topologically offset corrected. Channel redirection mapping is then performed in conjunction with the link connectivity weights in the dynamic data flow coupling topology graph to obtain the final channel matching identifier. Based on this final channel matching identifier, the target data transmission channel splitting command is dynamically triggered. Specifically, this includes: determining the initial data transmission channel associated with each target transition trajectory point. The initial data transmission channel is the transmission channel corresponding to the link segment where the trajectory point is located in the gradient diffusion path. Each target transition trajectory point corresponds to a unique initial data transmission channel. The channel identifier is completely consistent with the link identifier in the dynamic data flow coupling topology graph, ensuring the accuracy of channel association and avoiding splitting anomalies caused by incorrect channel matching. Based on the scheduling inertia compensation control, the initial data transmission channels are topologically offset corrected. The core purpose of offset correction is to adjust the spatial orientation of the data transmission channels, offsetting the channel offset caused by scheduling inertia, ensuring that the data transmission channels can match the dynamic changes in link load, and improving channel transmission efficiency. The magnitude of the topology offset correction is determined by the scheduling inertia compensation control; the larger the control, the larger the offset correction, and the faster it can offset the deviation caused by scheduling inertia, ensuring the accuracy of channel orientation.

[0062] After completing the topology offset correction, the corrected channels are redirected and mapped using the link connectivity weights in the dynamic data flow coupling topology graph. First, all available transmission channels adjacent to the corrected channels are extracted from the dynamic data flow coupling topology graph. Available channels with a link connectivity weight ≥ 0.3 (dimensionless) are selected; this threshold ensures channel connectivity and transmission stability. Then, the spatial matching degree between each available channel and the corrected channel is calculated. The spatial matching degree is calculated using the cosine of the angle between the link direction corresponding to the channel and the unit normal vector direction. The smaller the angle, the higher the spatial matching degree. The matching degree calculation formula is as follows: ,in For the first After the correction of the target jump trajectory point, the channel is the same as the first target jump trajectory point. Spatial matching degree of available channels (dimensionless, value range [0, 1]). unit normal vector With available channel link direction vector The included angle (in radians). For the first Link direction vectors of available channels (dimensionless). , are the magnitudes (dimensionless) of the unit normal vector and the link direction vector, respectively. =1. Select the available channel with the highest spatial matching degree as the final matching channel, and generate a corresponding final channel matching identifier. The channel matching identifier is a unique digital code that corresponds one-to-one with each transmission channel to ensure the uniqueness of channel identification. Based on the final channel matching identifier, dynamically trigger the diversion command for the target data transmission channel. The diversion command includes three core parameters: channel matching identifier, diversion ratio, and diversion timing. The diversion ratio is determined based on the scheduling inertia compensation control amount. The larger the compensation control amount, the higher the diversion ratio, ensuring that diversion can alleviate the link load pressure in a timely manner. The triggering timing of the diversion command is synchronized with the scheduling cycle to avoid the diversion process affecting the continuity of data transmission. Process all target transition trajectory points point by point to complete the corresponding channel offset correction, redirection mapping, and diversion command triggering.

[0063] Based on the gradient diffusion path and the bandwidth allocation change rate vector, the physical topology span parameter of each link segment in the path is extracted. The bandwidth allocation change rate vector is mapped to the instantaneous effective throughput capacity threshold of the corresponding link segment. Combining the physical topology span parameter and the instantaneous effective throughput capacity threshold, the data flow spatial propagation time and the queuing dwell time of the link nodes are superimposed to obtain the expected transmission delay benchmark sequence for each link. Specifically, this includes: extracting the physical topology span parameter of each link segment along the node extension order of the gradient diffusion path. The physical topology span parameter is the physical straight-line distance of each link segment, in meters, calculated using the two-dimensional plane coordinates of the topological vertices at both ends of the link segment. The calculation formula is as follows: ,in Let be the physical topological span of the link segment between the p-th node and the q-th node in the gradient diffusion path. Let be the two-dimensional planar coordinates (in meters) of the p-th node. Let be the two-dimensional planar coordinates (in meters) of the q-th node, consistent with the node coordinates in the dynamic data flow coupled topology diagram to ensure the accuracy of span calculation. The bandwidth allocation change rate vector is mapped to the instantaneous effective throughput capacity threshold of the corresponding link segment. The instantaneous effective throughput capacity threshold refers to the upper limit of the amount of data that a link segment can stably transmit within the current scheduling cycle, and is positively correlated with the bandwidth allocation change rate. The mapping calculation formula is: ,in This is the instantaneous effective throughput capacity threshold (in Mbps) for the pq-th link segment. This is the maximum throughput capacity of the link segment (in Mbps, set according to the link hardware specifications, uniformly 100Mbps). The component corresponding to the pq-th link segment in the bandwidth allocation rate vector (dimensionless, with a value range of [0, 1]) is assigned. The larger the component, the higher the instantaneous effective throughput capacity threshold, adapting to the dynamic changes in link load.

[0064] By combining the physical topology span parameter and the instantaneous effective throughput capacity threshold, the expected transmission delay of each link segment is obtained by superimposing the data flow spatial propagation time and the queuing dwell time of the link nodes. The data flow spatial propagation time refers to the time it takes for data to travel from one end of the link segment to the other, and is directly proportional to the physical topology span and inversely proportional to the transmission rate. The transmission rate is set at 80% of the instantaneous effective throughput capacity threshold (to ensure transmission stability). The propagation time calculation formula is as follows: ,in The time taken for the data stream to propagate in space in the pq-th link segment. This represents the actual transmission rate of the link segment. The rate conversion factor (valued at 100) is used to convert the throughput capacity threshold into the transmission rate, ensuring unit matching. Link node queuing dwell time refers to the time data waits for transmission in the buffer queues of nodes at both ends of the link segment. It is related to the link load status and the instantaneous effective throughput capacity threshold; the higher the load and the lower the throughput capacity, the longer the queuing dwell time. The formula for calculating queuing dwell time is... ,in The queuing dwell time of the link nodes in the pq-th link segment. This represents the average amount of data (in MB) in the cache queues of the nodes at both ends of this link segment. The instantaneous effective throughput capacity threshold (in Mbps) is defined by a coefficient of 1000 to convert the time unit from seconds to milliseconds, ensuring dimensional consistency. The expected transmission delay for each link segment is obtained by superimposing the data stream spatial propagation time with the queuing dwell time of the link nodes. The calculation formula is as follows: ,in is the expected transmission delay (in milliseconds) for the pq-th link segment. The time taken for the data stream to propagate in space (in milliseconds). Given the queuing dwell time of the link nodes (in milliseconds), the expected transmission delays of each link segment are arranged in chronological order according to the link segment sequence of the gradient diffusion path, forming a baseline sequence of expected transmission delays for each link.

[0065] Using the time boundaries of each expected delay node in the expected transmission delay benchmark sequence as the monitoring trigger window, the system acquires the data flow ingress trigger identifier and egress confirmation identifier of each IoT terminal after executing the current diversion command in real time. Differential accumulation is performed on the timestamps associated with the ingress trigger identifier and egress confirmation identifier to extract the actual transmission delay observation sequence after each node executes the current diversion command. Specifically, the monitoring trigger window uses the time boundaries of each expected delay node in the expected transmission delay benchmark sequence as the monitoring trigger window. The start time of the monitoring trigger window is the start time of the scheduling cycle corresponding to each expected delay node, and the window length is consistent with the scheduling cycle (100ms) to ensure that the monitoring timing corresponds to the expected delay time node, avoiding actual delay calculation errors caused by monitoring time deviations. Within the monitoring trigger window, the monitoring module of the edge scheduling center acquires the data flow ingress trigger identifier and egress confirmation identifier of each IoT terminal after executing the current diversion command in real time. The ingress trigger identifier refers to the trigger signal identifier sent by the terminal when data enters the transmission channel, including the data entry timestamp. The exit confirmation identifier is a confirmation signal sent by the terminal when data transmission is completed and the data leaves the transmission channel. It includes the data departure timestamp. Both identifiers correspond one-to-one with the data stream to ensure the correlation of the timestamps.

[0066] For each data stream, a differential accumulation operation is performed on the timestamps associated with the ingress trigger flag and the egress confirmation flag to extract the actual transmission delay of that data stream. Specifically, first, the difference between the egress timestamp and the ingress timestamp is calculated to obtain the transmission delay of a single data stream. Then, the arithmetic mean of the transmission delays of all data streams within the same link segment and the same monitoring window is calculated to obtain the average actual transmission delay of that link segment within the current monitoring window, eliminating errors caused by instantaneous fluctuations in a single data stream. The formula for calculating the average actual transmission delay is as follows: ,in is the average actual transmission delay (in milliseconds) of the pq-th link segment within the t-th monitoring window. For the first The exit timestamp (in milliseconds) of the data stream within the t-th monitoring window. For the first The entry timestamp (in ms) of each data stream within the t-th monitoring window is used to extract the average actual transmission delay of each link segment window by window according to the time sequence of the monitoring window. All actual transmission delays are arranged by time sequence number to form the actual transmission delay observation sequence of each link.

[0067] The actual delay fluctuation characteristics of each time node in the actual transmission delay observation sequence are extracted. These characteristics are then aligned with the expected transmission delay benchmark sequence through point-by-point topological mapping. Numerical difference calculation and polarity calibration are performed to obtain the delay deviation feedback quantity, which characterizes the degree of deviation in the execution of the current traffic splitting command. Specifically, this includes: extracting the actual delay fluctuation characteristics of each time node in the actual transmission delay observation sequence. These characteristics mainly include two core indicators: the fluctuation amplitude and the fluctuation frequency. The fluctuation amplitude is the absolute value of the difference in actual delay between two adjacent monitoring windows in the same link segment, and the fluctuation frequency is the number of actual delay fluctuations per unit time. These two indicators characterize the stability of the actual delay and assist in the calculation of the deviation feedback quantity. The actual transmission delay observation sequence is then aligned with the expected transmission delay benchmark sequence through point-by-point topological mapping. The alignment rule is a one-to-one correspondence based on time sequence number. The actual delay of the t-th monitoring window in the actual transmission delay observation sequence is matched with the expected delay of the t-th time node in the expected transmission delay benchmark sequence to ensure that the time nodes of each set of comparison data are consistent, avoiding errors in deviation calculation caused by time misalignment.

[0068] The actual latency after alignment is compared with the expected latency to calculate the latency difference. The sign of the latency difference indicates the direction of deviation in the execution of the traffic splitting command. A positive value indicates that the actual latency is greater than the expected latency, indicating insufficient traffic splitting (e.g., a low splitting ratio leading to excessive link load). A negative value indicates that the actual latency is less than the expected latency, indicating excessive traffic splitting (e.g., a high splitting ratio leading to wasted bandwidth resources). The formula for calculating the latency difference is as follows: ,in This represents the delay difference (in milliseconds) of the pq-th link segment within the t-th monitoring window. This represents the actual transmission delay (in milliseconds) of the link segment within the t-th monitoring window. For the expected transmission delay (in milliseconds) of this link segment at time node t, the calculated delay difference is calibrated with positive or negative polarity. The calibration rule is as follows: When ≥0, it is calibrated as a positive deviation and marked as +1. When the value is less than 0, it is calibrated as a negative deviation and marked as −1. The polarity mark and the delay difference together constitute the delay deviation feedback quantity. The delay deviation feedback quantity fully characterizes the degree and direction of deviation of the current split instruction execution. The calculation formula is as follows: ,in Let pq be the delay deviation feedback value of the pq-th link segment within the t-th monitoring window. The delay difference (in milliseconds) is calculated for each link segment and each monitoring window to form a set of delay deviation feedback values.

[0069] The delay deviation feedback quantity is compared with a preset deviation convergence threshold cycle by cycle to identify abnormal link nodes whose delay deviation feedback quantity exceeds the deviation convergence threshold, and the historical correction trajectory of the scheduling inertial compensation control quantity corresponding to the abnormal link node is extracted. Specifically, based on the transmission delay requirements of the factory IoT terminals, a preset delay deviation convergence threshold is set. The convergence threshold is divided into a positive convergence threshold and a negative convergence threshold. The positive convergence threshold is set to 5ms (i.e., the actual delay is no more than 5ms greater than the expected delay), and the negative convergence threshold is set to -3ms (i.e., the actual delay is no more than 3ms less than the expected delay). The threshold setting takes into account the stability of transmission delay and the utilization of bandwidth resources, avoiding frequent corrections due to overly strict thresholds or excessive delay deviations due to overly lenient thresholds. Each delay deviation feedback quantity in the delay deviation feedback quantity set is compared with the preset deviation convergence threshold cycle by cycle. The comparison cycle is consistent with the monitoring window (100ms). The comparison rule is that when the delay deviation feedback quantity of a certain link segment exceeds the threshold, the error rate of the feedback quantity is 100ms. If the deviation exceeds the positive convergence threshold or falls below the negative convergence threshold, the link node corresponding to the link segment is determined to be an abnormal link node. This means the deviation of the traffic splitting command execution at this node exceeds the allowable range, requiring correction of the scheduling inertia compensation control quantity. Identified abnormal link nodes are screened, eliminating those misjudged due to transient interference (such as sudden traffic spikes or temporary link failures), while retaining nodes whose deviations consistently exceed the convergence threshold to ensure accurate identification of abnormal nodes. Historical correction trajectories of the scheduling inertia compensation control quantities corresponding to the abnormal link nodes are extracted. The extraction range covers the most recent 10 scheduling cycles prior to the current scheduling cycle. The extracted content includes the scheduling inertia compensation control quantity, delay deviation feedback quantity, and correction adjustment quantity corresponding to the abnormal node in each historical scheduling cycle, arranged chronologically to form a sequence of historical correction trajectories for that abnormal node.

[0070] Based on historical correction trajectories, a gradient descent direction search is performed on the scheduling inertial compensation control quantity. An adaptive step size adjustment factor is then superimposed on the scheduling inertial compensation control quantity along the gradient descent direction to obtain the corrected scheduling inertial compensation control update quantity. Specifically, this includes: based on the historical correction trajectory sequence of abnormal link nodes, a gradient descent direction search is performed on the scheduling inertial compensation control quantity. The gradient descent direction is used to determine the correction direction of the control quantity, ensuring that the corrected control quantity can bring the time delay deviation closer to the convergence threshold. The formula for calculating the gradient descent direction is as follows: ,in For the first The gradient descent adjustment amount (dimensionless) of the scheduling inertial compensation control quantity for each abnormal link node. The gradient descent learning rate (dimensionless, with a value of 0.05) is used to control the step size of gradient descent. The smaller the value, the smoother the correction, avoiding fluctuations caused by overcorrection. The partial derivative (in milliseconds / dimensionless) of the time delay deviation feedback quantity with respect to the scheduling inertial compensation control quantity is used to characterize the degree of influence of control quantity changes on the time delay deviation. The partial derivative is calculated by the numerical difference between the control quantity and the deviation feedback quantity in the historical correction trajectory sequence. To avoid slow convergence or oscillations during gradient descent, an adaptive step size adjustment factor is superimposed on the scheduling inertial compensation control quantity along the gradient descent direction. The adaptive step size adjustment factor is dynamically adjusted according to the absolute value of the time delay deviation feedback quantity; the larger the deviation, the larger the step size adjustment factor, and the faster the correction speed. The smaller the deviation, the smaller the step size adjustment factor, and the smoother the correction. The formula for calculating the adaptive step size adjustment factor is: ,in For the first The adaptive step size adjustment factor (dimensionless, with a value range of [0.1, 0.15]) for each abnormal link node. It is the absolute value of the time delay difference in the time delay deviation feedback quantity (in milliseconds). Given the expected transmission delay (in milliseconds) corresponding to the abnormal node, and ensuring that the step size adjustment factor matches the delay deviation and the expected delay, the gradient descent adjustment amount is multiplied by the adaptive step size adjustment factor to obtain the correction increment of the scheduling inertial compensation control amount. This correction increment is then added to the current scheduling inertial compensation control amount to obtain the corrected scheduling inertial compensation control update amount. The calculation formula is as follows: ,in For the first The updated scheduling inertia compensation control quantity after correction of each abnormal link node (dimensionless, with a value range of [0, 0.5]). This is the current scheduling inertia compensation control quantity (dimensionless). This is the adaptive step size adjustment factor (dimensionless). The gradient descent adjustment amount (dimensionless) is used to perform the above gradient iterative correction process on the scheduling inertia compensation control amount of all abnormal link nodes one by one to obtain the control update amount of all abnormal nodes. For non-abnormal link nodes, their scheduling inertia compensation control amount remains unchanged and is directly used as the control update amount. The control update amounts of all nodes are organized into a set of scheduling inertia compensation control update amounts.

[0071] The corrected scheduling inertia compensation control update quantity is mapped to the collaborative sensing scheduling execution strategy of each IoT terminal. This includes data acquisition frequency adjustment parameters for terminal nodes, data encapsulation format conversion instructions, and transmission channel priority remapping tables. The collaborative sensing scheduling execution strategy is encapsulated into control command frames and broadcast to each IoT terminal node through the control channel of the edge scheduling center to complete collaborative sensing and intelligent scheduling. Specifically, this includes mapping the corrected scheduling inertia compensation control update quantity to the collaborative sensing scheduling execution strategy of each IoT terminal. The mapping rules are determined based on the magnitude and polarity of the control update quantity and the final channel matching identifier, and are divided into three aspects: First, the channel switching strategy, which determines the target transmission channel of each terminal based on the final channel matching identifier, clarifies the timing and order of channel switching, and ensures that data transmission is not interrupted during the switching process. Second, the bandwidth allocation strategy, which adjusts the bandwidth allocation ratio of each link segment according to the magnitude of the control update quantity. The larger the control update quantity, the larger the bandwidth adjustment range, prioritizing the bandwidth requirements of high-load links. Third, the traffic splitting control strategy, which adjusts the traffic splitting ratio according to the polarity of the control update quantity. Positive deviations correspond to an increase in the traffic splitting ratio, and negative deviations correspond to a decrease in the traffic splitting ratio, so that the actual transmission delay converges to the expected delay.

[0072] The collaborative sensing scheduling execution strategy is encapsulated into control command frames. These frames employ a standardized frame structure, comprising five parts: a frame header, a command identifier, a terminal address, scheduling parameters, and a frame trailer. The frame header, occupying 2 bytes, identifies the start of the command frame. The command identifier distinguishes different scheduling commands (such as channel switching, bandwidth adjustment, and traffic splitting control), occupying 1 byte. The terminal address specifies the IoT terminal receiving the command, occupying 4 bytes. The scheduling parameters include core parameters such as control update amount, final channel matching identifier, bandwidth allocation ratio, and traffic splitting ratio, occupying 8 bytes. The frame tail is used to verify the integrity of the instruction frame, occupying 2 bytes, to ensure that no data loss or bit errors occur during the transmission of the instruction frame. The encapsulated control instruction frame is broadcast to each IoT terminal node through the control channel of the edge scheduling center. The control channel adopts a dedicated low-latency transmission channel with a transmission rate of 10Mbps to ensure that the instruction frame can be transmitted to each terminal quickly and stably. After receiving the instruction frame, each IoT terminal node verifies and parses the instruction frame, extracts the scheduling execution strategy, and adjusts its own transmission channel, bandwidth allocation, and traffic splitting ratio according to the strategy requirements to complete the execution of collaborative sensing scheduling.

[0073] This enables the determination of link load status, effective compensation for scheduling inertia, dynamic adaptation of transmission channels, real-time feedback of latency deviation, and iterative optimization of scheduling strategies, thereby improving the accuracy of collaborative perception and the real-time performance of intelligent scheduling for IoT terminals.

[0074] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0075] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A large-scale model-based IoT terminal collaborative sensing and intelligent scheduling system, characterized in that, include: The module is used to acquire multidimensional data traffic characteristics and business flow metadata of each terminal node in the IoT cluster. It converts the multidimensional data traffic characteristics into data flow polarity vectors through the feature mapping network of the large model and constructs a dynamic data attribute representation matrix. The generation module is used to calculate the data link congestion degree and data encapsulation protocol matching degree between each node and the edge scheduling center based on the dynamic data attribute representation matrix, and to construct a dynamic data flow coupling topology diagram simulating the potential field distribution. The analysis module is used to extract the time-varying characteristics of data link congestion between topology nodes and the distribution of protocol matching degree based on the dynamic data flow coupling topology graph, and obtain the logical space state offset and service flow diffusion gradient; based on the logical space state offset and service flow diffusion gradient, it calculates the gradient diffusion path and bandwidth allocation change rate vector of service flow metadata in the dynamic data flow coupling topology graph, and obtains the predicted scheduling trajectory sequence. The correction module is used to determine the state transition threshold and control the scheduling inertia compensation for each trajectory point according to the predicted scheduling trajectory sequence, and dynamically trigger the diversion command of the target data transmission channel. Based on the gradient diffusion path and the bandwidth allocation change rate vector, the expected transmission delay of each link is calculated, and the actual transmission delay of each node after executing the current diversion command is collected in real time. The difference between the actual transmission delay and the expected transmission delay is used as the delay deviation feedback quantity. The delay deviation feedback quantity is used to perform gradient iterative correction on the scheduling inertial compensation control, determine the collaborative sensing scheduling execution strategy, and send it to each IoT terminal.

2. The IoT terminal collaborative sensing and intelligent scheduling system based on a large model according to claim 1, characterized in that, Obtain multi-dimensional data traffic characteristics and business flow metadata of each terminal node within the IoT cluster. Transform the multi-dimensional data traffic characteristics into data flow polarity vectors using a feature mapping network of a large model, and construct a dynamic data attribute representation matrix, including: The system collects raw data packets sent from each terminal node within a continuous time window from each terminal node in the IoT cluster. Based on the transmission rate sequence, data packet length sequence, and time interval sequence of adjacent data packets in the raw data packets, it extracts multidimensional data traffic characteristics of each terminal node. At the same time, it extracts service flow metadata of each terminal node based on the service type code, source terminal identifier, destination edge scheduling center address, and data encapsulation format label in the raw data packets. The multidimensional data traffic features of each terminal node are input into the feature mapping network, and a nonlinear polarity transformation is performed on the transmission rate, packet length and time interval to map the multidimensional data traffic features into a data flow polarity vector with directional and magnitude attributes. Alignment is performed on the data flow polarity vector and the timestamp field carried in the business flow metadata, one time point at a time with the terminal node. The aligned timestamp is used as the row index of the matrix, and the terminal node number and business attribute are used as the column index of the matrix. The direction component and magnitude component of the aligned data flow polarity vector are filled into the corresponding cells of the matrix to construct a dynamic data attribute representation matrix.

3. The IoT terminal collaborative sensing and intelligent scheduling system based on a large model according to claim 2, characterized in that, Based on the dynamic data attribute representation matrix, the data link congestion degree and data encapsulation protocol matching degree between each node and the edge scheduling center are calculated, and a dynamic data flow coupling topology diagram simulating the potential field distribution is constructed, including: Based on the dynamic data attribute representation matrix, the magnitude component sequence and direction component deflection trajectory of each terminal node in a continuous time window are extracted, as well as the data encapsulation format label sequence corresponding to each terminal node. The cumulative amplitude of the modulus component sequence is compared with the preset throughput threshold of the edge scheduling center. The dispersion index of the directional component deflection trajectory is combined for weighted aggregation to calculate the data link congestion degree and data encapsulation protocol matching degree between each node and the edge scheduling center. Data link congestion and data encapsulation protocol matching degree are mapped to the potential field strength scalar of the topology vertex and the coupling weight coefficient of the edge between the vertex, respectively. Based on the potential field strength scalar, coupling weight coefficient and the identifier of each terminal node, the spatial topology relationship is initialized and the initial node link mapping network is constructed. In the initial node link mapping network, the potential field strength scalar is used as the reference radiation source. The virtual gravitational and repulsive field distribution is constructed along the decay direction of the coupling weight coefficient, and the resultant force direction and displacement step size of each vertex are calculated. The spatial positions of the vertices in the initial node link mapping network are iteratively rearranged according to the resultant force direction and displacement step size to obtain the dynamic data flow coupling topology.

4. The IoT terminal collaborative sensing and intelligent scheduling system based on a large model according to claim 3, characterized in that, Based on the dynamic data flow coupling topology diagram, the time-varying characteristics of data link congestion between topology nodes and the distribution of protocol matching degree are extracted to obtain the logical space state offset and service flow diffusion gradient, including: Based on the dynamic data flow coupled topology graph, a joint time series analysis is performed on the coupling weight coefficients between topology nodes and the potential field strength scalar of the corresponding vertices to extract the time-varying feature sequence of data link congestion. Based on the time-varying feature sequence, and combined with the data encapsulation format label sequence corresponding to each terminal node, logical space coordinate mapping is performed to obtain the protocol matching degree distribution density among topology nodes. Perform orthogonal feature coupling operation on the time-varying feature sequence of data link congestion degree and the protocol matching degree distribution density to obtain the logical space state offset; The logic space state offset is mapped by a gradient operator along the scalar decay direction of the potential field strength in the dynamic data flow coupling topology graph to obtain the business flow diffusion gradient.

5. The IoT terminal collaborative sensing and intelligent scheduling system based on a large model according to claim 4, characterized in that, Based on the logical space state offset and the service flow diffusion gradient, the gradient diffusion path and bandwidth allocation change rate vector of the service flow metadata in the dynamic data flow coupling topology are calculated to obtain the predicted scheduling trajectory sequence, including: Based on the logical space state offset and the business flow diffusion gradient, the logical space state offset is mapped to the starting reference point for topology optimization. The vector direction of the business flow diffusion gradient is used as the potential energy conduction guide. In the dynamic data flow coupling topology graph, the link connectivity of adjacent topology nodes is traversed step by step along the conduction guide. The decay sequence of link coupling weights is extracted and the cumulative convergence judgment is performed. The continuous links that fall into the preset convergence interval are combined into gradient diffusion paths. The physical topology spacing parameters of each link segment in the path are extracted in the order of node extension along the gradient diffusion path. The physical topology spacing parameters are perturbed and coupled with the potential field strength scalar of the corresponding node on the gradient diffusion path to obtain an intermediate adjustment sequence that characterizes the dynamic expansion and contraction of the link bandwidth resources. The intermediate adjustment sequence is subjected to time-series difference processing and vector normalization mapping to obtain the bandwidth allocation change rate vector. Extract the temporal evolution nodes of the dynamic scaling amplitude of each bandwidth resource in the bandwidth allocation change rate vector, and superimpose the temporal evolution nodes with the spatial topological coordinates of the gradient diffusion path for displacement compensation to construct a spatiotemporal coupled state matrix. Discretize and reassemble the spatiotemporal coupled state matrix according to a preset scheduling time window to obtain the predicted scheduling trajectory sequence.

6. The IoT terminal collaborative sensing and intelligent scheduling system based on a large model according to claim 5, characterized in that, Based on the predicted scheduling trajectory sequence, state transition threshold determination and scheduling inertia compensation control are performed for each trajectory point, dynamically triggering the diversion command of the target data transmission channel, including: Based on the predicted scheduling trajectory sequence, the link load status parameters corresponding to each trajectory point in the sequence are extracted point by point. The link load status parameters are compared with the preset state transition threshold on a cycle-by-cycle basis to identify the critical trajectory points where the load state change rate crosses the state transition threshold. The critical trajectory points are arranged in time sequence and combined into a target transition trajectory point set. Based on the set of target transition trajectory points, extract the state offset residual of each target transition trajectory point in the set between the current scheduling cycle and the previous scheduling cycle, perform momentum decay filtering and damping smoothing calculation on the state offset residual, and obtain the scheduling inertia compensation control quantity. Based on the scheduling inertia compensation control quantity, the initial data transmission channel associated with the target jump trajectory point is topologically offset corrected. Combined with the link connectivity weight in the dynamic data flow coupling topology graph, channel redirection mapping is performed to obtain the final channel matching identifier. Based on the final channel matching identifier, the target data transmission channel splitting command is dynamically triggered.

7. The IoT terminal collaborative sensing and intelligent scheduling system based on a large model according to claim 6, characterized in that, Based on the gradient diffusion path and the bandwidth allocation change rate vector, the expected transmission delay of each link is calculated, and the actual transmission delay of each node after executing the current traffic splitting command is collected in real time. The difference between the actual transmission delay and the expected transmission delay is used as the delay deviation feedback quantity, including: Based on the gradient diffusion path and the bandwidth allocation change rate vector, the physical topology span parameter of each link segment in the path is extracted. The bandwidth allocation change rate vector is mapped to the instantaneous effective throughput capacity threshold of the corresponding link segment. The data flow spatial propagation time and the link node queuing dwell time are superimposed and extrapolated by combining the physical topology span parameter and the instantaneous effective throughput capacity threshold to obtain the expected transmission delay benchmark sequence of each link. Using the time boundary of each expected delay node in the expected transmission delay reference sequence as the monitoring trigger window, the data stream inlet trigger identifier and outlet confirmation identifier of each IoT terminal after executing the current diversion command are obtained in real time. Differential accumulation operation is performed on the timestamp associated with the inlet trigger identifier and the outlet confirmation identifier to extract the actual transmission delay observation sequence of each node after executing the current diversion command. The actual delay fluctuation characteristics of each time node in the actual transmission delay observation sequence are extracted. The actual delay fluctuation characteristics are aligned with the expected transmission delay reference sequence through point-by-point topological mapping. The numerical difference calculation and positive and negative polarity calibration are performed to obtain the delay deviation feedback quantity that characterizes the degree of deviation of the current diversion command execution.

8. The IoT terminal collaborative sensing and intelligent scheduling system based on a large model according to claim 7, characterized in that, The scheduling inertial compensation control is iteratively corrected using the time delay deviation feedback to determine the cooperative sensing scheduling execution strategy and distribute it to each IoT terminal, including: The delay deviation feedback quantity is compared with the preset deviation convergence threshold cycle by cycle to identify abnormal link nodes whose delay deviation feedback quantity exceeds the deviation convergence threshold, and the historical correction trajectory of the scheduling inertial compensation control quantity corresponding to the abnormal link node is extracted. Based on the historical correction trajectory, a gradient descent direction search is performed on the scheduling inertial compensation control quantity. An adaptive step size adjustment factor is superimposed on the scheduling inertial compensation control quantity along the gradient descent direction to obtain the corrected scheduling inertial compensation control update quantity. The corrected scheduling inertia compensation control update quantity is mapped to the collaborative sensing scheduling execution strategy of each IoT terminal. The collaborative sensing scheduling execution strategy is encapsulated into a control command frame and broadcast to each IoT terminal node through the control channel of the edge scheduling center to complete collaborative sensing and intelligent scheduling.

9. The IoT terminal collaborative sensing and intelligent scheduling system based on a large model according to claim 8, characterized in that, The collaborative sensing scheduling execution strategy includes data acquisition frequency adjustment parameters for terminal nodes, data encapsulation format conversion instructions, and transmission channel priority remapping tables.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the system as described in any one of claims 1 to 9.

Images