A remote data transmission method for unmanned automatic weather stations based on cloud-network collaboration

By using a cloud-network collaborative data backhaul method, combined with an improved SOFTS model and the Black-winged Kite optimization algorithm, the problems of data continuity and reliability in the data backhaul of automatic weather stations in uninhabited areas were solved, achieving efficient retransmission and resource optimization.

CN123123474APending Publication Date: 2026-11-17CHENGDU ZHIHENG TIANSHENG TECHNOLOGY CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611268158.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing data transmission methods for automatic weather stations in uninhabited areas struggle to achieve continuous data recovery in weak link environments, resulting in low transmission efficiency, insufficient reliability of remote monitoring, and a lack of precise correlation between cloud reception status and station-side local storage indexes.

Method used

A cloud-network collaborative approach is adopted, which combines a data index table to be retransmitted, an improved SOFTS model and the Blackwing Kite optimization algorithm to identify unclosed back transmission segments, establish a retransmission index, generate a retransmission status control sequence, and perform data loading and compressed packet back transmission under the constraints of satellite short message capacity and station energy.

Benefits of technology

It improves the ability to restore the continuity of meteorological data in weak link environments, enhances the efficiency of retransmission and the reliability of remote backhaul, and optimizes the utilization of short message resources and station-end energy consumption control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123474A_ABST
    Figure CN123123474A_ABST
Patent Text Reader

Abstract

This invention discloses a remote data transmission method for unmanned automatic weather stations based on cloud-network collaboration, relating to the field of communication network data transmission control technology. The method includes the following steps: S1. Collecting operational data and writing it to local storage according to the collection sequence to form a station-end transmission object; S2. Receiving satellite short messages on the cloud and organizing them to form a complete message state; S3. Returning unclosed transmission segments and establishing a data index table to be retransmitted; S4. Extracting data segments to be retransmitted at the station end and forming a retransmission queue; S5. Improving the SOFTS model by rewriting the global core representation of the STAR module into an executable core state and generating state transition markers; S6. Running the Black-winged Kite optimization algorithm to generate a retransmission state control sequence; S7. Compressing and segmenting packets at the station end and transmitting them in batches; S8. Verifying messages, updating the index table, and feeding back the model on the cloud. This invention improves retransmission efficiency and transmission reliability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of communication network data transmission control technology, and in particular to a method for remote data backhaul from an unmanned automatic weather station based on cloud-network collaboration. Background Technology

[0002] Existing automatic weather stations in uninhabited areas are typically deployed in regions with insufficient public network coverage, such as plateaus, deserts, and canyons. The meteorological data collected at these stations mostly relies on satellite short messages for remote transmission. While this method can achieve basic reporting, it is prone to problems such as missing messages, broken sequences, and verification failures due to limitations in short message payload capacity, link latency, terrain obstruction, and station power supply conditions.

[0003] Existing backhaul methods typically employ timed reporting, fixed retransmission, or simple threshold triggering for retransmission processing. They lack precise correlation between cloud reception status, local storage index at the station, and retransmission queue, making it difficult to accurately locate unclosed backhaul segments and dynamically determine the retransmission order under conditions of limited short message capacity and station energy.

[0004] In addition, traditional cloud-based receipts are mostly used for receiving confirmation and fail to participate in retransmission queue updates and retransmission status control, resulting in insufficient data continuity recovery capability in weak link environments in uninhabited areas, and low retransmission efficiency and remote monitoring reliability.

[0005] Therefore, how to provide a remote data transmission method for unmanned automatic weather stations based on cloud-network collaboration 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 remote data backhaul method for unmanned automatic weather stations based on cloud-network collaboration. This invention employs a combination of a data backhaul index table, an improved SOFTS model, and the Black-winged Kite optimization algorithm. By identifying unclosed backhaul segments in the cloud and establishing a backhaul index, the global core representation of the STAR module is rewritten into an achievable core state. Then, based on the cloud achievable results and station-side energy constraints, a backhaul state control sequence is generated. This enables dynamic loading, compression, packetization, and closed-loop backhaul of data segments to be backhauled in weak link environments. It has the advantages of high backhaul efficiency, good data continuity, and strong remote backhaul reliability.

[0007] A remote data transmission method for an automatic weather station in an uninhabited area based on cloud-network collaboration according to an embodiment of the present invention includes the following steps: S1. Automatic weather stations in uninhabited areas collect operational data, write it to local storage according to the collection sequence, and organize it into station-end transmission objects. S2. The cloud platform receives satellite short messages from the target object at the receiving station, performs decoding, verification, and continuity processing to form a complete message status; S3. Read the complete status of the message in the cloud, perform index merging on unclosed return segments, and establish an index table of data to be retransmitted. S4. The station receives the index table of data to be retransmitted, extracts the data segment to be retransmitted from the index item from the local storage, writes it into the retransmission queue, and forms a set of data segments to be retransmitted. S5. Improve the SOFTS model to read the index table of data to be retransmitted and the set of data segments to be retransmitted, rewrite the global core representation of the STAR module to the corresponding position of the index item, form an achievable core state, and generate a state transition flag based on the cloud achievable result. S6. The Black-winged Kite optimization algorithm runs in the cloud. Based on the retrievable core state and state transition marker, it generates a retransmission state control sequence under the constraints of satellite short message capacity and station energy, and writes it in reverse to the STAR module. S7. The station loads the data segments to be retransmitted from the retransmission queue according to the retransmission status control sequence, and transmits them back in batches via satellite short messages after compression and packetization. S8 verifies the satellite short message transmitted back to the cloud, generates a cloud receipt result and updates the data index table to be retransmitted, and feeds the cloud receipt result back to the improved SOFTS model to control the next round of retransmission queue updates.

[0008] Optionally, S1 specifically includes: S11. The station controller of the uninhabited automatic weather station calls the acquisition process to collect operational data and writes the acquisition sequence according to the station clock to form operational data with acquisition sequence. S12. The station controller sequentially writes the operating data with the acquisition timing sequence into the local storage according to the acquisition timing sequence, and records the write address range to form a local timing index. S13. The station controller reads the local timing index and merges the running data that has been written but not yet entered into the backhaul record into the station backhaul record. S14. The station controller binds the station backhaul records to the local time-series index, retains the write address range in the local storage, and organizes them into station backhaul objects.

[0009] Optionally, S2 specifically includes: S21. The satellite short message corresponding to the object returned by the cloud platform receiving station is written into the receiving buffer according to the receiving time sequence to form a satellite short message receiving record. S22. The cloud platform performs protocol decapsulation and payload decoding on the satellite short message reception record to restore the station-end return record and the bound local time sequence index. S23. The cloud platform performs verification and check on the station-end return records based on the load decoding results, and writes the verification and check results into the corresponding local time series index. S24. The cloud platform organizes the reception continuity of the station's backhaul records according to the local time sequence index to form a complete message status.

[0010] Optionally, S3 specifically includes: S31. The cloud platform reads the complete status of the message, checks the station's back transmission record according to the local time sequence index, and marks the discontinuous or unverified local time sequence index as a reception breakpoint. S32. The cloud platform reads adjacent received indices based on the receiving breakpoint, determines the start and end indices of unclosed back transmission segments, and forms a missing segment index. S33. The cloud platform merges adjacent or overlapping missing test segment indexes and removes the index range that has been covered by the verified station-end return records to form a merged index. S34. The cloud platform will associate the merged index with the back-up object from the site and create an index table of data to be retransmitted.

[0011] Optionally, S4 specifically includes: S41. The station controller parses the index table of data to be retransmitted, reads the local time-series index corresponding to the index item, and forms an index query sequence. S42. The station controller retrieves the local time-series index based on the index query sequence, obtains the write address range, and reads the running data from the local storage. S43. The station controller restores the acquisition order of the read running data according to the local time sequence index and binds it with the source index item to form a data segment to be transmitted. S44. The station controller writes the data segments to be retransmitted into the retransmission queue according to the source index item, forming a set of data segments to be retransmitted.

[0012] Optionally, the improved SOFTS model includes a reversible instance normalization layer, a sequence embedding layer, a multi-layer STAR module, and a linear predictor. The STAR module includes a first multilayer perceptron, a random pooling layer, a repetition and splicing operation layer, a second multilayer perceptron, and residual connections; The STAR module has a core state storage area that is connected to the index entries of the index table for the data to be retransmitted. The core state storage area holds the retrievable core state, and the state transition flag is bound to the retrievable core state.

[0013] Optionally, S5 specifically includes: S51. The sequence embedding layer reads the set of data segments to be retransmitted, maps the data segments to be retransmitted into sequence representations, and writes the source index item into the index position of the sequence representation. S52. The first multilayer perceptron reads the sequence representation and generates a hidden representation, and the hidden representation maintains the correspondence between the source index item and the hidden representation channel; S53. The core status storage area establishes the connection relationship between the corresponding position of the index item and the hidden representation channel according to the data to be retransmitted index table; S54. The random pooling layer reads the hidden representation channel according to the source index item, writes the pooling result of the same source hidden representation to the corresponding position of the index item, and forms an achievable core state. S55. When the source index item does not hit the hidden representation channel in this round, the core state storage area retains the written executable core state along the corresponding position of the index item, and the sequence representation corresponding to the source index item that has not formed an executable core state is retained along the residual connection. S56. The repeat and splice operation layer reads the retrievable core state according to the source index item and repeatedly splices the retrievable core state to the same source sequence representation. S57. The second multilayer perceptron fuses the homologous sequence representations after repeated splicing, and the residual connection writes the fusion result back to the sequence representations of the same source index items; S58. The core state storage area writes the cloud receipt results to the corresponding position of the index item and binds them with the receivable core state to generate a state transition marker.

[0014] Optionally, S6 specifically includes: S61. The cloud platform runs the Blackwing Kite optimization algorithm, reads the retrievable core state and state transition marker, and encodes the position corresponding to the index item as the individual position. S62. The fitness of an individual location is calculated using the completeness of the index item retransmission corresponding to the retrievable core status as the benefit term, and the length of the data segment to be retransmitted, the upper limit of the satellite short message payload, and the remaining energy at the station as the constraint terms. S63. Execute the Blackwing Kite optimization algorithm to update attack behavior and migration behavior according to fitness, and adjust the control order of index items corresponding to individual positions; S64. When the updated individual position exceeds the index range of the data to be retransmitted index table, the individual position will be rolled back to the corresponding position of the adjacent index item that has not been retransmitted. S65. Generate a supplementary transmission state control sequence according to the updated individual position, and write the supplementary transmission state control sequence into the core state storage area of ​​the STAR module.

[0015] Optionally, S7 specifically includes: S71. The station controller parses the retransmission status control sequence, matches the retransmission queue according to the corresponding position of the index item, reads the data segment to be retransmitted, and removes the sent data segment according to the source index item and the local sending record to form the current retransmission data segment. S72. The current retransmission data segment is arranged according to the acquisition time sequence and then compressed and encoded. The source index item and the acquisition time sequence are retained with the compression result to form a compressed retransmission payload. S73. The compressed retransmission payload is packaged and encapsulated according to the upper limit of the satellite short message payload. The packet order is established to correspond with the source index item to form a retransmission satellite short message. S74. The satellite short message is sent via the satellite short message communication link, and the sending result, source index item and packet order are written into the local sending record.

[0016] Optionally, S8 specifically includes: S81. The cloud platform receives the retransmitted satellite short message, performs decapsulation according to the source index item and packet order, and forms a retransmission reception record. S82. After integrity verification and decompression, the retransmission received record is reassembled into a retransmission restored data segment according to the source index item and the acquisition time sequence. S83. The retransmission and restoration data segment is matched with the index range of the data index table to be retransmitted. The index range that has been matched is written to the retransmission completion status, and the index range that has not been matched is retained as an unclosed index range. S84. Generate cloud receipt results from the completed retransmission status and unclosed index range. Write the cloud receipt results into the core state storage area of ​​the improved SOFTS model and bind them to the receivable core state. S85. The core state storage area updates the state transition flag based on the cloud receipt results, and writes the unclosed index range back to the data to be retransmitted index table, forming the basis for updating the next round of retransmission queue.

[0017] The beneficial effects of this invention are: This invention, through collaborative processing between a cloud platform and local storage at the station, can index operational data collected by automatic weather stations in uninhabited areas according to the collection sequence. After satellite short message transmission, the cloud performs decoding, verification, and continuity processing to accurately identify unclosed transmission segments and establish an index table of data to be retransmitted. This avoids the problems of unclear missing measurement locations and excessive retransmissions in traditional timed reporting or fixed retransmission methods, allowing the retransmission targets at the station to directly correspond to the actual data segments in local storage, thus improving the continuity recovery capability of meteorological data in weak link environments.

[0018] This invention introduces an improved SOFTS model, rewriting the global core representation of the STAR module to the corresponding index entry in the data to be retransmitted index table, forming a verifiable core state, and enabling the cloud-based verification result to enter the core state storage area to generate a state transition marker. In this way, the retransmission state no longer remains at the level of external confirmation information, but participates in the updating and redistribution of the model's internal core representation. This allows index entries from different sources to maintain, update, or continue to participate in retransmission control according to the verification state, improving the adaptability of retransmission decisions to link status and data integrity status.

[0019] This invention also incorporates the Black-winged Kite optimization algorithm to generate a retransmission state control sequence under the constraints of satellite short message capacity and station energy. This sequence guides the station to load the data segments to be retransmitted from the retransmission queue and compress and retransmit them back. After verification in the cloud, a receipt result is generated and written back to the data index table to be retransmitted, forming a closed-loop control process of station data retrieval, short message back transmission, cloud verification, and the next round of retransmission updates. This process has the advantages of high retransmission efficiency, high short message resource utilization, strong station energy consumption control capability, and good remote back transmission reliability. 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 an overall flowchart of a remote data transmission method for an automatic weather station in an uninhabited area based on cloud-network collaboration, as proposed in this invention. Figure 2 This is a schematic diagram of the improved SOFTS model for a remote data transmission method for unmanned automatic weather stations based on cloud-network collaboration, as proposed in this invention. Figure 3 This is a data flow diagram of the supplementary transmission index and receipt closed loop for a remote data backhaul method for unmanned automatic weather stations based on cloud-network collaboration 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 Figures 1-3 A method for remote data transmission from an automatic weather station in an uninhabited area based on cloud-network collaboration includes the following steps: S1. Automatic weather stations in uninhabited areas collect operational data, write it to local storage according to the collection sequence, and organize it into station-end transmission objects. S2. The cloud platform receives satellite short messages from the target object at the receiving station, performs decoding, verification, and continuity processing to form a complete message status; S3. Read the complete status of the message in the cloud, perform index merging on unclosed return segments, and establish an index table of data to be retransmitted. S4. The station receives the index table of data to be retransmitted, extracts the data segment to be retransmitted from the index item from the local storage, writes it into the retransmission queue, and forms a set of data segments to be retransmitted. S5. Improve the SOFTS model to read the index table of data to be retransmitted and the set of data segments to be retransmitted, rewrite the global core representation of the STAR module to the corresponding position of the index item, form an achievable core state, and generate a state transition flag based on the cloud achievable result. S6. The Black-winged Kite optimization algorithm runs in the cloud. Based on the retrievable core state and state transition marker, it generates a retransmission state control sequence under the constraints of satellite short message capacity and station energy, and writes it in reverse to the STAR module. S7. The station loads the data segments to be retransmitted from the retransmission queue according to the retransmission status control sequence, and transmits them back in batches via satellite short messages after compression and packetization. S8 verifies the satellite short message transmitted back to the cloud, generates a cloud receipt result and updates the data index table to be retransmitted, and feeds the cloud receipt result back to the improved SOFTS model to control the next round of retransmission queue updates.

[0023] In this embodiment, S1 specifically includes: After power-on self-test, the station controller of the uninhabited automatic weather station starts the data acquisition process. The acquisition process reads meteorological data, station operation status data, communication status data and power supply status data. The meteorological data comes from the rain sensor and the supporting environmental sensor. The station operation status data comes from the controller self-test results. The communication status data comes from the satellite short message module. The power supply status data comes from the power management unit.

[0024] The station clock is continuously timed by the station real-time clock. When the timing module outputs valid timing information, it performs calibration on the real-time clock. The acquisition process writes the acquisition sequence each time it reads running data. The acquisition sequence consists of the station number and the acquisition timestamp. The station number is used to distinguish the meteorological station source, and the acquisition timestamp is used to determine the order of data within the same station.

[0025] In one embodiment, the rainfall collection interval can be 5 minutes, the tipping bucket resolution can be 0.2 mm, and the collection process reads the cumulative rainfall increment within the interval at the end of a collection interval, and at the same time reads the battery voltage, communication module working status and storage status to form operation data with collection timing.

[0026] Local storage uses FLASH or non-volatile storage. The station controller sequentially writes the running data with acquisition timing to the storage page. The write address range consists of the start address, the end address, and the data length. The data length is obtained by subtracting the start address from the end address and adding the length of the first byte. After the write verification is successful, it is registered in the local timing index.

[0027] The local time series index uses the acquisition time series as the index key. The index key corresponds to the write address range and the write completion flag. The write completion flag is set to complete after the write to the storage page is finished and the check value is consistent. When power is off and the system is restored, the station controller reads the index key backward from the most recent completion flag. The running data corresponding to the incomplete flag is not included in the station backhaul object.

[0028] When the station controller reads the local time sequence index, it also checks the return record flag. The running data that has been written and whose return record flag is empty is merged into the station return record. The merging order is determined by the collection timestamp from earliest to latest. Running data that has already been registered as a return record is not merged again.

[0029] When the station-side backhaul record is bound to the local time series index, the binding content includes the collection time series, write address range, data length and data integrity status. The data integrity status is determined by the write verification result. The station-side backhaul object consists of the station-side backhaul record and the corresponding binding content, and serves as the basic data for cloud-based continuous processing and unclosed backhaul segment positioning.

[0030] In this embodiment, S2 specifically includes: The cloud platform sets up a receiving cache to receive satellite short messages corresponding to the objects returned by the receiving station. The receiving cache uses the cloud receiving timestamp as the receiving sequence. The receiving sequence is only used to distinguish the order of arrival of satellite short messages. The station's collection sequence is used to restore the continuous collection order of running data.

[0031] When a satellite short message enters the receiving buffer, the cloud platform writes a receiving timestamp, link receiving result, and buffer writing position for each message. The buffer writing position is determined by the current tail address of the receiving queue. After writing is completed, a satellite short message receiving record is generated. Messages that have not been written do not enter the payload decoding process.

[0032] The protocol decapsulation process first reads the header, payload length, and check field of the satellite short message. If the payload length matches the actual payload byte count, the payload data is extracted. If the length does not match, the check failure result is written to the message to prevent incomplete payloads from being restored as station-end return records.

[0033] Payload decoding is performed after the protocol decapsulation is completed. The cloud platform restores the station-end return records and bound local time sequence index according to the encapsulation format of the station-end return object. The local time sequence index carries the collection timing and write address range, and the cloud platform determines the station-end collection location corresponding to the message based on this.

[0034] The verification check processes the integrity of the payload, the consistency between the station-end backhaul records and the local time series index, and whether the same index range is repeatedly received. The integrity of the payload is determined by the decapsulation verification result, the consistency is determined by whether the acquisition time series and the index range match, and the repeated reception is determined by whether the same station-end number and the same local time series index range have been registered.

[0035] The verification results are written to the corresponding local time-series index. The written content includes status flags for valid reception, no reception, verification failure, and duplicate reception. For duplicate receptions that pass verification, the first valid reception record is retained, and the remaining duplicate records are only saved as reception logs and do not participate in the generation of missing test segments.

[0036] The continuous sorting process uses the local time-series index as the sorting benchmark. The cloud platform checks the coverage of the records returned from the station from the earliest index to the latest index. The index range that is continuously covered and passes the verification is written to the received status, while the index range that is missing valid records or fails the verification is written to the sequence gap status.

[0037] The complete status of a message is formed by the local time sequence index, the coverage result of the station's backhaul record, the verification result, and the sequence gap status. The cloud platform saves the complete status of the message as a searchable record, which serves as the input for locating unclosed backhaul segments and establishing the index table of data to be retransmitted.

[0038] In this embodiment, S3 specifically includes: After the cloud platform reads the complete status of the message, it distinguishes different uninhabited automatic weather stations by station number and arranges the station-transmitted records from morning to night by the collection timestamp of the local time sequence index. The arranged data is used to check for reception breakpoints.

[0039] The reception breakpoint is determined by the continuity of the local time-series index and the verification result. When the current index lacks a valid reception status or a write verification failure status, it is marked as a reception breakpoint. When the same index range is repeatedly received but a valid reception status exists, it is not marked as a reception breakpoint.

[0040] The continuity of the local time sequence index is calculated by the difference between adjacent acquisition timestamps. When the difference is equal to the acquisition interval, it is considered continuous. When the difference is greater than the acquisition interval, the missing index range in the middle is marked as a reception breakpoint. When the difference is less than the acquisition interval, it is considered as repeated reception or out-of-order reception. Index ranges that are repeatedly received but have a valid reception status are not included in the missing test segment index.

[0041] The cloud platform reads adjacent received indices with the receiving breakpoint as the boundary. The index after the last valid received index before the breakpoint is used as the starting index of the unclosed back transmission segment, and the index before the first valid received index after the breakpoint is used as the ending index. The missing segment index is formed between the starting index and the ending index.

[0042] When the receiving breakpoint is before a received index, the earliest unreceived index is returned by the starting index retrieval station; when the receiving breakpoint is after a received index, the latest registered index is returned by the terminating index retrieval station. If the boundary index cannot be determined, the breakpoint record is retained and the station waits for the next round of message integrity status update.

[0043] When merging missing test segment indexes, the cloud platform first sorts them by starting index from smallest to largest, then compares the start and end ranges of adjacent missing test segments. When the starting index of the next segment is not greater than the next index of the ending index of the previous segment, they are merged into a continuous index range.

[0044] The index range that has been covered by the verified station-end return records is removed from the merge results. The removal is based on the consistency of the station-end number, local time-series index, and collection timestamp. The remaining uncovered index range forms the merged index.

[0045] When the merge index is associated with the site-side return object, the cloud platform writes the merge index to the corresponding site-side number and retains the local time-series index range corresponding to the merge index. After the association is completed, the site can return to local storage to read the corresponding write address range.

[0046] The data to be retransmitted index table consists of a merge index, the corresponding station number, the local time-series index range, and the retransmission status. The initial write of the retransmission status is not closed. After the index range that has completed retransmission is verified in the cloud, it is rewritten as retransmission completed. The index range that has not completed retransmission continues to be retained as the basis for the next round of retransmission.

[0047] In this embodiment, S4 specifically includes: After receiving the data index table to be retransmitted from the cloud, the station controller first verifies whether the station number is consistent with the local number, and then reads the local time-series index range in the index item. The index range is limited by the start collection time sequence and the end collection time sequence. After parsing, an index query sequence is formed.

[0048] The index query sequence is arranged from earliest to latest according to the collection time sequence. The arrangement is based on the numerical value of the collection timestamp. The index entries with smaller timestamps are placed at the front. When multiple sub-package records correspond to the same timestamp, the order is determined by the starting value of the write address in the local time sequence index to avoid the time sequence of the read running data being reversed.

[0049] The station controller uses the index query sequence to retrieve the local time-series index. The local time-series index uses the acquisition time sequence as the index key. After the retrieval is successful, the corresponding write address range is read. The write address range includes the start address, end address, and data length. The data length is used to limit the range of bytes read from local storage this time.

[0050] When reading operational data from local storage, the station controller reads continuously from the start address to the end address and compares the actual number of bytes read with the data length. If the two are consistent and the storage check value is consistent, the read is considered valid, and the valid operational data enters the acquisition sequence recovery process.

[0051] When a local time-series index exists but reading from the address range fails, the station controller writes the corresponding source index item into the local missing status. The local missing status is bound to the source index item and retained in the retransmission queue control record, which is used by the cloud receipt to distinguish between communication missing and local data missing.

[0052] Read valid running data and restore the collection order according to the local time-series index. The restoration process uses the collection timestamp as the primary sorting object and the write address as the secondary sorting object under the same collection timestamp. After sorting, bind it with the source index item to form the data segment to be transmitted.

[0053] When the data segment to be retransmitted is written to the retransmission queue, the source index is written along with the data segment. The queue entry records the source index, the acquisition timing range, the data length, and the reading status. The retransmission queue does not change the acquisition timing and only saves the data segment entry that can be called by the retransmission status control sequence.

[0054] After the retransmission queue is completed, the station controller summarizes all the valid data segments to be retransmitted and forms a set of data segments to be retransmitted. The set of data segments to be retransmitted retains the source index and the acquisition time range, which serve as the input for generating the retrievable core state of the improved SOFTS model.

[0055] In this embodiment, the improved SOFTS model includes a reversible instance normalization layer, a sequence embedding layer, a multi-layer STAR module, and a linear predictor: The improved SOFTS model adopts the basic time series modeling framework of SOFTS. The model includes a reversible instance normalization layer, a sequence embedding layer, a multi-layer STAR module, and a linear predictor. The STAR module includes a first multi-layer perceptron, a random pooling layer, a repetition and splicing operation layer, a second multi-layer perceptron, and residual connections.

[0056] In the forward computation, the original STAR module projects the channel sequence representations through the first multilayer perceptron, and then the global core representation shared by the random pooling layer is generated. This representation is then fed back to the channel sequence representations through the repetition and splicing operation layer. After fusion by the second multilayer perceptron, it is written back through residual connections.

[0057] The global core representation shared by the module only participates in the fusion temporarily in a single forward computation. It cannot fall into the corresponding position of the index item in the data to be retransmitted index table. The cloud receipt result cannot change the source of the core representation in the next round. The retransmission status of completed, unclosed, and waiting for receipt cannot enter the internal computation path of the STAR module.

[0058] The improved SOFTS model adds a core state storage area inside the STAR module. The core state storage area is connected to the output of the random pooling layer, the input of the repetition and splicing operation layer, and the cloud receipt result writing end. The core representation of the index item output by the random pooling layer is written into the core state storage area to form a receivable core state.

[0059] The core state storage area uses state tensors to carry retrievable core states. The rows of the state tensor correspond to the source index entries in the data to be transmitted index table, and the columns correspond to the hidden representation channels. Each position is written with the core state component under the corresponding source index entry, and the state transition flag is bound to the retrievable core state in the same row.

[0060] In one implementation, the input to the improved SOFTS model can be in the format B×I×30×4, where B represents the training batch, I represents the number of source index items, 30 represents the number of consecutive sampling points covered by each source index item, and 4 represents the four types of features: meteorological data collection, station operation, communication status, and power supply status.

[0061] The reversible instance normalization layer calculates the mean and standard deviation of each source index item and each class feature along the sampling point dimension. The normalized value is obtained by subtracting the mean from the original feature and dividing by the sum of the standard deviation and 1×10-5. The mean and standard deviation are saved to the current forward batch and used for denormalization at the output end.

[0062] The sequence embedding layer projects four types of features into a sequence representation of the hidden dimension D. In one implementation, D can be 64. The projection process involves multiplying the input feature vector with the embedding weight matrix and then superimposing the embedding bias. The output format is B×I×30×64, and the position of the source index term in the I dimension remains unchanged.

[0063] The first multilayer perceptron receives a sequence representation of B×I×30×64, first maps it to a 128-dimensional intermediate representation through the first linear layer, then compresses the nonlinear response through the GELU function, and finally maps it back to a 64-dimensional hidden representation through the second linear layer, maintaining the output format as B×I×30×64.

[0064] In the improved SOFTS model, the random pooling layer only reads the hidden representation channels within index terms from the same source, calculates the average pooling value along 30 sampling points, and forms the core representation of the index terms in B×I×64 format, instead of mixing index terms from different sources into a single shared core representation.

[0065] The core state storage area receives the core representation of the index item in B×I×64 format. The 64-dimensional vector corresponding to the i-th source index item is written to the i-th row. Rows marked as incomplete by the cloud acknowledgment results retain the original acknowledgment core state. Rows that do not hit the hidden representation in this round are marked with a valid flag 0 and the source index item is retained.

[0066] The repetition and splicing operation layer reads the B×I×64 retrievable core state from the core state storage area, repeats the 64-dimensional state vector of each source index item 30 times along the sampling point dimension, and then splices it with the same source sequence representation to form a spliced ​​representation in B×I×30×128 format.

[0067] The second multilayer perceptron compresses the B×I×30×128 format spliced ​​representation back to 64 dimensions. The residual connection adds the compressed result to the B×I×30×64 sequence representation before entering the STAR module position by position. The addition positions have the same batch number, the same source index item, and the same sampling point number.

[0068] The linear predictor reads the index item update representation output by the last STAR module, averages it along the sampling point dimension to obtain a B×I×64 readout vector, multiplies the readout vector by the completeness weight vector and adds the completeness bias, and then obtains the supplementary completeness estimate through the Sigmoid function. The readout vector is then passed through a state readout linear layer to obtain the readout probabilities of the completed class, the unclosed class and the waiting class. The state readout probabilities are used to constrain the model representation during the training phase. Finally, the state transition marker is determined after the cloud acknowledgment result is written to the core state storage area.

[0069] The training data comes from historical transmission records of automatic weather stations in uninhabited areas. The station's local storage records, satellite short message reception records, cloud verification results, and cloud receipt results are aligned according to the source index item. Each training sample corresponds to a sequence of running data covered by one source index item.

[0070] Training labels are automatically constructed from the actual cloud acknowledgment results. Source index items that have been retransmitted are written to the completed category label. Source index items with missing packets, failed verification, or discontinuous collection timing are written to the unclosed category label. Source index items that have entered the retransmission queue but have not received cloud acknowledgment are written to the waiting category label.

[0071] The completeness label for retransmission is obtained by dividing the number of sample points that pass the verification by the number of sample points that should be retransmitted in the source index item. The value range is 0 to 1. The sample points that pass the verification must simultaneously meet the requirements of consistent source index item, consistent collection timing, and valid verification result. Source index items with a denominator of 0 will not be included in the training batch.

[0072] The loss function is obtained by weighted summation of state readout cross-entropy loss, mean squared error loss of retransmission completeness, and index item ranking loss. In one embodiment, the weights of the three items can be 0.5, 0.3, and 0.2. The state readout cross-entropy loss is calculated from the three types of acknowledgment state labels and state readout probabilities. The mean squared error loss of retransmission completeness is calculated from the squared difference between the retransmission completeness label and the estimated retransmission completeness. The index item ranking loss is calculated in pairs within the same training batch. A loss is generated when the readout response of a source index item with a higher degree of non-closure is more than 0.1 lower than the readout response of a source index item with a lower degree of non-closure.

[0073] Model parameters are obtained through training. The training parameters include sequence embedding layer weights, first multilayer perceptron weights, second multilayer perceptron weights, and linear predictor weights. The core state storage area is not stored as fixed training weights, but is dynamically generated during inference from the data index table to be supplemented and the cloud receipt results.

[0074] In one implementation, the AdamW optimizer is used for training, with a learning rate of 1×10⁻³, a batch size of 64, a maximum number of training epochs of 120, and a scaling method that scales all gradients to a norm of no more than 1.0 when the gradient norm exceeds 1.0.

[0075] Convergence determination is performed on the validation set. Training stops when the validation set loss decreases by less than 1×10-3 for 10 consecutive rounds. Training also stops when the maximum number of training rounds is reached. After training stops, the model parameters with the lowest validation set loss are loaded and the inference process begins.

[0076] During inference, the current set of data segments to be retransmitted enters the reversible instance normalization layer and sequence embedding layer. The weights and biases obtained during training complete the forward calculation. The core state storage area allocates the corresponding position of the index item according to the index table of the current data to be retransmitted. The retrieval core state and state transition flag are written along with the current cloud retrieval result and transmitted to the retransmission control calculation.

[0077] In this embodiment, S5 specifically includes: The original STAR module projects the sequence representation of all channels to generate a global core representation shared by the module. This global core representation is then copied back to all channels. The acknowledgment differences between source index items cannot enter the random pooling entry and residual write-back boundary. The cloud acknowledgment results can only remain in the external record of the model.

[0078] Before the set of data segments to be retransmitted enters the sequence embedding layer, the data segments are organized into input tensors of batch dimension, index term dimension, sampling point dimension and feature dimension according to the source index item. In one embodiment, the input tensor can be B×I×30×4, where B represents the batch size, I represents the number of source index items in the current data to be retransmitted index table, 30 represents the number of continuous sampling points covered by each source index item, and 4 represents the number of upper-level feature categories of the running data.

[0079] After the sequence embedding layer reads the set of data segments to be supplemented, it performs linear projection on the 4-dimensional feature vector of each sampling point. The linear projection is obtained by multiplying the feature vector with the embedding weight matrix and superimposing the embedding bias. In one embodiment, the embedding dimension can be 64, and the output forms a sequence representation in the format of B×I×30×64.

[0080] When a source index item is written to the index position represented by the sequence, the index position is determined by the arrangement order of the source index items in the data to be supplemented index table. The i-th source index item is written to the i-th position in dimension I. The source index item is saved along with the collection time series range, so that the data segments of the same source index item maintain the same index position within the model.

[0081] The first multilayer perceptron consists of an input linear layer, a GELU nonlinear layer, and an output linear layer. The input linear layer maps the 64-dimensional sequence representation to a 128-dimensional intermediate representation. The GELU nonlinear layer compresses anomalous amplitude responses. The output linear layer then maps back to the 64-dimensional hidden representation, which maintains the B×I×30×64 format.

[0082] The correspondence between the hidden representation channel and the source index item is maintained by the I dimension. All 30 sampling points under the i-th source index item are written into the i-th hidden representation channel. The first multilayer perceptron only changes the values ​​in the feature dimension, without changing the batch number, the position of the source index item, or the order of the sampling points.

[0083] The core state storage area includes an index mapping table, a core state tensor, valid markers, and receipt records. The index mapping table records the correspondence between source index items and I-dimensional positions. The core state tensor stores the retrievable core state in B×I×64 format. The valid marker records whether there is already an available state at that position. The receipt record stores the cloud receipt results.

[0084] When establishing the connection between the corresponding position of the index item and the hidden representation channel in the core state storage area, the index mapping table first reads the position of the source index item in the data to be transmitted index table, and then connects the position to the I-dimensional channel of the hidden representation. The i-th source index item can only read the i-th group of hidden representation channels and write the i-th row of the core state tensor.

[0085] The random pooling layer performs sampling and aggregation within the same source index. The sampling mask is obtained by inputting the source index, batch number and the current training round into the pseudo-random generator. During the inference phase, a fixed sampling mask is generated by the source index and the current number of times the retransmission queue is updated. In one implementation, the sampling ratio can be 60%, with 30 sampling points corresponding to 18 sampling positions.

[0086] The same-source hidden representation is a hidden representation with the same source index. The random pooling layer only reads the valid sampling points in the same-source hidden representation, accumulates the sampled 64-dimensional vector along the sampling point dimension, and then divides it by the number of valid sampling points to obtain the 64-dimensional pooling result. When the number of valid sampling points is 0, no new pooling result is generated, and the core state storage area enters the miss processing.

[0087] When the source index item does not hit the hidden representation channel in this round, the core state storage area first reads the valid flag. When the valid flag is 1, the original retrieval core state is retained and the original retrieval core state is continued to be provided to the repetition and splicing operation layer. When the valid flag is 0, no new retrieval core state is generated, the sequence representation of the corresponding source index item does not perform core state splicing, and is only retained along the residual connection and enters the STAR module output side.

[0088] When the repetition and splicing operation layer reads the retrievable core state, it first determines the position of the source index item by the index mapping table, and then reads the 64-dimensional state vector at the same position in the core state tensor. The state vector is repeated along 30 sampling points to form a repetitive state representation in B×I×30×64 format.

[0089] Repeated state representations and homologous sequence representations are concatenated along the feature dimension. Each position after concatenation contains a 64-dimensional original sequence representation and a 64-dimensional reproducible core state. The output format is B×I×30×128. Positions that do not form reproducible core states enter the residual path with the original sequence representation, and the valid label remains 0.

[0090] The second multilayer perceptron consists of a fusion linear layer, a GELU nonlinear layer, and a write-back linear layer. The fusion linear layer maps the 128-dimensional spliced ​​representation to a 128-dimensional fusion representation. The GELU nonlinear layer compresses the abrupt response, and the write-back linear layer compresses the fusion representation into a 64-dimensional write-back representation.

[0091] The residual connection adds the 64-dimensional write-back representation to the 64-dimensional sequence representation before entering the STAR module position by position. Position-by-position addition only occurs between the same batch number, the same source index item, and the same sampling point number. The sequence representation that does not hit the source index item retains the original residual output, thus avoiding state crosstalk between different source index items.

[0092] After receiving the cloud receipt, the core state storage area reads the source index item, packet order, and retransmission completion status from the receipt record. The source index item is used to locate the core state tensor line, the packet order is used to verify the retransmission fragments under the same source index item, and the retransmission completion status is used to update the validity flag and state transition flag.

[0093] The state transition marker is generated by the cloud receipt result and the valid marker. In one embodiment, the state transition marker adopts three types of encoding: 100 is written after the retransmission is completed, 010 is written if there is still an unclosed index range, and 001 is written if it has been loaded but no receipt has been received. The encoding is written to the corresponding position of the source index item and bound to the achievable core state.

[0094] After the core state and state transition marker are bound together, the model outputs a core state tensor in B×I×64 format and a state transition marker tensor in B×I×3 format. The two are then used together to generate the supplementary state control sequence, enabling the Blackwing Kite optimization algorithm to read the internal model state and cloud receipt state of each source index item.

[0095] Compared to the original STAR module, the improvements are as follows: the core representation is changed from a one-time shared vector to the internal state occupied by the source index item; the random pooling entry is changed from global channel mixing to homo-source hidden representation aggregation; the repetition and splicing exit is changed from full-channel copying to homo-source redistribution; the residual write-back boundary is changed from module-level write-back to index item-level write-back; and the cloud receipt result is changed from external confirmation information to the model's internal state that can change the core state's path. As a result, the differences in completion, incomplete closure, and waiting for receipt during the weak link retransmission process directly change the pooling range, splicing object, and retransmission control input of the next round of STAR module.

[0096] In this embodiment, S6 specifically includes: During the retransmission state control sequence generation stage, the retrievalable core state and state transition flag are received. The retrievalable core state comes from the B×I×64 state tensor in the core state storage area, and the state transition flag comes from the B×I×3 encoding of the same source index entry position. I represents the number of source index entries in the current data to be retransmitted index table.

[0097] The capacity of a satellite short message is determined by the upper limit of the payload per packet and the number of packets currently allowed to be sent. The upper limit of the payload per packet comes from the link configuration of the satellite short message communication module. The number of packets currently allowed to be sent is obtained by dividing the communication window length issued by the cloud by the transmission duration of a single packet and rounding down. In one embodiment, the payload per packet can be 116 bytes.

[0098] The station-end energy is determined by the station-end's most recent power supply status data. The remaining energy is obtained by multiplying the battery's rated capacity by the current state of charge. The available communication energy is obtained by subtracting the energy required for the acquisition process and hibernation from the remaining energy. If the available communication energy is less than the energy transmitted per packet, the source index item will not be included in the current loading calculation.

[0099] When the available communication energy is less than the single packet transmission energy, no new loading order is generated in this round. The core state storage area will keep the corresponding source index item in a waiting state. The data to be retransmitted index table retains the unclosed index range. The station will re-enter the retransmission state control sequence generation process after the next power supply status data update.

[0100] Individual location is not a geographic location. Individual location represents the location corresponding to the selectable index item in the data index table to be supplemented. Each individual consists of a set of real number locations. The location value is rounded to obtain the source index item number. When the number is within the range of 1 to I, fitness calculation is performed. When the number is less than 1 or greater than I, out-of-bounds rollback is performed.

[0101] In one implementation, the population size can be 20, the maximum number of iterations can be 50, and the number of positions of a single individual is equal to the upper limit of the number of index items that can be loaded in this round. The upper limit of the number of index items that can be loaded is determined by the smaller of the number of packets that can be sent supported by available communication energy and the number of packets that can be sent supported by satellite short message capacity.

[0102] The index item retransmission completeness is obtained by reading the retrievable core state. The reading process involves multiplying the 64-dimensional retrievable core state and the completeness weight vector dimension by dimension, summing the results, adding the completeness bias, and inputting the result into the Sigmoid function. The output value is limited to the range of 0 to 1. The unclosed benefit is equal to 1 minus the retransmission completeness. The lower the retransmission completeness, the greater the impact of missing data on the source index item in the current rainfall process.

[0103] The length of the data segment to be retransmitted comes from the data length of the index item with the same source in the retransmission queue. The capacity usage is obtained by dividing the length of the data segment to be retransmitted by the upper limit of the single packet payload and then rounding up. The energy usage is obtained by multiplying the capacity usage by the single packet transmission energy and adding the compression processing energy.

[0104] The score for a single source index item is calculated from the unclosed revenue, capacity utilization, and energy utilization. Capacity utilization equals the number of packets loaded divided by the number of packets currently allowed to be sent, and energy utilization equals the energy utilization divided by the available communication energy. The score equals the revenue weight multiplied by the unclosed revenue, minus the capacity weight multiplied by the capacity utilization, and then minus the energy weight multiplied by the energy utilization.

[0105] Individual fitness is obtained by summing the scores of all source index items corresponding to the individual's location. For repeated source index items, only the score of the first occurrence is calculated. Source index items with a capacity utilization rate or energy utilization rate exceeding 1 are not included in the summation. In one embodiment, the revenue weight, capacity weight, and energy weight can be 0.6, 0.25, and 0.15, respectively, and the sum of the three weights is 1. The weights are determined by the parameter group with the lowest average retransmission completion latency in the historical backhaul record playback test.

[0106] The attack behavior update is used to move the current individual closer to the individual with the highest fitness. The attack direction is obtained by subtracting the current individual's position vector from the optimal individual's position vector. The attack direction is multiplied by the attack step size and a uniformly random number between 0 and 1 and then added to the current individual's position. The attack step size is equal to 1 minus the current round divided by the maximum number of iterations.

[0107] The migration behavior update is used to expand the search range of incomplete index items. First, the migration target position is selected from the source index items with unclosed benefits greater than 0. Then, the migration direction is obtained by subtracting the updated individual position from the migration target position. The migration direction is multiplied by the migration coefficient and the Cauchy sampling value and then superimposed on the individual position. The migration coefficient is equal to 1 minus the current round divided by the maximum iteration round. The Cauchy sampling value is limited to the range of -2 to 2.

[0108] After the attack behavior update and migration behavior update are completed, the individual position is decoded into the source index item control order. During decoding, the source index items are read first according to the order of the individual position in the vector. For repeated source index items, only the first occurrence position is retained, and source index items that have been retransmitted are removed.

[0109] When the updated individual location exceeds the index range of the data to be retransmitted index table, first read the source index item corresponding to the location before the update, then search forward and backward along the index item sequence number for adjacent source index items that have not been retransmitted. If the distances before and after are the same, select the source index item with higher unclosed benefit.

[0110] The retransmission status control sequence consists of the source index item control order, the number of packets loaded for each source index item, the data length before compression, and the write position of the core status storage area. The number of packets loaded is obtained by dividing the length of the data segment to be retransmitted by the upper limit of the single packet payload and then rounding up. Tail source index items that exceed the current allowed number of packets to be sent will not enter the current round sequence.

[0111] When the retransmission status control sequence is written to the core status storage area of ​​the STAR module, the control sequence number and the current loading status are recorded at the corresponding position of the same source index item. The control sequence number is used by the station to read the data segment to be retransmitted from the retransmission queue, and the current loading status is used by the STAR module to identify the source index item that has entered the transmission process but still needs to wait for the cloud acknowledgment in the next round.

[0112] When the cloud receipt has not arrived, the core status storage area retains the achievable core status and current loading status of the source index item. When the next round of retransmission status control sequence is generated, the source index item will continue to be treated as a waiting receipt object, so as to avoid the station repeatedly compressing and loading the same data segment when the receipt is missing.

[0113] In this implementation, the individual positions of the Blackwing Kite optimization algorithm are kept consistent with the corresponding positions of the index entries in the core state storage area. The optimization results are not saved separately as external sorting results, but are written into the internal control state that the STAR module can read in the next round. As a result, the satellite short message capacity, station energy and cloud acknowledgment jointly change the reading range of the retransmission queue and the receiving path of the acknowledgment core state.

[0114] In this embodiment, S7 specifically includes: The retransmission status control sequence is read from the communication receive buffer by the station controller. The control sequence includes the source index item, the corresponding position of the index item, the control sequence number, and the loading status of this round. The station controller first verifies the station number and the sequence verification value. Control sequences that fail the verification are written into the receive error record and do not enter the retransmission queue for matching.

[0115] The index entry is used to match the queue entry in the retransmission queue. The retransmission queue entry stores the source index entry, the acquisition time range, the data length, and the reading status. The station controller uses the same source index entry and overlapping acquisition time range as the matching basis. After a successful match, it reads the corresponding data segment to be retransmitted.

[0116] The local transmission record is used to determine whether the data segment to be retransmitted has been sent. The local transmission record saves the source index item, packet order, transmission time, transmission result and receipt association status. When the source index item and packet order both exist and the transmission result is successful, the corresponding data segment to be retransmitted is removed from the loading objects in this round.

[0117] The current supplementary data segment consists of unsent supplementary data segments. When multiple data segments enter the same round of transmission, they are first arranged in ascending order of the acquisition timestamp. If the acquisition timestamps are the same, they are then arranged in order of the starting value of the local storage write address. The arrangement result is used as the input order for compression encoding.

[0118] The compression encoding process handles the running data content in the current retransmitted data segment. Before encoding, the source index item and the collection time range are retained as auxiliary index information. Meteorological data and power supply status data are first differentially compressed according to the difference between adjacent sampling points. Continuous sampling points with a difference of 0 are written into the repetition count, and sampling points with a difference of non-zero are written with the difference and the sampling point position. Station-end running status data and communication status data are first converted into status codes, and then consecutive identical status codes are written into the repetition count.

[0119] In one implementation, the current retransmission data segment can be compressed into a compressed window with 30 sampling points. Each window first writes the window start acquisition timing and then writes the differential compression result. The compressed retransmission payload generated after compression still carries the source index item and acquisition timing, ensuring that it can return to the index range of the data to be retransmitted index table after decompression in the cloud.

[0120] The upper limit of satellite short message payload comes from the link configuration of the communication module. Before packet encapsulation, the number of bytes of compressed retransmission payload is calculated. If the number of bytes does not exceed the upper limit of payload, it is encapsulated into a single retransmission satellite short message. If the number of bytes exceeds the upper limit of payload, it is divided according to the remaining number of bytes after deducting the message header, packet order and check field from the upper limit of payload.

[0121] In one implementation, the payload of a single satellite short message packet can be 116 bytes. When the message header, source index, packet order, and check field together occupy 16 bytes, each packet can carry a maximum of 100 bytes of compressed data. The number of packets is obtained by dividing the number of compressed data bytes by 100 bytes and then rounding up.

[0122] The packet order starts from 1 and increments. All packets under the same source index are written to the same source index and the total number of packets. The last packet is written with an end marker. The cloud uses this to determine whether the same source index has received the complete set of packets.

[0123] Before transmitting the satellite short message, the station controller calculates the CRC check value for the message header, source index, packet order, and compressed payload. In one embodiment, the CRC check can use a 16-bit cyclic redundancy check, with a polynomial of 0x1021 and an initial value of 0xFFFF. The check value is obtained by shifting and XORing byte by byte and then writing it to the end of the message. After the satellite short message communication link returns the transmission acceptance result, the transmission acceptance result is written to the local transmission record.

[0124] When the satellite short message communication link does not return the transmission acceptance result, the station controller will write the corresponding sub-packet into the pending confirmation status. The pending confirmation status is not the same as the transmission success. When parsing the retransmission status control sequence in the next round, it is still necessary to combine the cloud receipt result to determine whether to reload.

[0125] After the local transmission record is written, the transmission results, packet order and collection time range under the same source index are retained as queryable records. These records are used both for the next round of station-side removal of transmitted data segments and for determining whether a packet has entered the satellite short message communication link when the cloud receipt has not arrived.

[0126] In this embodiment, S8 specifically includes: After receiving the retransmitted satellite short message, the cloud platform writes it into the retransmission receiving cache. The source index item locates the index range in the data index table to be retransmitted, and the packet order indicates the arrangement position of each packet under the same source index item. The receiving cache establishes the retransmission receiving record based on the source index item and the packet order.

[0127] When writing the retransmission receive record, record the message receive timestamp, source index item, packet order, total number of packets, payload length, and check field. The receive timestamp is only used to distinguish the order of arrival, while the source index item and packet order are used to reassemble the retransmission data. The receive timestamp does not participate in index range matching.

[0128] The decapsulation process first reads the source index and packet order in the message header, then reads the payload length and check field. The cloud platform uses the same CRC calculation method as the station to recalculate the check value for the message header, source index, packet order, and compressed payload. When the payload length is equal to the actual payload bytes and the recalculated check value is consistent with the check field, a valid reception status is written. If any result is inconsistent, a check failure status is written.

[0129] The reorganization of sub-packages under the same source index is performed in ascending order of sub-packages. The total number of sub-packages is used to determine whether a complete set of sub-packages has been collected. If any sub-package is missing, decompression and restoration will not be performed. Missing sub-packages will be written to the exception record of the unclosed index range.

[0130] After the retransmission received record passes the integrity check, it is decompressed and restored. The decompression and restoration reads the initial acquisition timing, differential compression value and repetition count of the window. First, the running data increment of each sampling point is restored, and then the data segments are arranged according to the acquisition timing.

[0131] When the data segment to be retransmitted and restored is matched with the index table of data to be retransmitted, the cloud platform first uses the source index item to locate the index range, and then compares whether the collection time sequence of the data segment to be retransmitted and restored falls within the index range. The part with continuous collection time sequence and valid verification result is written into the retransmission completion status.

[0132] The writing range for the retransmission completion status is determined by the start and end positions of the data acquisition timing that have passed verification. The number of sampling points between the start and end positions is equal to the number of sampling points in the corresponding range in the data to be retransmitted index table, and the difference between the sampling timestamps is equal to the acquisition interval. When these sampling timestamps are all equal to the acquisition interval, the index range is recorded as retransmission completion. If the number is insufficient or the difference between adjacent sampling timestamps is not equal to the acquisition interval, only the sub-ranges that have passed verification consecutively are written to the retransmission completion status.

[0133] The incomplete matching index range is generated by four types of results: missing sub-packets, verification failure, source index items cannot be matched, and collection time sequence is discontinuous. The cloud platform retains the incomplete part as an unclosed index range and does not delete it from the data to be transmitted index table.

[0134] The cloud receipt result is generated by the retransmission completion status and the unclosed index range. The retransmission completion status is written with a completion flag, the unclosed index range is written with an unclosed flag, and the index range waiting for packet splitting or waiting for verification completion is written with a waiting flag. All three types of flags are bound to the source index item.

[0135] When the cloud receipt result is written to the improved SOFTS model, the core state storage area first locates the corresponding position of the index item through the source index item, and then writes the completion mark, unclosed mark or waiting mark to the receipt record at that position, and binds it with the receivable core state at the same position.

[0136] When updating the state transition flags in the core state storage area, the completion flag corresponds to the completion class code, the unclosed flag corresponds to the unclosed class code, and the waiting flag corresponds to the waiting class code. The code is written into the state transition flag tensor in B×I×3 format, and the I-dimensional position is consistent with the source index item.

[0137] When writing back an unclosed index range to the data index table to be retransmitted, the cloud platform retains the original source index entries and updates the start and end positions of the unclosed range. Completed continuous sub-ranges are written to the retransmission completion status. The start and end positions of the unclosed range serve as the input for the next round of intercepting the data segment to be retransmitted and updating the retransmission queue at the station end.

[0138] After the cloud receipt is sent to the site, the site updates the local sending record with the source index item and the packet order. Packets that have been retransmitted will no longer enter the retransmission queue, and data segments corresponding to unclosed index ranges will continue to be retained. Data segments that are waiting to be marked will remain in a pending confirmation state.

[0139] This implementation process ensures that the retransmission of satellite short messages forms a closed path from reception, verification, reassembly, index matching to model writeback. The cloud-based acknowledgment not only confirms the communication reception status but also changes the state transition flag in the core state storage area and reconnects the unclosed index range to the next round of retransmission queue update.

[0140] Example 1: To verify the feasibility of this invention in practice, it was applied to a remote data transmission scenario of an automatic rainfall monitoring station in an uninhabited area. The station was deployed using a drone-based system with automatic leveling after deployment. The station was equipped with a tipping bucket rain sensor, local FLASH storage, a lithium battery power supply unit, and a satellite short message communication module. There was no mains power or public network available on-site, and rainfall events were prone to link blockage and low-power sleep switching. Traditional timed reporting methods generate numerous short messages during sudden increases in rainfall intensity. If some messages fail verification, the cloud can only see a broken receiving sequence, making it difficult to determine the specific missing data segments in the local acquisition sequence. This easily leads to problems such as duplicate retransmissions and delayed transmission of critical rainfall segments.

[0141] During implementation, the station recorded rainfall increments, station operation status, and communication status at 5-minute intervals. The bucket resolution was configured at 0.2mm. After each data entry was written to the local FLASH memory, a local time-series index and the address range were simultaneously registered. During a single weak link backhaul, the station generated 463 backhaul payloads corresponding to the collected data. The cloud actually received 428 payload records, of which 19 failed integrity verification. The cloud then organized the data according to the local time-series index, merging the missing and failed verification entries into 7 unclosed backhaul segments and issuing a data index table to be retransmitted. The station read the corresponding address segments from the FLASH memory according to the index entries, restored the collection order, and formed a set of data segments to be retransmitted, avoiding the retransmission of messages that had already passed verification.

[0142] The improved SOFTS model writes the seven index items into the corresponding positions of the core state storage area in the STAR module. The random pooling layer only aggregates the hidden representations under the same source index item to generate a retrievable core state. After the cloud retrievable result is written, the state transition flag distinguishes between index items that have been retransmitted, those that are not yet closed, and those waiting to be reloaded. In this embodiment, the satellite short message single packet payload is configured as 116B. When the station-end battery voltage is in the range of 3.71V to 3.84V, the Black-winged Kite optimization algorithm prioritizes index items that have a greater impact on the continuity of rainfall changes, generating 14 retransmitted satellite short messages. The station compresses and divides the packets before sending them in batches. To verify the feasibility of this invention in implementation, the implementers selected 8640 station-end collection records for comparative testing. The statistical results are shown in Table 1 below.

[0143] Table 1. Comparison of Closed-Loop Effects of Satellite Short Message Retransmission from Automatic Weather Stations in Uninhabited Areas

[0144] As can be seen from the data in Table 1 above, the method of the present invention outperforms the timed reporting combined with sequential retransmission method and the rule-based retransmission method based on cloud-based missing data index in terms of missing segment location accuracy, effective retransmission completion rate, duplicate retransmission ratio, average retransmission completion delay, number of short messages, communication energy consumption, and rainfall process restoration integrity rate. The timed reporting combined with sequential retransmission method has a missing segment location accuracy of 79.4%, an effective retransmission completion rate of 77.8%, and a duplicate retransmission ratio of 21.6%, indicating that this method mainly relies on fixed-order retransmission, making it difficult to accurately distinguish between truly missing data segments and effectively received data segments, easily leading to duplicate transmissions and retransmission delays. The rule-based retransmission method based on cloud-based missing data index improves the missing segment location accuracy to 88.9% and the effective retransmission completion rate to 86.5% through cloud-based missing data location. However, because its retransmission order still mainly relies on rule judgment, it lacks sufficient joint utilization of station-end energy status, short message payload constraints, and receipt status, resulting in a duplicate retransmission ratio of 11.3% and an average retransmission completion delay of 31.5 minutes. The method of this invention achieves a missing segment location accuracy of 96.2%, an effective retransmission completion rate of 94.7%, reduces the proportion of repeated retransmissions to 4.8%, shortens the average retransmission completion delay to 18.6 minutes, reduces the number of short messages per 1000 collection records to 45, reduces communication energy consumption to 0.41 Wh, and improves the completeness rate of rainfall process restoration to 97.3%. This demonstrates that the present invention can more accurately locate missing segments and complete retransmission more efficiently under weak link and low energy consumption constraints, while reducing the occupation of invalid communication resources.

[0145] This embodiment utilizes the correspondence between the local time-series index at the station and the complete status of messages in the cloud. It first converts missing messages, failed verifications, and broken sequences into a data index table to be retransmitted. Then, the station extracts the data segments to be retransmitted from its local FLASH memory based on the index entries, avoiding the problem of blindly retransmitting in the receiving order in traditional methods. Based on this, the SOFTS model is improved by rewriting the global core representation of the STAR module to the corresponding position of the index entries, enabling each source index entry to form a retrievable core state. After the cloud's retrievable result is written to the core state storage area, a state transition flag is generated, allowing states such as retransmission complete, incomplete, and waiting to be reloaded to enter the model's internal calculation process. Subsequently, the Blackwing Kite optimization algorithm generates a retransmission state control sequence based on the retrievable core state, state transition flag, satellite short message payload limit, and remaining energy at the station. The station then compresses, packages, and retransmits the data in batches according to this sequence. After receiving the retransmitted satellite short messages, the cloud performs verification, decompression, and reassembly. The unclosed index range is then written back to the data index table to be retransmitted, thus forming a closed-loop process of data retrieval at the station, cloud verification, model status updates, and adjustment of the next round of retransmission queues. This embodiment demonstrates that the present invention can improve the continuity and reliability of data backhaul from automatic weather stations in uninhabited areas under conditions of no public network, weak satellite links, and limited energy.

[0146] The above are merely preferred embodiments 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 method for remote data transmission from an unmanned automatic weather station based on cloud-network collaboration, characterized in that, Includes the following steps: S1. Automatic weather stations in uninhabited areas collect operational data, write it to local storage according to the collection sequence, and organize it into station-end transmission objects. S2. The cloud platform receives satellite short messages from the target object at the receiving station, performs decoding, verification, and continuity processing to form a complete message status; S3. Read the complete status of the message in the cloud, perform index merging on unclosed return segments, and establish an index table of data to be retransmitted. S4. The station receives the index table of data to be retransmitted, extracts the data segment to be retransmitted from the index item from the local storage, writes it into the retransmission queue, and forms a set of data segments to be retransmitted. S5. Improve the SOFTS model to read the index table of data to be retransmitted and the set of data segments to be retransmitted, rewrite the global core representation of the STAR module to the corresponding position of the index item, form an achievable core state, and generate a state transition flag based on the cloud achievable result. S6. The Black-winged Kite optimization algorithm runs in the cloud. Based on the retrievable core state and state transition marker, it generates a retransmission state control sequence under the constraints of satellite short message capacity and station energy, and writes it in reverse to the STAR module. S7. The station loads the data segments to be retransmitted from the retransmission queue according to the retransmission status control sequence, and transmits them back in batches via satellite short messages after compression and packetization. S8 verifies the satellite short message transmitted back to the cloud, generates a cloud receipt result and updates the data index table to be retransmitted, and feeds the cloud receipt result back to the improved SOFTS model to control the next round of retransmission queue updates.

2. The method for remote data transmission from an unmanned automatic weather station based on cloud-network collaboration according to claim 1, characterized in that, S1 specifically includes: S11. The station controller of the uninhabited automatic weather station calls the acquisition process to collect operational data and writes the acquisition sequence according to the station clock to form operational data with acquisition sequence. S12. The station controller sequentially writes the operating data with the acquisition timing sequence into the local storage according to the acquisition timing sequence, and records the write address range to form a local timing index. S13. The station controller reads the local timing index and merges the running data that has been written but not yet entered into the backhaul record into the station backhaul record. S14. The station controller binds the station backhaul records to the local time-series index, retains the write address range in the local storage, and organizes them into station backhaul objects.

3. The method for remote data transmission from an unmanned automatic weather station based on cloud-network collaboration according to claim 1, characterized in that, S2 specifically includes: S21. The satellite short message corresponding to the object returned by the cloud platform receiving station is written into the receiving buffer according to the receiving time sequence to form a satellite short message receiving record. S22. The cloud platform performs protocol decapsulation and payload decoding on the satellite short message reception record to restore the station-end return record and the bound local time sequence index. S23. The cloud platform performs verification and check on the station-end return records based on the load decoding results, and writes the verification and check results into the corresponding local time series index. S24. The cloud platform organizes the reception continuity of the station's backhaul records according to the local time sequence index to form a complete message status.

4. The method for remote data transmission from an unmanned automatic weather station based on cloud-network collaboration according to claim 1, characterized in that, S3 specifically includes: S31. The cloud platform reads the complete status of the message, checks the station's back transmission record according to the local time sequence index, and marks the discontinuous or unverified local time sequence index as a reception breakpoint. S32. The cloud platform reads adjacent received indices based on the receiving breakpoint, determines the start and end indices of unclosed back transmission segments, and forms a missing segment index. S33. The cloud platform merges adjacent or overlapping missing test segment indexes and removes the index range that has been covered by the verified station-end return records to form a merged index. S34. The cloud platform will associate the merged index with the back-up object from the site and create an index table of data to be retransmitted.

5. A method for remote data transmission from an unmanned automatic weather station based on cloud-network collaboration as described in claim 1, characterized in that, S4 specifically includes: S41. The station controller parses the index table of data to be retransmitted, reads the local time-series index corresponding to the index item, and forms an index query sequence. S42. The station controller retrieves the local time-series index based on the index query sequence, obtains the write address range, and reads the running data from the local storage. S43. The station controller restores the acquisition order of the read running data according to the local time sequence index and binds it with the source index item to form a data segment to be transmitted. S44. The station controller writes the data segments to be retransmitted into the retransmission queue according to the source index item, forming a set of data segments to be retransmitted.

6. The method for remote data transmission from an unmanned automatic weather station based on cloud-network collaboration according to claim 1, characterized in that, The improved SOFTS model includes an invertible instance normalization layer, a sequence embedding layer, a multi-layer STAR module, and a linear predictor. The STAR module includes a first multilayer perceptron, a random pooling layer, a repetition and splicing operation layer, a second multilayer perceptron, and residual connections; The STAR module has a core state storage area that is connected to the index entries of the index table for the data to be retransmitted. The core state storage area holds the retrievable core state, and the state transition flag is bound to the retrievable core state.

7. A method for remote data transmission from an automatic weather station in an unmanned area based on cloud-network collaboration as described in claim 1, characterized in that, S5 specifically includes: S51. The sequence embedding layer reads the set of data segments to be retransmitted, maps the data segments to be retransmitted into sequence representations, and writes the source index item into the index position of the sequence representation. S52. The first multilayer perceptron reads the sequence representation and generates a hidden representation, and the hidden representation maintains the correspondence between the source index item and the hidden representation channel; S53. The core status storage area establishes the connection relationship between the corresponding position of the index item and the hidden representation channel according to the data to be retransmitted index table; S54. The random pooling layer reads the hidden representation channel according to the source index item, writes the pooling result of the same source hidden representation to the corresponding position of the index item, and forms an achievable core state. S55. When the source index item does not hit the hidden representation channel in this round, the core state storage area retains the written executable core state along the corresponding position of the index item, and the sequence representation corresponding to the source index item that has not formed an executable core state is retained along the residual connection. S56. The repeat and splice operation layer reads the retrievable core state according to the source index item and repeatedly splices the retrievable core state to the same source sequence representation. S57. The second multilayer perceptron fuses the homologous sequence representations after repeated splicing, and the residual connection writes the fusion result back to the sequence representations of the same source index items; S58. The core state storage area writes the cloud receipt results to the corresponding position of the index item and binds them with the receivable core state to generate a state transition marker.

8. A method for remote data transmission from an automatic weather station in an uninhabited area based on cloud-network collaboration, as described in claim 1, is characterized in that... S6 specifically includes: S61. The cloud platform runs the Blackwing Kite optimization algorithm, reads the retrievable core state and state transition marker, and encodes the position corresponding to the index item as the individual position. S62. The fitness of an individual location is calculated using the completeness of the index item retransmission corresponding to the retrievable core status as the benefit term, and the length of the data segment to be retransmitted, the upper limit of the satellite short message payload, and the remaining energy at the station as the constraint terms. S63. Execute the Blackwing Kite optimization algorithm to update attack behavior and migration behavior according to fitness, and adjust the control order of index items corresponding to individual positions; S64. When the updated individual position exceeds the index range of the data to be retransmitted index table, the individual position will be rolled back to the corresponding position of the adjacent index item that has not been retransmitted. S65. Generate a supplementary transmission state control sequence according to the updated individual position, and write the supplementary transmission state control sequence into the core state storage area of ​​the STAR module.

9. A method for remote data transmission from an automatic weather station in an unmanned area based on cloud-network collaboration as described in claim 1, characterized in that, Specifically, S7 includes: S71. The station controller parses the retransmission status control sequence, matches the retransmission queue according to the corresponding position of the index item, reads the data segment to be retransmitted, and removes the sent data segment according to the source index item and the local sending record to form the current retransmission data segment. S72. The current retransmission data segment is arranged according to the acquisition time sequence and then compressed and encoded. The source index item and the acquisition time sequence are retained with the compression result to form a compressed retransmission payload. S73. The compressed retransmission payload is packaged and encapsulated according to the upper limit of the satellite short message payload. The packet order is established to correspond with the source index item to form a retransmission satellite short message. S74. The satellite short message is sent via the satellite short message communication link, and the sending result, source index item and packet order are written into the local sending record.

10. A method for remote data transmission from an unmanned automatic weather station based on cloud-network collaboration as described in claim 1, characterized in that, S8 specifically includes: S81. The cloud platform receives the retransmitted satellite short message, performs decapsulation according to the source index item and packet order, and forms a retransmission reception record. S82. After integrity verification and decompression, the retransmission received record is reassembled into a retransmission restored data segment according to the source index item and the acquisition time sequence. S83. The retransmission and restoration data segment is matched with the index range of the data index table to be retransmitted. The index range that has been matched is written to the retransmission completion status, and the index range that has not been matched is retained as an unclosed index range. S84. Generate cloud receipt results from the completed retransmission status and unclosed index range. Write the cloud receipt results into the core state storage area of ​​the improved SOFTS model and bind them to the receivable core state. S85. The core state storage area updates the state transition flag based on the cloud receipt results, and writes the unclosed index range back to the data to be retransmitted index table, forming the basis for updating the next round of retransmission queue.