A spatial internet of things device multipath routing control system and method

Abstract

The application discloses a kind of space Internet of Things equipment multi-path routing control system and method, comprising: collecting space and ground node link state data, pre-processing generates link state set;Extract Doppler shift and visible angle sequence, input improved SCINet generates link duration fragment;Based on the relationship between fragment connectivity and time overlap, build the relationship diagram and generate time window;To service data execution fragmentation processing, establish life trajectory record and generate fragmented set;Based on time window and fragment priority, perform path allocation and parallel transmission and replication;According to link state, perform path migration, complete data reorganization and output service result.The application predicts the availability of space link and builds cross-path connection, realizes the multi-path collaborative scheduling and millisecond-level seamless migration transmission of Internet of Things data fragmentation.

Description

Technical Field

[0001] This invention relates to the field of space Internet of Things (IoT) communication technology, and in particular to a multi-path routing control system and method for space IoT devices. Background Technology

[0002] With the development of integrated space-air-ground networks, space-based IoT devices are increasingly being applied to scenarios such as marine monitoring, emergency communication, remote sensing data transmission, and data collection in remote areas. These networks typically consist of space nodes, ground nodes, and IoT terminals, with data requiring forwarding via inter-satellite links, space-to-ground links, or space-to-ground links. Due to the high-speed movement of space nodes, the periodic changes in coverage, and the long propagation distances of the links, the link status fluctuates continuously over time.

[0003] Existing spatial IoT routing control methods typically select paths based on link latency, bandwidth, packet loss rate, or signal strength. Some methods employ link prediction or multi-path backup to improve transmission reliability. When the quality of the primary path deteriorates, the system switches to a backup path or distributes data across multiple available paths for transmission. These methods can alleviate single-link outages to some extent, but they largely rely on static link scoring or passive fault triggering.

[0004] However, space links are significantly affected by Doppler shift, changes in visibility angle, and Earth's obstruction. Existing methods struggle to accurately identify the sustainable transmission time periods of links and lack routing organization methods based on the temporal continuity between link segments. When a link experiences a momentary failure or a sudden performance degradation, data fragmentation is prone to waiting, repeated transmission, or switching delays, making it difficult to simultaneously guarantee service continuity, low latency, and low-power transmission.

[0005] Therefore, how to provide a multi-path routing control system and method for spatial IoT devices is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a multi-path routing control system and method for space IoT devices. This invention comprehensively utilizes space IoT communication, time series prediction, and intelligent multi-path routing control technologies. By constructing an improved SCINet network that incorporates Doppler-visible angle quadruple convolution and occlusion memory feedback gating mechanisms to predict persistent link segments, and combining cross-path connection graphs, connection time windows, data slice life trajectory recording structures, and hierarchical elastic transmission windows, it achieves segmented intelligent scheduling, continuous path migration, and seamless reassembly transmission of space IoT service data. It has the advantages of high link prediction accuracy, low path switching latency, strong service continuity, high transmission reliability, and high network resource utilization.

[0007] A multi-path routing control method for space IoT devices according to an embodiment of the present invention includes: Collect multi-source link status data between space nodes and ground nodes in the space Internet of Things network, preprocess the multi-source link status data, and generate a link status set; Based on the link state set, the corresponding instantaneous Doppler frequency shift sequence and visible angle sequence are extracted synchronously. A complex-quaternion coupled signal matrix is ​​constructed and input into the improved SCINet network, which introduces Doppler-visible angle quaternion convolution and occlusion memory feedback gating mechanism, to generate a set of link persistent segments. Based on the set of persistent link segments, a cross-path connection graph is constructed according to the spatial node connectivity, time interval overlap, and link communication capability constraints. A connection time window is generated based on the time overlap interval between persistent link segments. The IoT business data to be transmitted is segmented, a data segment life trajectory record structure is established for each data segment, and the data segments are prioritized according to the data segment life trajectory record structure to generate a data segment set. Based on the acceptance time window and data fragment priority, a hierarchical elastic transmission window is constructed on the cross-path acceptance relationship graph. Path allocation, parallel transmission and risk-weighted replication are performed on each data fragment, and the data fragment life trajectory record structure is updated synchronously. During data transmission, path migration is performed based on the cross-path connection diagram. When the continuous segment of the link corresponding to the current path ends or the link status is lower than the preset threshold, the incomplete data fragments are switched to the subsequent path, and the completed data fragments are reassembled and verified to generate complete business data output.

[0008] Optionally, the multi-source link status data includes link delay data, link bandwidth data, link packet loss rate data, link retransmission rate data, signal strength data, link connection status data, node location information, node orbit parameters, node attitude parameters, atmospheric environment parameters, and timestamp data.

[0009] Optionally, the preprocessing of multi-source link status data to generate a link status set includes time alignment, anomaly removal, and unified scale processing of the multi-source link status data to generate a link status set containing inter-satellite links, satellite-to-ground links, and air-to-ground links.

[0010] Optionally, the generated persistent link segment set includes: Read the node location, node speed, carrier frequency, signal strength, signal-to-noise ratio, link delay, link bandwidth, link packet loss rate, and sampling time corresponding to each candidate link from the link state set, and generate the link state time sequence according to the link number and sampling time; The instantaneous Doppler frequency shift sequence is extracted based on the relative motion relationship between the transmitting node and the receiving node, and the visible angle sequence is extracted based on the geometric visibility relationship between the transmitting node, the receiving node and the Earth's occlusion boundary. The instantaneous Doppler frequency shift sequence, the visible angle sequence and the link state timing are synchronized and aligned. Link delay, link bandwidth, link packet loss rate, signal strength and signal-to-noise ratio are mapped to complex link state channels, instantaneous Doppler frequency shift sequence is mapped to quaternion phase perturbation channel, and visible angle sequence is mapped to quaternion visible constraint channel to construct complex-quaternion coupled signal matrix; The complex-quaternion coupled signal matrix is ​​input into the improved SCINet network. The complex-quaternion coupled signal matrix is ​​sampled in layers. In each layer slice branch, the complex link state features, Doppler phase drift features and visible angle change features are extracted simultaneously by Doppler-visible angle quaternion convolution. Before the feature integration of adjacent layers, the link features of the shared space nodes are subjected to symmetric differential interaction and orbital phase offset compensation. After each slice output, an occlusion determination result is generated based on the visible angle sequence. The invalid feature channels of the occlusion time slice are gated and suppressed through the occlusion memory feedback gating mechanism. The quadruple convolution features of the most recent consecutive visible time slices are called for replacement compensation. The compensated features are input into the reconstruction module, and the delay prediction sequence, bandwidth prediction sequence, packet loss rate prediction sequence and connection duration prediction sequence of each candidate link are output. The time interval that continuously meets the preset stable communication conditions is marked as the link persistence segment, and a link persistence segment set is generated.

[0011] Optionally, the gating suppression of invalid feature channels in the occlusion time slice through the occlusion memory feedback gating mechanism includes: The output features of each slice are input into an improved SCINet network with an occlusion memory feedback gating mechanism, which includes a gating generation stage, a historical feature caching stage, and a feedback compensation stage. The gating generation stage receives the visible angle sequence and the output features of each slice. Based on the change state of the visible angle, it generates the gating weights of the corresponding time slices, enabling the improved SCINet network to distinguish between visible time slices, critical time slices, and occluded time slices according to the gating weights. It also performs attenuation control on the feature channels affected by occlusion in the occluded time slice. The historical feature caching stage receives the features output by the Doppler-visible angle quadruple convolution and writes the quadruple convolution features corresponding to the most recent consecutive visible time slices into the historical feature caching area according to the link number, slice level and sampling time, so that the improved SCINet network retains the effective link state expression before the occlusion occurs. During the feedback compensation phase, when an occlusion time slice is detected, the system reads the closest continuous visible quadruple convolutional features from the historical feature cache that belong to the same candidate link as the current occlusion time slice and are in the same time distance. It then performs replacement compensation on the invalid feature channels that are gated and attenuated, merges the replacement and compensated feature channels with the valid feature channels of the current time slice that are not affected by occlusion, and inputs the merged features into the reconstruction module.

[0012] Optionally, the construction of the cross-path connection graph, generating a connection time window based on the time overlap interval between continuous link segments, includes: Read the link number, start time, end time, sending node, receiving node, predicted latency, predicted bandwidth, predicted packet loss rate, and connection duration corresponding to each link persistence segment in the link persistence segment set, and build a link persistence segment index according to the link number and start time; According to the spatial node connectivity, the link persistent segments are matched for succession. Segment combinations in which the receiving node of the previous link persistent segment and the sending node of the next link persistent segment have a forwarding connection relationship are marked as candidate succession segment pairs. Overlap detection is performed on the start and end times of candidate segment pairs. The time interval in which the preceding link continuous segment and the following link continuous segment are both in a transmittable state is extracted, and candidate segment pairs that do not have time overlap intervals are removed. Based on the predicted latency, predicted bandwidth, predicted packet loss rate, and connection duration, candidate segment pairs with overlapping time intervals are screened for link communication capabilities. Segment pairs that meet the link communication capability constraints are identified as acceptable segment pairs, and the start and end points of the corresponding overlapping time intervals are written into the acceptance time window. Using the preceding and subsequent path segments as path nodes, and the transmission acceptance relationship between acceptable segment pairs as the acceptance connection relationship, the acceptance time window, spatial node identifier, and link communication capability constraints are written into the graph structure record to generate a cross-path acceptance relationship graph.

[0013] Optionally, the step of prioritizing data fragments based on the data fragment life trajectory recording structure and generating a data fragment set includes: Read the IoT business data to be transmitted, extract the business source, business type, data length, target receiving node, service quality requirements, generation time and expiration time, and establish a business data description record; The IoT business data is segmented based on the data length, service quality requirements, and expiration time in the business data description record. Each data segment is written with a segment number, the business number to which it belongs, the segment order, the segment length, the generation time, the expiration time, and the target receiving node. A data slice life trajectory record structure is established for each data slice, and the slice number, entry scheduling time, current path status, path migration record, cumulative waiting time, remaining valid time, amount of data transmitted, amount of data not transmitted, number of path switching and verification status are written into the data slice life trajectory record structure. Based on the data slice life trajectory record structure, read the remaining effective time, cumulative waiting time, amount of untransmitted data, number of path switching and service quality requirements of each data slice. According to the preset time remaining threshold, waiting time threshold, amount of untransmitted data threshold and number of switching threshold, mark the data slices with real-time protection priority, reliable protection priority or energy-saving transmission priority. The priority labeling results are written into the corresponding data slice life trajectory record structure, and all data slices are aggregated according to their business number, slice order, data slice priority and target receiving node to generate a data slice set.

[0014] Optionally, the construction of a hierarchical elastic transmission window, which performs path allocation, parallel transmission, and risk-weighted replication for each data shard, includes: Read the path nodes, connection relationships, connection time windows, predicted bandwidth, predicted latency, and predicted packet loss rate from the cross-path connection graph, and read the data fragment priority, remaining valid time, amount of untransmitted data, and target receiving node from the data fragment set; A path-level transmission window is constructed based on the acceptance time window. Path nodes that are in the same acceptance time window and meet the connectivity conditions of the target receiving node are written into the candidate path window. The transmission range that each candidate path can carry is determined based on the predicted bandwidth and connection duration. Based on data fragment priority, a fragment-level scheduling window is constructed. Data fragments with real-time guarantee priority are placed in the time-limit constraint window, data fragments with reliable guarantee priority are placed in the replication transmission window, and data fragments with energy-saving transmission priority are placed in the low-load transmission window. Map the fragment-level scheduling window to the path-level sending window. Perform path allocation based on the remaining valid time of the data fragment, the predicted path latency, the predicted path packet loss rate, and the transmission range that the path can carry. For data fragments with real-time guarantee priority, allocate paths that meet the expiration time limit. For data fragments with reliability guarantee priority, allocate the main path and the replication path. For data fragments with energy-saving transmission priority, allocate paths with sufficient remaining bandwidth. Parallel transmission and risk-weighted replication are performed on the data fragments according to the path allocation results, and the allocation path, replication path, sending time, transmission status, amount of data transmitted, amount of data not transmitted, and window adjustment record corresponding to each data fragment are written into the data fragment life trajectory record structure.

[0015] Optionally, generating complete business data output includes: During data transmission, the system reads the link persistence segment, take-up time window, link status, and data slice life trajectory record structure corresponding to the current transmission path, and continuously records the current path, transmission status, amount of data transmitted, and amount of data not transmitted for each data slice. Migration determination is performed based on the end time of the link segment corresponding to the current transmission path, the real-time link status, and the preset threshold. When the link segment ends or the link status is lower than the preset threshold, the current path is marked as a path to be migrated. Based on the cross-path connection diagram, read the connection relationship and connection time window corresponding to the path to be migrated, and determine the successor path from the path nodes that meet the constraints of node connectivity, time connection and link communication capability. The incomplete data fragments in the path to be migrated are switched to the subsequent path for continued transmission, and the switching time, original path identifier, subsequent path identifier, amount of untransmitted data and migration status are written into the corresponding data fragment life trajectory record structure. After the target receiving node receives the data fragments that have been transmitted, it reassembles and verifies the consistency of the data fragments according to their service number, fragment order, and verification status, and combines the data fragments that pass the verification into complete service data for output.

[0016] A multi-path routing control system for space IoT devices according to an embodiment of the present invention includes: The link acquisition module is used to collect and preprocess multi-source link status data between space nodes and ground nodes to generate a link status set. The segment prediction module is used to extract Doppler frequency shift and visible angle sequences based on the link state set, and generate a set of persistent link segments through an improved SCINet network. The connection graph module is used to construct a cross-path connection relationship graph based on the set of persistent links and generate connection time windows; The fragmentation and labeling module is used to fragment IoT business data, establish a data fragment life trajectory record structure and label priorities, and generate a data fragment set. The window scheduling module is used to construct a hierarchical elastic transmission window based on the acceptance time window and the fragmentation priority, and to perform path allocation, parallel transmission and risk-weighted replication. The migration and reorganization module is used to perform path migration based on the cross-path connection diagram, reorganize and verify the completed data fragments, and generate complete business data output.

[0017] The beneficial effects of this invention are: This invention improves the SCINet network by introducing Doppler-visible angle quadruple convolution and occlusion memory feedback gating mechanism to predict the state changes of spatial links under the influence of Doppler frequency shift, visible angle change and occlusion. It can more accurately identify continuous links, reduce the interference of occlusion time slices and link abrupt changes on the prediction results, and improve the accuracy of link availability judgment.

[0018] This invention constructs a cross-path connection relationship diagram and generates a connection time window based on the time overlap interval between continuous link segments. This allows path switching to no longer rely on temporary recalculation after link failure, but instead perform continuous migration based on the pre-formed connection relationship, thereby reducing path switching latency and improving the transmission continuity of spatial IoT service data in a dynamic link environment.

[0019] This invention establishes a data fragment life trajectory recording structure for data fragments and combines it with hierarchical elastic transmission window execution path allocation, parallel transmission, and risk-weighted replication to enable data fragments of different priorities to be dynamically scheduled according to remaining time limits, transmission status, and path risks, thereby improving data transmission reliability, timeliness, and network resource utilization. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a multi-path routing control method for space IoT devices proposed in this invention; Figure 2 This is a comparison diagram of link prediction results and actual link status for a multi-path routing control method for space IoT devices proposed in this invention. Figure 3 This is a schematic diagram of the structure of a multi-path routing control system for space Internet of Things (IoT) devices proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1 and Figure 2 A multi-path routing control method for space IoT devices, comprising: Collect multi-source link status data between space nodes and ground nodes in the space Internet of Things network, preprocess the multi-source link status data, and generate a link status set; Based on the link state set, the corresponding instantaneous Doppler frequency shift sequence and visible angle sequence are extracted synchronously. A complex-quaternion coupled signal matrix is ​​constructed and input into the improved SCINet network, which introduces Doppler-visible angle quaternion convolution and occlusion memory feedback gating mechanism, to generate a set of link persistent segments. Based on the set of persistent link segments, a cross-path connection graph is constructed according to the spatial node connectivity, time interval overlap, and link communication capability constraints. A connection time window is generated based on the time overlap interval between persistent link segments. The IoT business data to be transmitted is segmented, a data segment life trajectory record structure is established for each data segment, and the data segments are prioritized according to the data segment life trajectory record structure to generate a data segment set. Based on the acceptance time window and data fragment priority, a hierarchical elastic transmission window is constructed on the cross-path acceptance relationship graph. Path allocation, parallel transmission and risk-weighted replication are performed on each data fragment, and the data fragment life trajectory record structure is updated synchronously. During data transmission, path migration is performed based on the cross-path connection diagram. When the continuous segment of the link corresponding to the current path ends or the link status is lower than the preset threshold, the incomplete data fragments are switched to the subsequent path, and the completed data fragments are reassembled and verified to generate complete business data output.

[0023] In this embodiment, the multi-source link status data includes link delay data, link bandwidth data, link packet loss rate data, link retransmission rate data, signal strength data, link connection status data, node location information, node orbit parameters, node attitude parameters, atmospheric environment parameters, and timestamp data.

[0024] In this embodiment, the preprocessing of multi-source link status data to generate a link status set includes time alignment, anomaly removal, and unified scale processing of the multi-source link status data to generate a link status set containing inter-satellite links, satellite-to-ground links, and air-to-ground links.

[0025] In this embodiment, the generation of the persistent link fragment set includes: The node location, node speed, carrier frequency, signal strength, signal-to-noise ratio, link delay, link bandwidth, link packet loss rate, and sampling time are read from the link state set for each candidate link. A link state time series is generated according to the link number and sampling time. Specifically, the generation of the link state time series according to the link number and sampling time is as follows: Group all records in the link status set according to the link number, with each group corresponding to a candidate link; Within each set of records, the nodes are arranged in ascending order of sampling time. The node location, node speed, carrier frequency, signal strength, signal-to-noise ratio, link delay, link bandwidth, and link packet loss rate are aligned sequentially to the same time axis to form a single link state sequence. If the time interval between adjacent sampling exceeds the preset interpolation threshold, the missing time points are filled in by time interpolation, and the interpolation flag is marked in the interpolation record; If duplicate records exist at the same sampling time point, the record with the highest signal-to-noise ratio is selected as the valid record, and the remaining records are archived to the redundant data table. The individual link status sequences are merged in ascending order of link number to generate a link status time series sorted by both link number and sampling time. The instantaneous Doppler frequency shift sequence is extracted based on the relative motion relationship between the transmitting node and the receiving node, and the visible angle sequence is extracted based on the geometric visibility relationship between the transmitting node, the receiving node and the Earth's occlusion boundary. The instantaneous Doppler frequency shift sequence, the visible angle sequence and the link state timing are synchronized and aligned. Link delay, link bandwidth, link packet loss rate, signal strength, and signal-to-noise ratio are mapped to complex link state channels. Instantaneous Doppler frequency shift sequences are mapped to quaternion phase perturbation channels, and visible angle sequences are mapped to quaternion visible constraint channels. A complex-quaternion coupled signal matrix is ​​constructed, specifically as follows: According to the link number and sampling time, the link delay, link bandwidth, link packet loss rate, signal strength and signal-to-noise ratio are read, and the five parameters are normalized to the same dimension and written into the link quality sequence table in sequence. For each sampling moment in the link quality sequence table, the real and imaginary parts of the five parameters are assigned respectively according to the order of delay, bandwidth, packet loss rate, signal strength and signal-to-noise ratio. The five parameters are then combined in pairs to form a ten-dimensional complex link state channel, and the channel is kept in one-to-one correspondence with the original link number and sampling moment. Read the instantaneous Doppler frequency shift value corresponding to the link number and sampling time, perform amplitude normalization on the value, and write it into the first imaginary axis of the quaternion phase perturbation channel; Read the visible angle value corresponding to the link number and sampling time, perform amplitude normalization on the value and write it into the second imaginary axis of the quaternion visible constraint channel. At the same time, determine whether the visible angle is lower than the preset occlusion threshold and mark the occlusion status bit. The preset occlusion threshold is set to 5°. The multidimensional features are spliced ​​into a three-dimensional coupling tensor under a unified time axis and a unified link number dimension according to the combination order of the link quality complex channel, phase perturbation channel, and visible constraint channel, and written into the input buffer of the improved SCINet network to form a complex-quaternion coupling signal matrix. The complex-quaternion coupled signal matrix is ​​input into the improved SCINet network. Hierarchical slicing sampling is performed on the complex-quaternion coupled signal matrix. In each slice branch, complex link state features, Doppler phase drift features, and visible angle variation features are simultaneously extracted through Doppler-visible angle quaternion convolution. Before integrating features from adjacent layers, symmetric differential interaction and orbital phase offset compensation are performed on the link features of shared space nodes. The complex-quaternion coupled signal matrix is ​​subjected to hierarchical slicing sampling. In each slice branch, Doppler-visible angle quaternion convolution is used to simultaneously extract complex link state features, Doppler phase drift features, and visible angle variation features. Specifically: The complex-quaternion coupled signal matrix is ​​recursively sliced ​​according to the increasing time window length. The first layer uses the minimum sampling interval as the unit, and the slice length of each subsequent layer increases by 2 times while retaining half of the overlap between adjacent slices. Three parallel convolutional branches are constructed within each slice. The first branch reads the complex link state channel, the second branch reads the quaternion phase perturbation channel, and the third branch reads the quaternion visibility constraint channel. Each branch shares the convolutional kernel size and stride. During the kernel initialization phase, the Doppler frequency shift is mapped as a rotation coefficient and injected into the weights of the second branch convolution kernel, and the visible angle is mapped as an amplitude coefficient and injected into the weights of the third branch convolution kernel, so that the three branches synchronously modulate the convolution kernel. During the convolution calculation stage, the three branches extract link quality features, phase drift features, and visible angle change features respectively, and at the end of the slice, the three types of features are aligned and stitched according to the time dimension. Before integrating features from adjacent levels, symmetric differential interaction and orbital phase offset compensation are performed on the link features of shared space nodes. Specifically: After each layer of slices is output, multiple link features of the same spatial node as the sender or receiver are aggregated according to the link number to form a shared spatial node feature group. For any two link feature vectors within the feature group of each shared space node, perform element swapping to generate symmetric feature pairs, calculate the difference vector between the two link feature vectors, and write the difference vector into the symmetric difference interaction cache. Read the orbit phase information of the shared space node at the corresponding sampling time, perform phase offset compensation on the differential vector, and obtain the orbit phase compensation vector; The orbital phase compensation vector is fused with the original symmetric feature pair to form an interactive compensation feature; The interactive compensation features are written into the feature integration queue according to the link number and sampling time, providing joint features including symmetric differential interaction and orbital phase compensation for the feature integration of adjacent levels; After each slice output, an occlusion determination result is generated based on the visible angle sequence. An occlusion memory feedback gating mechanism is used to suppress invalid feature channels in occluded time slices. Quadruple convolution features from the most recent consecutive visible time slices are used for replacement compensation. The compensated features are input into the reconstruction module, which outputs the latency prediction sequence, bandwidth prediction sequence, packet loss rate prediction sequence, and connection duration prediction sequence for each candidate link. Time intervals that continuously meet preset stable communication conditions are marked as link duration segments, generating a link duration segment set. Specifically, the latency prediction sequence, bandwidth prediction sequence, packet loss rate prediction sequence, and connection duration prediction sequence for each candidate link are output as follows: The multi-channel feature tensor, after compensation by the occlusion memory feedback gating mechanism, is input into the first fusion unit of the reconstruction module, and is sequentially expanded according to the time dimension. The original amplitude information of each channel is maintained by residual connection. In the first fusion unit, multi-path separable convolutional branches are used to perform local temporal aggregation on the features after the sequence is expanded. The branches share the same time step identifier and are spliced ​​by channel after being aligned at the feature fusion end to obtain the fused feature sequence. The fused feature sequence is input into the temporal attention unit, and redundant information is filtered based on the importance weights of the features at each time step to generate a weighted feature sequence. The weighted feature sequence is input into the multi-task mapping unit and passes through four parallel linear projection channels to obtain the latency prediction vector, bandwidth prediction vector, packet loss rate prediction vector, and connection duration prediction vector of the candidate link, respectively. The four types of prediction vectors are aligned with the link number and sampling time, reorganized into corresponding delay prediction sequences, bandwidth prediction sequences, packet loss rate prediction sequences, and connection duration prediction sequences, and written into the prediction result buffer. The preset stable communication conditions are: Within the decision window, the instantaneous increment of the predicted latency does not exceed 10% of the previous time step, and the average predicted latency does not exceed 150% of the link's rated latency. The predicted bandwidth is continuously greater than 75% of the historical peak value of the link, maintaining a positive trend or with fluctuations of less than 5%; The predicted packet loss rate is consistently below 1%, and the increase in packet loss rate between two consecutive predictions does not exceed 0.2%. The remaining connection duration calculated based on phase perturbation and visibility angle must be no less than 1.5 times the maximum duration of the service segment; The visible angle at the corresponding sampling time must be more than twice the safety margin of the occlusion threshold of 5°, that is, the visible angle should not be less than 10°.

[0026] This invention structurally improves the SCINet network, proposing an improved SCINet network for spatial IoT link prediction scenarios. In terms of input modeling, link delay, link bandwidth, link packet loss rate, signal strength, and signal-to-noise ratio are used to form link state features; instantaneous Doppler frequency shift is used as a phase perturbation feature; and the visibility angle is used as a visibility constraint feature. A complex-quaternion coupled signal matrix is ​​constructed as the input network, enabling the network to simultaneously obtain link communication quality, spatial motion perturbation, and geometric visibility state before entering hierarchical slice sampling, effectively avoiding the predictive bias caused by solely relying on link quality parameters.

[0027] To address the link phase drift problem caused by high-speed motion of spatial nodes, this invention introduces a Doppler-visible angle quadruple convolution structure into the improved SCINet network. Complex link state features, Doppler phase drift features, and visible angle variation features are extracted simultaneously in each layer slice branch. By embedding Doppler frequency shift and visible angle variation into the quadruple convolution process, the network can learn the coupling relationship between link quality changes, orbital motion, and visible state in the same convolution operation, improving its ability to identify abrupt changes in link state, variations in connection duration, and the boundaries of link stability intervals.

[0028] To address the link state disruption caused by Earth occlusion or a rapid decrease in visibility angle, this invention incorporates an occlusion memory feedback gating mechanism into the improved SCINet network. This mechanism generates occlusion determination results based on the visibility angle sequence, gates and suppresses invalid feature channels in the occlusion time slice, and uses the quadruple convolutional features from the most recent consecutive visible time slices for replacement compensation. This allows the valid link state before occlusion to be fed back into the feature representation of the occlusion time slice. When reconstructing the prediction sequence, the network does not directly treat occlusion noise as valid link change learning, thus reducing the interference of the occlusion time slice on the prediction results for latency, bandwidth, packet loss rate, and connection duration.

[0029] To address the issue of state linkage when multiple candidate links share a space node, this invention introduces symmetrical differential interaction and orbital phase offset compensation before integrating features at adjacent levels. This involves differential comparison and phase correction of the features of multiple links sharing a space node, enabling the network to identify common disturbances and differential changes of the same space node on different links. This enhances the ability of the improved SCINet network to express the linkage relationships between inter-satellite links, satellite-to-ground links, and space-to-ground links, making the generated link continuous segments more consistent with the dynamic topology changes of actual space networks.

[0030] The improved SCINet network not only retains the original SCINet's ability to perform hierarchical slicing and multi-scale reconstruction of time series, but also incorporates Doppler phase perturbation, visible angle occlusion constraints, and shared node link linkage relationships into the prediction process. This makes the output delay prediction sequence, bandwidth prediction sequence, packet loss rate prediction sequence, and connection duration prediction sequence more stable and continuous. The link duration segments obtained by labeling based on the prediction results can provide more accurate timing information for cross-path connection relationship diagrams, connection time windows, and hierarchical elastic transmission windows, improving the continuity, reliability, and switching timeliness of multi-path routing control for space IoT devices.

[0031] In this embodiment, the gating suppression of invalid feature channels in the occlusion time slice through the occlusion memory feedback gating mechanism includes: The output features of each slice are input into an improved SCINet network with an occlusion memory feedback gating mechanism, which includes a gating generation stage, a historical feature caching stage, and a feedback compensation stage. The gating generation stage receives the visible angle sequence and the output features of each slice. Based on the change state of the visible angle, it generates the gating weights of the corresponding time slices, enabling the improved SCINet network to distinguish between visible time slices, critical time slices, and occluded time slices according to the gating weights. It also performs attenuation control on the feature channels affected by occlusion in the occluded time slice. The historical feature caching stage receives the features output by the Doppler-visible angle quadruple convolution and writes the quadruple convolution features corresponding to the most recent consecutive visible time slices into the historical feature caching area according to the link number, slice level and sampling time, so that the improved SCINet network retains the effective link state expression before the occlusion occurs. During the feedback compensation phase, when an occlusion time slice is detected, the system reads the closest continuous visible quadruple convolutional features from the historical feature cache that belong to the same candidate link as the current occlusion time slice and are in the same time distance. It then performs replacement compensation on the invalid feature channels that are gated and attenuated, merges the replacement and compensated feature channels with the valid feature channels of the current time slice that are not affected by occlusion, and inputs the merged features into the reconstruction module.

[0032] In this embodiment, the construction of the cross-path connection graph, which generates a connection time window based on the time overlap interval between continuous link segments, includes: Read the link number, start time, end time, sending node, receiving node, predicted latency, predicted bandwidth, predicted packet loss rate, and connection duration corresponding to each link persistence segment in the link persistence segment set, and build a link persistence segment index according to the link number and start time; Link persistence segments are matched for succession based on spatial node connectivity. Segment pairs where the receiving node of the preceding link persistence segment and the sending node of the following link persistence segment have a forwarding connection are marked as candidate successor segment pairs. Specifically, the matching of link persistence segments based on spatial node connectivity is as follows: For each link persistent segment, read the sending node number and receiving node number, and generate sending-receiving key-value pairs in the link persistent segment index table; Retrieve the forwarding connectivity between the sending node and the receiving node in the node topology table. If there is a direct inter-satellite connection or a valid route via a gateway from the sending node to the receiving node, mark the connectivity and forwarding status as key-value pairs in the link persistent fragment index table. According to the sampling time order, connectivity matching is performed on any two adjacent continuous links. If the receiving node of the current segment and the sending node of the next segment are both marked as connected and forwardable and the time intervals of the two segments overlap, then these two segments are combined into a candidate receiving segment pair. Write all segments that meet the connectivity and forwarding conditions and whose time intervals overlap into the candidate segment pair list, and record the preceding path identifier, the succeeding path identifier, and the start and end times of the overlap for each pair to complete the matching of continuous links. Overlap detection is performed on the start and end times of candidate segment pairs. The time interval in which the preceding link continuous segment and the following link continuous segment are both in a transmittable state is extracted, and candidate segment pairs that do not have time overlap intervals are removed. Based on predicted latency, predicted bandwidth, predicted packet loss rate, and connection duration, candidate segment pairs with overlapping time intervals are screened for link communication capabilities. Segment pairs that meet the link communication capability constraints are identified as acceptable segment pairs, and the start and end points of the corresponding overlapping time intervals are written into the acceptance time window, where: The link communication capability constraints are: Latency constraints: The predicted end-to-end latency of a candidate successor segment shall not exceed 1.5 times the predicted latency of the preceding segment, and shall not exceed the maximum acceptable latency declared by the service. Bandwidth constraints: The predicted available bandwidth of the candidate successor segment must be no less than 120% of the remaining rate requirement of the current data fragment to be migrated, and at the same time no less than 80% of the predicted bandwidth of the preceding segment, in order to ensure stable throughput after the switchover. Packet loss rate constraint: The predicted packet loss rate of the candidate successor segment must be less than 1% and must not be higher than the predicted packet loss rate of the preceding segment; Continuous duration constraint: The remaining connection duration of the candidate successor segment must be no less than 1.5 times the estimated time required for the data fragment to be migrated to complete the transmission, and greater than the remaining duration of the preceding segment, to ensure that the data will not be migrated again immediately after the path switch. Using pre-order and subsequent path segments as path nodes, and the transmission acceptance relationship between acceptable segment pairs as the acceptance connection relationship, the acceptance time window, spatial node identifier, and link communication capability constraints are written into the graph structure record to generate a cross-path acceptance relationship graph. Specifically, the generation of the cross-path acceptance relationship graph is as follows: Read the preceding segment number, succeeding segment number, start and end times of the acceptance time window, start node number and end node number of each acceptable segment pair, establish path node pairs for the segment pairs in the order of preceding and succeeding numbers, and assign a send-receive binary link label to each node. The preceding segment number and the succeeding segment number are concatenated into a 1×2 path index vector, which is then written into the path index matrix in chronological order to ensure that the row numbers of the matrix correspond one-to-one with the candidate successor segment pairs. For two adjacent rows in the path index matrix, calculate the difference flag between the send-receive tags of the preceding node and the succeeding node. If the difference flag of the two rows is zero and the two rows have overlapping time windows, record the starting point and ending point of the overlap and generate the acceptance window vector. Using the start and end points of the receiving window vector as edge attributes, directed edges are generated for the corresponding row numbers of the path index matrix; at the same time, four link communication capability indicators are read. If the latency, bandwidth, packet loss rate and connection duration all meet the link communication capability constraints, then a receiving identifier is written for the directed edge. Write the start and end nodes of the directed edges that have been marked as acceptable into the graph node set, and write the accepting window vector and link communication capability constraints of the directed edges into the graph edge set; generate an adjacency list structure according to the node set and edge set to form a cross-path accepting relationship graph.

[0033] In this embodiment, the step of prioritizing data fragments based on the data fragment life trajectory recording structure and generating a data fragment set includes: Read the IoT business data to be transmitted, extract the business source, business type, data length, target receiving node, service quality requirements, generation time and expiration time, and establish a business data description record; The IoT business data is segmented based on the data length, service quality requirements, and expiration time in the business data description record. Each data segment is written with a segment number, the business number to which it belongs, the segment order, the segment length, the generation time, the expiration time, and the target receiving node. A data slice lifecycle record structure is established for each data slice. The slice number, entry scheduling time, current path status, path migration record, cumulative waiting time, remaining valid time, amount of data transmitted, amount of data not transmitted, number of path switches, and verification status are written into the data slice lifecycle record structure. Specifically, establishing a data slice lifecycle record structure for each data slice involves: Request a unique 64-bit fragment number for the data to be transmitted, write the fragment number into the fragment number field, and write the current system time into the entry scheduling time field; Initialize the path migration record as an empty linked list, write the initial allocated path number into the current path status field, and simultaneously append a record to the path migration record linked list. The record content includes timestamp, initial path, and 0 bytes transferred. Set the cumulative waiting time field to zero, set the amount of data transmitted to zero, write the amount of data not transmitted to the fragment length field, set the path switching count field to zero, and set the verification status field to unverified. Calculate the difference between the expiration time of the fragment and the time it enters the scheduling process, and write the result into the remaining valid time field; if the difference is negative, mark the fragment as an expired fragment and write it into the scheduling exception queue. The ten fields are written into the shard-index hash table in a fixed order and registered in the scheduling memory mapping area in the form of shard number and structure pointer, thus completing the establishment of the data shard life trajectory record structure; Based on the data slice life trajectory record structure, the remaining effective time, cumulative waiting time, amount of untransmitted data, number of path switching times and service quality requirements of each data slice are read. According to the preset time reserve threshold, waiting time threshold, untransmitted data threshold and switching time threshold, the data slices are marked with real-time guarantee priority, reliable guarantee priority or energy-saving transmission priority. The preset time reserve threshold is 15s, the waiting time threshold is 8s, the untransmitted data threshold is 32kB, and the switching time threshold is 2 times. Data fragments are labeled with real-time guarantee priority, reliability guarantee priority, or energy-saving transmission priority, specifically as follows: Read four real-time indicators from the data slice life trajectory record structure: remaining valid time, cumulative waiting time, amount of untransmitted data, and number of path switching times. Align the fields with the expiration time limit, reliability level, and energy saving level corresponding to the service quality requirements to generate a slice real-time evaluation table. Iterate through the shard real-time evaluation table. When the remaining valid time is less than or equal to 15 seconds, or the cumulative waiting time is greater than or equal to 8 seconds, write the data shard into the real-time protection queue and set the priority flag field in the record structure to real-time protection. For fragments that have not entered the real-time assurance queue, if the amount of untransmitted data is greater than or equal to 32kB, or the number of path switching times is greater than or equal to 2, and the reliability level of the service quality requirement is higher than the energy-saving level, then the fragment will be written into the reliable assurance queue and the priority flag field will be set to reliable assurance. For fragments that have not entered the real-time assurance queue and the reliable assurance queue, they are directly written to the energy-saving transmission queue, and the priority flag field is set to energy-saving transmission. Following the order of real-time support queue, reliable support queue, and energy-saving transmission queue, the three types of queues, along with their corresponding fragment numbers and priority flags, are synchronously written into the scheduling mapping table to complete the data fragment priority labeling. The priority labeling results are written into the corresponding data slice life trajectory record structure, and all data slices are aggregated according to their business number, slice order, data slice priority and target receiving node to generate a data slice set.

[0034] In this embodiment, the construction of a hierarchical elastic transmission window, which performs path allocation, parallel transmission, and risk-weighted replication for each data shard, includes: Read the path nodes, connection relationships, connection time windows, predicted bandwidth, predicted latency, and predicted packet loss rate from the cross-path connection graph, and read the data fragment priority, remaining valid time, amount of untransmitted data, and target receiving node from the data fragment set; A path-level transmission window is constructed based on the acceptance time window. Path nodes within the same acceptance time window that meet the connectivity conditions of the target receiving node are written into the candidate path window. The transmission range that each candidate path can carry is determined based on the predicted bandwidth and connection duration. Specifically, the path-level transmission window is constructed based on the acceptance time window as follows: Read the start and end times of the connection time windows of all edges in the cross-path connection graph, and merge them according to the time sequence to obtain a non-overlapping global time window sequence. Retrieve the window segment into which the current system time falls from the global time window sequence, record the window segment as the activity takeover time window, and read the start time, end time, and associated preceding and succeeding path nodes; For each preceding and succeeding path associated with the activity takeover time window, node connectivity verification is performed. If there is a direct inter-satellite connection, ground gateway forwarding, or air-to-ground backhaul between the receiving and sending nodes of the two paths, the target receiving node connectivity condition is deemed met. For paths that have passed connectivity verification, read their predicted available bandwidth and remaining connection duration, calculate the available transmission duration of the path within the active carrying time window, and multiply the predicted available bandwidth by the available transmission duration to obtain the transmission range that the path can carry. Write the path number, start and end times of the acceptance time window, and the transmission range that meet the connectivity conditions and have completed the calculation of the transmission range into the candidate path window to complete the construction of the path-level sending window. The transmission range that each candidate path can carry is determined based on the predicted bandwidth and connection duration, specifically as follows: Read the predicted available bandwidth value of the candidate path; Read the remaining connection duration of the candidate path and take the smaller value between it and the remaining duration of the active takeover time window as the available transmission time; Multiply the predicted available bandwidth value by the available transmission time, convert the result to megabytes, and obtain the theoretical data transmission capacity of the path; Based on the preset protocol overhead factor, which is 0.92 by default, the theoretically transmittable data volume is calculated, and signaling and retransmission overhead is eliminated to obtain the net effective transmission capacity. Write the net effective transmission capacity into the transmittable range field of the candidate path window, and archive it together with the path number and the start and end times of the acceptance time window; A shard-level scheduling window is constructed based on data shard priority. Data shards with real-time guarantee priority are placed in the time-constraint window, data shards with reliability guarantee priority are placed in the replication transmission window, and data shards with energy-saving transmission priority are placed in the low-load transmission window. Specifically, the shard-level scheduling window based on data shard priority is constructed as follows: Read the fragment number, remaining valid time, amount of untransmitted data and target receiving node of the three priority queues: real-time guarantee, reliable guarantee and energy-saving transmission. Merge the three queues into the fragment scheduling temporary storage table and add a priority flag field to each record in the table. Based on the priority flag, allocate time limit constraint windows to the real-time support queue, calculate the difference between the current system time and the remaining valid time of each segment record to obtain the available scheduling margin, place the segment with a scheduling margin of less than 5 seconds at the top of the time limit constraint window, and write the remaining real-time support segments into the window in ascending order of remaining valid time. Perform a replication requirement assessment on the reliable guarantee queue, read the transmission range that can be carried by all available paths in the cross-path connection graph, if the spare capacity of the transmission range is greater than 1.2 times the amount of untransmitted data in the fragment, add a replication flag to the fragment, write the main path and replication path number into the replication path list, and write the fragment into the replication transmission window. The energy-saving transmission queue is grouped into segments according to the target receiving node. The ratio of the current path load to the transmission range that can be carried is read. If the path load is less than 60% of the transmission range that can be carried, the segment is written into the low-load transmission window of the corresponding path. If the path load is greater than 60%, the segment is written into the low-load transmission window of the backup path. In the three types of window structures generated by the time-constrained window, the replication transmission window, and the low-load transmission window, a window sequence number, a scheduling start time, and an expected completion time field are added to each shard record. The window structure is then written into the scheduling mapping table to complete the construction of the shard-level scheduling window. Map the fragment-level scheduling window to the path-level sending window. Perform path allocation based on the remaining valid time of the data fragment, the predicted path latency, the predicted path packet loss rate, and the transmission range that the path can carry. For data fragments with real-time guarantee priority, allocate paths that meet the expiration time limit. For data fragments with reliability guarantee priority, allocate the main path and the replication path. For data fragments with energy-saving transmission priority, allocate paths with sufficient remaining bandwidth. Parallel transmission and risk-weighted replication are performed on data fragments according to the path allocation results. The allocation path, replication path, transmission time, transmission status, amount of data transmitted, amount of data not transmitted, and window adjustment record for each data fragment are written into the data fragment lifecycle record structure. Specifically, the parallel transmission and risk-weighted replication are performed on the data fragments according to the path allocation results as follows: For each data fragment, obtain the main path number, available replication path number, transmittable range and remaining data volume of the fragment, sort the replication paths from low to high according to the predicted packet loss rate and write them into the replication path priority queue in sequence. For each data fragment, calculate the difference between the remaining transmission range that the main path can carry and the remaining data volume of the fragment. If the difference is positive and the replication path priority queue is not empty, start the replication scheduling thread. The thread first selects the first replication path at the top of the replication path priority queue, reads the predicted packet loss rate of the replication path and calculates its reciprocal, and uses the reciprocal value as the replication weight. Based on the replication weight and the remaining transmission range that the main path can carry, the remaining data volume of the fragment is risk-weighted and split. The remaining data volume of the fragment is multiplied by the replication weight to obtain the data volume of the replication segment. The data volume of the replication segment is sent to the replication path, and the remaining transmission range that the replication path can carry is updated synchronously. Simultaneously, the remaining data volume of the fragment minus the data volume of the replicated segment is written into the main path sending buffer and sent asynchronously and in parallel with the replicated path. Both the main path and the replicated path enable the ACK-NACK feedback mechanism, compare and confirm according to the data block sequence number, and send back the transmission status. After receiving a complete ACK receipt on the main path or the replication path, the real-time update values ​​of the transmitted data volume and untransmitted data volume fields are immediately written into the corresponding data slice life trajectory record structure. The transmission status field is marked as completed or in progress, and the sending time, completion time, main path number, replication path number, and window adjustment number are recorded. If both the main path and the replication path have completed transmission, the window scheduling module is called to release the transmission window resources that have been occupied, thus completing the parallel transmission and risk-weighted replication process.

[0035] In this embodiment, generating complete business data output includes: During data transmission, the system reads the link persistence segment, take-up time window, link status, and data slice life trajectory record structure corresponding to the current transmission path, and continuously records the current path, transmission status, amount of data transmitted, and amount of data not transmitted for each data slice. Migration determination is performed based on the end time of the link segment corresponding to the current transmission path, the real-time link status, and the preset threshold. When the link segment ends or the link status is lower than the preset threshold, the current path is marked as a path to be migrated. The preset threshold is when the remaining duration of the link segment does not exceed 2 seconds, or the average real-time round-trip delay reaches 800ms, or the predicted available bandwidth drops to less than 60% of the historical peak, or the average packet loss rate of the last three samples rises to more than 2%, or the real-time signal-to-noise ratio drops to less than 8dB. If any of these conditions are met, the current path is considered to have insufficient communication capacity and is marked as a path to be migrated. Based on the cross-path connection diagram, read the connection relationship and connection time window corresponding to the path to be migrated, and determine the successor path from the path nodes that meet the constraints of node connectivity, time connection and link communication capability. The incomplete data fragments in the path to be migrated are switched to the subsequent path for continued transmission, and the switching time, original path identifier, subsequent path identifier, amount of untransmitted data and migration status are written into the corresponding data fragment life trajectory record structure. After the target receiving node receives the data fragments that have been transmitted, it reassembles and verifies the consistency of the data fragments according to their service number, fragment order, and verification status, and combines the data fragments that pass the verification into complete service data for output.

[0036] refer to Figure 3 A multi-path routing control system for space IoT devices, comprising: The link acquisition module is used to collect and preprocess multi-source link status data between space nodes and ground nodes to generate a link status set. The segment prediction module is used to extract Doppler frequency shift and visible angle sequences based on the link state set, and generate a set of persistent link segments through an improved SCINet network. The connection graph module is used to construct a cross-path connection relationship graph based on the set of persistent links and generate connection time windows; The fragmentation and labeling module is used to fragment IoT business data, establish a data fragment life trajectory record structure and label priorities, and generate a data fragment set. The window scheduling module is used to construct a hierarchical elastic transmission window based on the acceptance time window and the fragmentation priority, and to perform path allocation, parallel transmission and risk-weighted replication. The migration and reorganization module is used to perform path migration based on the cross-path connection diagram, reorganize and verify the completed data fragments, and generate complete business data output.

[0037] Example 1: In a continuous space IoT transmission cycle, the system connects 48 space nodes, 6 ground nodes, and 320 IoT terminals. The terminals upload status data, environmental data, and alarm data within an observation area without fixed communication infrastructure coverage. The system collected 96,000 status records for inter-satellite links, satellite-to-ground links, and air-to-ground links, with a sampling interval of 1 second. Each record includes link number, node location, node speed, carrier frequency, signal strength, signal-to-noise ratio, link delay, link bandwidth, link packet loss rate, and sampling time. The raw data contained 17.8% of link fluctuation segments, including 6,420 records with obstruction edges at visibility angles below 5°, 5,180 records with signal-to-noise ratios below 8dB, and 3,960 records with packet loss rates exceeding 2%, indicating that the links in this scenario do indeed experience second-level sudden drops and short-term interruptions.

[0038] After the data enters the processing flow, the system first groups the 96,000 link status records according to the link number, forming 126 candidate link record groups. Within each group, the records are arranged in ascending order of sampling time, aligning node location, node speed, carrier frequency, signal strength, signal-to-noise ratio (SNR), link delay, link bandwidth, and link packet loss rate to the same time axis. For records with adjacent sampling intervals exceeding 2 seconds, the system uses linear interpolation to fill in missing points, completing 1384 sampling points in this round. For records reported repeatedly at the same sampling point, the system retains the record with the highest SNR, removing 692 duplicate records. After preprocessing, the effective records in the link status set are 96,692, the delay data range converges from the original 78–1280ms to 82–1175ms, and the proportion of abnormal spikes decreases from 4.6% to 0.9%.

[0039] During the training phase, the system used the first 72,000 records as training samples, 12,000 records as validation samples, and 12,692 records as test samples. In the training samples, one inter-satellite link had signal-to-noise ratios (SNRs) of 18.5dB, 18.1dB, 17.8dB, 17.2dB, and 16.9dB at five consecutive sampling points, corresponding to delays of 132ms, 136ms, 139ms, 145ms, and 151ms, bandwidths of 4.8Mbps, 4.7Mbps, 4.5Mbps, 4.4Mbps, and 4.2Mbps, and visibility angles of 36°, 34°, 32°, 30°, and 28°. This segment was marked as a stable segment. Another satellite-to-ground link experienced an SNR decrease from 11.2dB to 7.6dB at five consecutive sampling points, a visibility angle decrease from 12° to 4°, and a packet loss rate increase from 0.8% to 2.7%. This segment was marked as a critical occlusion segment.

[0040] The system synchronously extracts instantaneous Doppler frequency shift and visible angle sequences from the link state set. It writes link delay, link bandwidth, link packet loss rate, signal strength, and signal-to-noise ratio into the link state feature channel, Doppler frequency shift into the phase perturbation feature channel, and visible angle into the visible constraint feature channel, forming a complex-quaternion coupled signal matrix. Taking link L36 as an example, its delay at 38s is 214ms, bandwidth is 3.6Mbps, packet loss rate is 0.62%, signal strength is -86dBm, signal-to-noise ratio is 14.2dB, Doppler frequency shift is 5.7kHz, and visible angle is 22°. The system writes this set of data into the same sampling row. When the visible angle of the same link drops to 4.6° at 52s, the system synchronously writes the occlusion status bit for use by the occlusion memory feedback gating mechanism.

[0041] After improving the SCINet network by inputting a complex-quaternion coupled signal matrix, the system performs slice sampling at three levels: 4s, 8s, and 16s, with each adjacent slice retaining 50% overlap. The first layer of slices captures short-term signal-to-noise ratio (SNR) spikes, the second layer captures bandwidth slippage trends, and the third layer captures connection duration changes. In the Doppler-visible angle quaternion convolution processing, the network simultaneously extracts link state features, Doppler phase drift features, and visible angle change features. Taking the occlusion segment from 52s to 59s as an example, the original SCINet directly uses low SNR and high packet loss rate as continuous input, causing the connection duration prediction to jump from 19s to 3s. The occlusion memory feedback gating mechanism of this invention suppresses invalid feature channels when the visible angle is below 5° and calls the quaternion convolution features from the most recent continuous visible time slice from 49s to 51s for replacement compensation. After compensation, the connection duration prediction smoothly decreases from 18s to 9s without any precipitous misjudgments.

[0042] After testing, the improved SCINet network achieved an average prediction error of 29ms for latency, 0.22Mbps for bandwidth, 0.17% for packet loss rate, and 2.6s for connection duration on 12,692 test samples. The corresponding errors for the traditional SCINet were 55ms, 0.43Mbps, 0.34%, and 6.1s, respectively. Regarding link persistence segment identification, this invention identified 418 stable communication segments, of which 394 were manually verified as correct, achieving an accuracy rate of 94.3%. The traditional method identified 376 stable communication segments, with 309 being correct, achieving an accuracy rate of 82.2%.

[0043] In the business transmission simulation, 320 IoT terminals generated a total of 20.48MB of business data. The system divided this data into 320 data fragments of 64kB each, including 46 alarm fragments, 174 status fragments, and 100 periodic monitoring fragments. After each fragment entered scheduling, a data fragment lifecycle record structure was established, recording the fragment number, scheduling entry time, current path status, path migration record, cumulative waiting time, remaining valid time, amount of data transmitted, amount of data not transmitted, number of path switching attempts, and verification status. The system set the time margin threshold to 15s, the waiting time threshold to 8s, the amount of data not transmitted to 32kB, and the number of switching attempts to 2. After labeling, there were 52 real-time priority fragments, 131 reliable priority fragments, and 137 energy-saving transmission priority fragments.

[0044] At 37s, the visibility angle of the primary satellite-to-ground link L36 decreased from 13° to 8°, and the bandwidth decreased from 3.2Mbps to 1.7Mbps. The system predicted that this persistent link segment would end at 41s. After reading the persistent link segment set, the system found that the subsequent paths L36-L82 and L82-L11 overlapped in time from 38s to 44s. The predicted bandwidth of the subsequent path was 3.9Mbps, the predicted packet loss rate was 0.48%, and the connection duration was 58s, which met the link communication capacity constraints. Therefore, the period from 38s to 44s was written into the acceptance time window, and a cross-path acceptance relationship diagram was generated. The traditional method only re-routes after detecting a packet loss rate exceeding 2% at 42s, during which time 27 fragments had already entered a waiting state.

[0045] Within the hierarchical elastic transmission window, real-time guaranteed fragments enter the time-constrained window with an average waiting time of 1.1 seconds; reliable guaranteed fragments enter the replication transmission window, with 63 fragments allocated to the primary path and replication paths; energy-saving transmission fragments enter the low-load transmission window, utilizing paths with loads below 60% for transmission. For the alarm fragment P118, the system reads its remaining valid time as 12 seconds and untransmitted data as 64kB, assigns it to the subsequent paths L82-L11, and simultaneously writes the path migration record into the lifecycle. For the reliable fragment P206, the system sends a 48kB primary segment to the primary path and a 16kB replication segment to the replication path with a lower predicted packet loss rate. Ultimately, the primary path completes acknowledgment in 0.72 seconds, and the replication path completes acknowledgment in 0.91 seconds, at which point the system releases replication window resources.

[0046] When the signal-to-noise ratio (SNR) of another air-to-ground link drops to 7.4 dB and the visibility angle drops to 3.8° at the 56th second, the system marks this path as a path to be migrated based on the migration threshold, because the real-time SNR has fallen below 8 dB and the packet loss rate has reached 2.3%. This invention completes the handover of 36 incomplete fragments within 0.19 seconds, while the traditional method has an average handover time of 1.31 seconds. After transmission, the target receiving node reassembles 320 data fragments according to their service number and fragmentation order. This invention successfully reassembles 318 fragments, achieving a reassembly success rate of 99.4%; the traditional method successfully reassembles 303 fragments, achieving a reassembly success rate of 94.7%.

[0047] In the same batch of test services, the results of comparing the method of this invention with the traditional single-path plus static backup method are as follows: the average end-to-end latency is 1.38s and 2.36s, respectively; the average arrival latency of alarm fragments is 0.82s and 1.69s, respectively; the on-time delivery rate of fragments is 98.7% and 90.9%, respectively; the average packet loss rate is 0.41% and 1.81%, respectively; the amount of retransmitted data is 0.58MB and 2.64MB, respectively; the cumulative waiting time caused by link interruption is 3.6s and 29.4s, respectively; and the energy consumption per unit of data transmission is 0.76 times and 1.00 times that of the traditional method, respectively.

[0048] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-path routing control method for space IoT devices, characterized in that, include: Collect multi-source link status data between space nodes and ground nodes in the space Internet of Things network, preprocess the multi-source link status data, and generate a link status set; Based on the link state set, the corresponding instantaneous Doppler frequency shift sequence and visible angle sequence are extracted synchronously. A complex-quaternion coupled signal matrix is ​​constructed and input into the improved SCINet network, which introduces Doppler-visible angle quaternion convolution and occlusion memory feedback gating mechanism, to generate a set of link persistent segments. Based on the set of persistent link segments, a cross-path connection graph is constructed according to the spatial node connectivity, time interval overlap, and link communication capability constraints. A connection time window is generated based on the time overlap interval between persistent link segments. The IoT business data to be transmitted is segmented, a data segment life trajectory record structure is established for each data segment, and the data segments are prioritized according to the data segment life trajectory record structure to generate a data segment set. Based on the acceptance time window and data fragment priority, a hierarchical elastic transmission window is constructed on the cross-path acceptance relationship graph. Path allocation, parallel transmission and risk-weighted replication are performed on each data fragment, and the data fragment life trajectory record structure is updated synchronously. During data transmission, path migration is performed based on the cross-path connection diagram. When the continuous segment of the link corresponding to the current path ends or the link status is lower than the preset threshold, the incomplete data fragments are switched to the subsequent path, and the completed data fragments are reassembled and verified to generate complete business data output.

2. The multi-path routing control method for a space IoT device according to claim 1, characterized in that, The multi-source link status data includes link delay data, link bandwidth data, link packet loss rate data, link retransmission rate data, signal strength data, link connection status data, node location information, node orbit parameters, node attitude parameters, atmospheric environment parameters, and timestamp data.

3. The multi-path routing control method for a space IoT device according to claim 1, characterized in that, The preprocessing of multi-source link status data to generate a link status set includes time alignment, anomaly removal, and unified scale processing of the multi-source link status data, generating a link status set containing inter-satellite links, satellite-to-ground links, and air-to-ground links.

4. The multi-path routing control method for space IoT devices according to claim 1, characterized in that, The generated persistent link fragment set includes: Read the node location, node speed, carrier frequency, signal strength, signal-to-noise ratio, link delay, link bandwidth, link packet loss rate, and sampling time corresponding to each candidate link from the link state set, and generate the link state time sequence according to the link number and sampling time; The instantaneous Doppler frequency shift sequence is extracted based on the relative motion relationship between the transmitting node and the receiving node, and the visible angle sequence is extracted based on the geometric visibility relationship between the transmitting node, the receiving node and the Earth's occlusion boundary. The instantaneous Doppler frequency shift sequence, the visible angle sequence and the link state timing are synchronized and aligned. Link delay, link bandwidth, link packet loss rate, signal strength and signal-to-noise ratio are mapped to complex link state channels, instantaneous Doppler frequency shift sequence is mapped to quaternion phase perturbation channel, and visible angle sequence is mapped to quaternion visible constraint channel to construct complex-quaternion coupled signal matrix; The complex-quaternion coupled signal matrix is ​​input into the improved SCINet network. The complex-quaternion coupled signal matrix is ​​sampled in layers. In each layer slice branch, the complex link state features, Doppler phase drift features and visible angle change features are extracted simultaneously by Doppler-visible angle quaternion convolution. Before the feature integration of adjacent layers, the link features of the shared space nodes are subjected to symmetric differential interaction and orbital phase offset compensation. After each slice output, an occlusion determination result is generated based on the visible angle sequence. The invalid feature channels of the occlusion time slice are gated and suppressed through the occlusion memory feedback gating mechanism. The quadruple convolution features of the most recent consecutive visible time slices are called for replacement compensation. The compensated features are input into the reconstruction module, and the delay prediction sequence, bandwidth prediction sequence, packet loss rate prediction sequence and connection duration prediction sequence of each candidate link are output. The time interval that continuously meets the preset stable communication conditions is marked as the link persistence segment, and a link persistence segment set is generated.

5. The multi-path routing control method for a space IoT device according to claim 4, characterized in that, The method of gating and suppressing invalid feature channels of occlusion time slots through occlusion memory feedback gating mechanism includes: The output features of each slice are input into an improved SCINet network with an occlusion memory feedback gating mechanism, which includes a gating generation stage, a historical feature caching stage, and a feedback compensation stage. The gating generation stage receives the visible angle sequence and the output features of each slice. Based on the change state of the visible angle, it generates the gating weights of the corresponding time slices, enabling the improved SCINet network to distinguish between visible time slices, critical time slices, and occluded time slices according to the gating weights. It also performs attenuation control on the feature channels affected by occlusion in the occluded time slice. The historical feature caching stage receives the features output by the Doppler-visible angle quadruple convolution and writes the quadruple convolution features corresponding to the most recent consecutive visible time slices into the historical feature caching area according to the link number, slice level and sampling time, so that the improved SCINet network retains the effective link state expression before the occlusion occurs. During the feedback compensation phase, when an occlusion time slice is detected, the system reads the closest continuous visible quadruple convolutional features from the historical feature cache that belong to the same candidate link as the current occlusion time slice and are in the same time distance. It then performs replacement compensation on the invalid feature channels that are gated and attenuated, merges the replacement and compensated feature channels with the valid feature channels of the current time slice that are not affected by occlusion, and inputs the merged features into the reconstruction module.

6. The multi-path routing control method for a space IoT device according to claim 1, characterized in that, The construction of the cross-path connection graph, which generates a connection time window based on the time overlap interval between continuous link segments, includes: Read the link number, start time, end time, sending node, receiving node, predicted latency, predicted bandwidth, predicted packet loss rate, and connection duration corresponding to each link persistence segment in the link persistence segment set, and build a link persistence segment index according to the link number and start time; According to the spatial node connectivity, the link persistent segments are matched for succession. Segment combinations in which the receiving node of the previous link persistent segment and the sending node of the next link persistent segment have a forwarding connection relationship are marked as candidate succession segment pairs. Overlap detection is performed on the start and end times of candidate segment pairs. The time interval in which the preceding link continuous segment and the following link continuous segment are both in a transmittable state is extracted, and candidate segment pairs that do not have time overlap intervals are removed. Based on the predicted latency, predicted bandwidth, predicted packet loss rate, and connection duration, candidate segment pairs with overlapping time intervals are screened for link communication capabilities. Segment pairs that meet the link communication capability constraints are identified as acceptable segment pairs, and the start and end points of the corresponding overlapping time intervals are written into the acceptance time window. Using the preceding and subsequent path segments as path nodes, and the transmission acceptance relationship between acceptable segment pairs as the acceptance connection relationship, the acceptance time window, spatial node identifier, and link communication capability constraints are written into the graph structure record to generate a cross-path acceptance relationship graph.

7. The multi-path routing control method for a space IoT device according to claim 1, characterized in that, The step of prioritizing data fragments based on their life trajectory record structure and generating a data fragment set includes: Read the IoT business data to be transmitted, extract the business source, business type, data length, target receiving node, service quality requirements, generation time and expiration time, and establish a business data description record; The IoT business data is segmented based on the data length, service quality requirements, and expiration time in the business data description record. Each data segment is written with a segment number, the business number to which it belongs, the segment order, the segment length, the generation time, the expiration time, and the target receiving node. A data slice life trajectory record structure is established for each data slice, and the slice number, entry scheduling time, current path status, path migration record, cumulative waiting time, remaining valid time, amount of data transmitted, amount of data not transmitted, number of path switching and verification status are written into the data slice life trajectory record structure. Based on the data slice life trajectory record structure, read the remaining effective time, cumulative waiting time, amount of untransmitted data, number of path switching and service quality requirements of each data slice. According to the preset time remaining threshold, waiting time threshold, amount of untransmitted data threshold and number of switching threshold, mark the data slices with real-time protection priority, reliable protection priority or energy-saving transmission priority. The priority labeling results are written into the corresponding data slice life trajectory record structure, and all data slices are aggregated according to their business number, slice order, data slice priority and target receiving node to generate a data slice set.

8. The multi-path routing control method for a space IoT device according to claim 1, characterized in that, The construction of a hierarchical elastic transmission window, which performs path allocation, parallel transmission, and risk-weighted replication for each data shard, includes: Read the path nodes, connection relationships, connection time windows, predicted bandwidth, predicted latency, and predicted packet loss rate from the cross-path connection graph, and read the data fragment priority, remaining valid time, amount of untransmitted data, and target receiving node from the data fragment set; A path-level transmission window is constructed based on the acceptance time window. Path nodes that are in the same acceptance time window and meet the connectivity conditions of the target receiving node are written into the candidate path window. The transmission range that each candidate path can carry is determined based on the predicted bandwidth and connection duration. Based on data fragment priority, a fragment-level scheduling window is constructed. Data fragments with real-time guarantee priority are placed in the time-limit constraint window, data fragments with reliable guarantee priority are placed in the replication transmission window, and data fragments with energy-saving transmission priority are placed in the low-load transmission window. Map the fragment-level scheduling window to the path-level sending window. Perform path allocation based on the remaining valid time of the data fragment, the predicted path latency, the predicted path packet loss rate, and the transmission range that the path can carry. For data fragments with real-time guarantee priority, allocate paths that meet the expiration time limit. For data fragments with reliability guarantee priority, allocate the main path and the replication path. For data fragments with energy-saving transmission priority, allocate paths with sufficient remaining bandwidth. Parallel transmission and risk-weighted replication are performed on the data fragments according to the path allocation results, and the allocation path, replication path, sending time, transmission status, amount of data transmitted, amount of data not transmitted, and window adjustment record corresponding to each data fragment are written into the data fragment life trajectory record structure.

9. The multi-path routing control method for a space IoT device according to claim 1, characterized in that, The generation of complete business data output includes: During data transmission, the system reads the link persistence segment, take-up time window, link status, and data slice life trajectory record structure corresponding to the current transmission path, and continuously records the current path, transmission status, amount of data transmitted, and amount of data not transmitted for each data slice. Migration determination is performed based on the end time of the link segment corresponding to the current transmission path, the real-time link status, and the preset threshold. When the link segment ends or the link status is lower than the preset threshold, the current path is marked as a path to be migrated. Based on the cross-path connection diagram, read the connection relationship and connection time window corresponding to the path to be migrated, and determine the successor path from the path nodes that meet the constraints of node connectivity, time connection and link communication capability. The incomplete data fragments in the path to be migrated are switched to the subsequent path for continued transmission, and the switching time, original path identifier, subsequent path identifier, amount of untransmitted data and migration status are written into the corresponding data fragment life trajectory record structure. After the target receiving node receives the data fragments that have been transmitted, it reassembles and verifies the consistency of the data fragments according to their service number, fragment order, and verification status, and combines the data fragments that pass the verification into complete service data for output.

10. A multi-path routing control system for space IoT devices, executing the multi-path routing control method for space IoT devices according to any one of claims 1 to 9, characterized in that, include: The link acquisition module is used to collect and preprocess multi-source link status data between space nodes and ground nodes to generate a link status set. The segment prediction module is used to extract Doppler frequency shift and visible angle sequences based on the link state set, and generate a set of persistent link segments through an improved SCINet network. The connection graph module is used to construct a cross-path connection relationship graph based on the set of persistent links and generate connection time windows; The fragmentation and labeling module is used to fragment IoT business data, establish a data fragment life trajectory record structure and label priorities, and generate a data fragment set. The window scheduling module is used to construct a hierarchical elastic transmission window based on the acceptance time window and the fragmentation priority, and to perform path allocation, parallel transmission and risk-weighted replication. The migration and reorganization module is used to perform path migration based on the cross-path connection diagram, reorganize and verify the completed data fragments, and generate complete business data output.

Images