Data transmission method and system based on semantic compression and dynamic decision control

Abstract

The invention relates to the technical field of data transmission, and provides a data transmission method and system based on semantic compression and dynamic decision control, high-dimensional original time series data is extracted as semantic features and compressed into a hidden state, only a small amount of representative high-value semantic information is transmitted, and the data transmission efficiency is improved. Meanwhile, a sending strategy is dynamically adjusted in combination with the data change density and the network condition in the transmission process, real-time feedback and strategy optimization are achieved through analysis of a receiving end on the link state, a closed-loop semantic transmission mechanism with the self-adaptive regulation and control capacity is constructed, and efficient transmission of data information is achieved.

Description

Data transmission method and system based on semantic compression and dynamic decision control Technical Field

[0001] This disclosure relates to the field of data transmission technology, specifically to a data transmission method and system based on semantic compression and dynamic decision control. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Remote energy storage power stations and new energy power plants are gradually adopting technologies such as cloud monitoring and digital twins to achieve remote sensing, intelligent modeling, and predictive maintenance of the operating status of critical equipment. During this process, numerous sensors and controllers on-site continuously collect high-frequency, multi-source time-series data such as voltage, current, temperature, pressure, and vibration, and upload it to the cloud in real time for equipment health assessment, fault precursor identification, and operation and maintenance strategy optimization. This type of industrial time-series data is not only massive in volume but also changes frequently and has high information density, serving as a core information carrier supporting safe system operation and intelligent decision-making. Especially in high-risk systems involving battery energy storage, any slight abnormal fluctuation may indicate potential thermal runaway or electrical faults. Such faults are often characterized by their suddenness and rapid evolution; if the response is not timely, they can easily lead to cell thermal propagation, equipment explosions, or even large-scale power outages. Therefore, only by synchronizing critical time-series data from the edge to the cloud with high fidelity and low latency can potential risks be effectively identified and effective intervention strategies formulated.

[0004] Traditional reliable communication protocols such as Transmission Control Protocol (TCP), which are widely used today, have certain advantages in ensuring data order and correctness. However, in practical industrial applications, several key technical bottlenecks have gradually emerged. First, these protocols employ a connection-based, fully reliable transmission mechanism, requiring all data packets to be successfully acknowledged in sequence before subsequent data can be sent. This leads to prominent head-of-line blocking problems, especially in scenarios with network fluctuations, high packet loss rates, or unstable link quality, where retransmission of a single data packet can delay the entire data flow. Second, while TCP's congestion control mechanism can maintain overall network stability, it lacks a mechanism to differentiate between high-priority data and cannot perform weighted scheduling of critical information during emergencies, resulting in data with higher diagnostic value not being transmitted in a timely manner. Furthermore, this protocol only focuses on the formal structure of data transmission (such as packet size and order) and cannot perceive the semantic features and importance of the data content. This makes it impossible for the system to intelligently filter and optimize the transmission of time-series data in the event of sudden anomalies.

[0005] Taking remote energy storage systems as an example, when the edge monitoring system detects early signs of thermal runaway, such as a sudden rise in cell temperature or voltage fluctuations, the related multi-channel sampling data and historical evolution trajectories will surge rapidly, forming a sudden surge in time-series data. If traditional reliable protocols are still used for transmission at this time, not only will the response speed not keep up with the pace of fault evolution, but the delayed uploading of core indicators may also prevent the cloud system from constructing a complete picture of the fault, identifying the risk level, and generating effective strategies within the golden time window. This could easily lead to missed opportunities for safe handling, ultimately inducing serious safety accidents such as the spread of thermal runaway and the explosion of energy storage units, resulting in equipment damage and significant economic losses. Summary of the Invention

[0006] To address the aforementioned issues, this disclosure proposes a data transmission method and system based on semantic compression and dynamic decision control. It extracts semantic features from high-dimensional raw time-series data and compresses them into hidden states, transmitting only a small amount of representative, high-value semantic information. Simultaneously, during transmission, the transmission strategy is dynamically adjusted based on data change density and network conditions. Real-time feedback and strategy optimization are achieved through link state analysis at the receiving end, constructing a closed-loop semantic transmission mechanism with adaptive control capabilities to realize efficient data transmission.

[0007] To achieve the above objectives, this disclosure adopts the following technical solution: One or more embodiments provide a data transmission method based on semantic compression and dynamic decision control, including the following steps: The data sending end adaptively segments the acquired multi-dimensional data to be transmitted and then fuses multi-scale spatiotemporal features to generate a first feature sequence; the first feature sequence is subjected to global semantic compression to generate a first hidden state sequence; based on the semantic change rate of the first hidden state sequence and the network channel quality, combined with a dynamic decision function, an optimal transmission strategy instruction is generated, and the first hidden state sequence is encapsulated into an application layer data stream for data transmission; The data receiving end sequentially performs streaming parsing and multi-dimensional cross-validation on the received application layer data stream; if the validation passes, a second hidden state sequence is generated; if the validation fails, an abnormal hidden state sequence is generated, and guided semantic repair is performed to generate the second hidden state sequence; based on the second hidden state sequence, the end-to-end transmission delay, semantic reconstruction confidence, and posterior anomaly score are calculated to identify the link operation status of data transmission and generate a corresponding data transmission strategy, which is then transmitted to the data sending end to achieve adaptive updating of the transmission strategy.

[0008] One or more embodiments provide a data transmission system based on semantic compression and dynamic decision control, including a data transmission end and a data receiving end. The data transmission end, used to process the received data and transmit it to the data receiving end, is configured to perform the following process: the data transmission end adaptively segments the acquired multi-dimensional data to be transmitted and fuses multi-scale spatiotemporal features to generate a first feature sequence; the first feature sequence is subjected to global semantic compression to generate a first hidden state sequence; based on the semantic change rate of the first hidden state sequence and the network channel quality, combined with a dynamic decision function, an optimal transmission strategy instruction is generated, and the first hidden state sequence is encapsulated into an application layer data stream for data transmission. Further, the data receiving end is configured to perform the following process: the data receiving end sequentially performs streaming parsing and multi-dimensional cross-validation on the received application layer data stream; if the validation passes, a second hidden state sequence is generated; if the validation fails, an abnormal hidden state sequence is generated; guided semantic repair is performed to generate the second hidden state sequence; based on the second hidden state sequence, the end-to-end transmission delay, semantic reconstruction confidence, and posterior anomaly score are calculated to identify the link operation status of data transmission and generate a corresponding data transmission strategy to be transmitted to the data transmission end to achieve adaptive updating of the transmission strategy.

[0009] Compared with existing technologies, the beneficial effects of this disclosure are as follows: By converting the raw sampling points in multi-dimensional industrial time-series data into a representative semantic latent state sequence, this invention significantly reduces the amount of data required for edge-to-cloud transmission. Taking a typical remote energy storage system as an example, the raw data may contain hundreds to thousands of data points in a single sampling period. However, after multi-scale spatiotemporal feature fusion and semantic compression processing, only a set of latent state vectors abstractly expressing its evolutionary trend and abnormal behavior needs to be transmitted, reducing the data volume by an order of magnitude or even more. Compared with the traditional mode of transmitting raw data packet by packet, this method effectively avoids the transmission bottleneck under big data surges and greatly improves the response speed of critical information in emergencies.

[0010] Meanwhile, by integrating semantic change rate and link status to construct a dynamic transmission strategy, high-semantic-value segments are prioritized for transmission during sudden increases in data density or network fluctuations, ensuring real-time delivery of core information. During stable periods, data frequency is proactively reduced, achieving intelligent scheduling and saving of bandwidth resources. Furthermore, the receiver's link status awareness and strategy feedback mechanism can adjust subsequent transmission strategies in real time based on indicators such as semantic reconstruction quality and anomaly distribution characteristics, constructing an efficient, stable, and self-learning data transmission closed-loop system. These improvements effectively solve key problems in existing technologies such as "severe redundant transmission of high-dimensional raw data, high network resource consumption, and delayed response to sudden risks," providing solid support for remote operation and maintenance security in high-risk new energy scenarios.

[0011] The advantages of this disclosure, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0012] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute a limitation thereof.

[0013] Figure 1 is a flowchart of the data transmission system method based on semantic compression and dynamic decision control according to Embodiment 1 of this disclosure; Figure 2 is a system architecture diagram of the data transmission system based on semantic compression and dynamic decision control according to Embodiment 2 of this disclosure. Detailed Implementation

[0014] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0016] It should be noted that the terminology used herein is for descriptive purposes only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments in this disclosure can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0017] Example 1: In one or more of the technical solutions disclosed in the embodiments, as shown in Figure 1, a data transmission method based on semantic compression and dynamic decision control includes the following steps: S1: The data sending end adaptively segments the acquired multi-dimensional data to be transmitted and then fuses multi-scale spatiotemporal features to generate a first feature sequence. Global semantic compression is then performed on the first feature sequence to generate a first hidden state sequence. S2: Based on the semantic change rate of the first hidden state sequence and the network channel quality, an optimal transmission strategy instruction is generated by combining a dynamic decision function. The first hidden state sequence is then encapsulated into an application layer data stream for data transmission. S3: The data receiving end sequentially performs streaming parsing and multi-dimensional cross-validation on the received application layer data stream. If the verification passes, a second hidden state sequence is generated; if the verification fails, an abnormal hidden state sequence is generated. Guided semantic repair is then performed to generate the second hidden state sequence. S4: Based on the second hidden state sequence, end-to-end transmission delay, semantic reconstruction confidence, and posterior anomaly score are calculated to identify the data transmission link's operational status and generate a corresponding data transmission strategy, which is then transmitted to the data sending end to achieve adaptive updates of the transmission strategy. This implementation first adaptively segments the multi-dimensional time-series data (such as temperature, current, and voltage) collected from the industrial site at the data sending end to adapt to differences in sampling density and time-series characteristics. Then, spatial correlation and temporal evolution trends are integrated to construct a unified first feature sequence. By introducing a global semantic compression mechanism, this sequence is further encoded into a first hidden state sequence with abstract expressive capabilities, thereby completing the compressed expression of key semantic information without transmitting the original data. In terms of transmission scheduling, the system combines the data change rate reflected in the hidden state sequence with the real-time channel state, using a dynamic decision function to output the optimal transmission strategy, achieving weighted priority scheduling of high-density data bursts or abnormal segments, and improving the transmission efficiency of key semantic data. At the receiving end, the system restores semantic information through streaming decoding and multidimensional cross-validation. If content is missing or erroneous, a guided semantic repair module is triggered to complete the semantics using context and prior model knowledge, ultimately restoring the second hidden state sequence. The system further evaluates end-to-end latency, semantic reconstruction confidence, and anomaly distribution based on this sequence, identifies the current link operating status, dynamically adjusts the transmission strategy, and constructs an adaptive closed-loop feedback mechanism.

[0018] This implementation significantly reduces the amount of data required for edge-to-cloud transmission by converting raw sampling points in multidimensional industrial time-series data into representative semantic latent state sequences. Taking a typical remote energy storage system as an example, the raw data may contain hundreds to thousands of data points in a single sampling period. However, after multi-scale spatiotemporal feature fusion and semantic compression, only a set of latent state vectors abstractly expressing its evolutionary trend and abnormal behavior needs to be transmitted, reducing the data volume by an order of magnitude or even more. Compared to the traditional method of transmitting raw data packet by packet, this method effectively avoids transmission bottlenecks under large data surges and greatly improves the response speed of critical information in emergencies.

[0019] Meanwhile, by integrating semantic change rate and link status to construct a dynamic transmission strategy, high-semantic-value segments are prioritized for transmission during sudden increases in data density or network fluctuations, ensuring real-time delivery of core information. During stable periods, data frequency is proactively reduced, achieving intelligent scheduling and saving of bandwidth resources. Furthermore, the receiver's link status awareness and strategy feedback mechanism can adjust subsequent transmission strategies in real time based on indicators such as semantic reconstruction quality and anomaly distribution characteristics, constructing an efficient, stable, and self-learning data transmission closed-loop system. These improvements effectively solve key problems in existing technologies such as "severe redundant transmission of high-dimensional raw data, high network resource consumption, and delayed response to sudden risks," providing solid support for remote operation and maintenance security in high-risk new energy scenarios.

[0020] Step S1 in this embodiment targets the raw multidimensional data vector collected in real time by the Energy Management System (EMS). Before entering the unstable transmission channel, it is preprocessed into a structured, high-information-density first hidden state sequence Z with inherent fault tolerance, providing a basis for proactive data reconstruction on the cloud receiving side and avoiding the delay problem caused by traditional retransmission.

[0021] In step S1, based on the data stream fluctuation index For the original multidimensional data vector Adaptive segmentation is performed to generate a first data sequence S. Based on the first data sequence S, multi-scale spatiotemporal features are fused to generate a first feature sequence. The first feature sequence is processed using a lightweight Transformer encoder. Perform global semantic compression to generate the first hidden state sequence Z; in step S1, the method for generating the first hidden state sequence Z includes the following steps: Step S11, based on the original multidimensional data vector Real-time calculation of data stream fluctuation indicators Based on the data stream fluctuation index The original multidimensional data vector is subjected to a composite logical condition consisting of two preset thresholds. Adaptive segmentation is performed to generate the first data sequence S. In this implementation, to address the information fragmentation or data redundancy problems inherent in the fixed-window data segmentation method, the collected original multidimensional data vectors are processed... Real-time dynamic segmentation is performed to divide the unbounded data vector into a first data sequence S with variable length but complete information. This ensures that each output first data sequence S is an independent analysis unit with high context relevance, providing the input with the optimal information value for subsequent semantic compression.

[0022] In the specific implementation process, this step adopts an adaptive serialization buffering method based on the dynamic characteristics of the data stream. This method is deployed in the computing unit at the edge and evaluates the dynamic salience of the data stream in real time, i.e., the data stream fluctuation index. This approach determines the optimal timing for sequence segmentation using an event-driven method, replacing the traditional fixed buffering strategy. The data stream fluctuation index... It is a quantitative parameter used to characterize the degree of change in the operating status of the energy management system (EMS) at any given time.

[0023] Specifically, based on data stream fluctuation indicators For the original multidimensional data vector The method for adaptive segmentation to generate the first data sequence S includes the following steps: Step S111: Construct a data flow fluctuation index by taking the moving average of the Euclidean distance increments of the obtained multidimensional data vectors. To ensure computational efficiency on resource-constrained edge devices, data flow fluctuation metrics are used. It is defined as the moving average of the recent Euclidean distance increments of the data vector, as shown in the following formula: ;in, This represents a data stream fluctuation metric, which is a scalar value calculated at timestamp t. This represents the moving average smoothing window, a preset integer used to eliminate the impact of noise or minor environmental disturbances on data stream fluctuations. Interference; This represents the time iteration index for the summation operation, and its value ranges from... arrive consecutive positive integers; Indicates the current timestamp index; The original multidimensional data vector represents the D-dimensional set of sensor readings collected in real time from various sensors of the Energy Management System (EMS) at discrete time point i. The original multidimensional data vector representing the previous time step is the immediately preceding discrete time step relative to discrete time step i. The corresponding original multidimensional data vector; The vector 2-norm operator is used to calculate the Euclidean distance between two continuous multidimensional data vectors in D-dimensional space, and is used to comprehensively measure the magnitude of change in the overall state.

[0024] Step S112: Set the quiet threshold and alertness threshold When the data flow fluctuation index is greater than the alert threshold, all data vectors in the data cache space at the current time (tend) are divided into a first data sequence S; otherwise, when the data flow fluctuation index is less than the tranquility threshold... Once the data meets the set data sequence length, it is then divided into a first data sequence S. Specifically, in this embodiment, to achieve precise segmentation control, two key thresholds are preset: a lower quiet threshold... and a higher alertness threshold Because the data from the device in a fault state differs significantly from that in a normal state, there is a certain interval between the two thresholds, i.e. Meanwhile, let the number of data points currently collected in the buffer pool be... The formation of the first data sequence S is triggered by the following composite logical condition, and the formula for triggering the segmentation is as follows: ;in, The trigger signal is a Boolean logic variable used to decide whether to perform the operation of integrating the data in the buffer pool into the first data sequence S. The alertness threshold is a preset floating-point hyperparameter whose value is set based on the specific application scenario and historical data analysis, and must be greater than the tranquility threshold. This is used to determine whether the system state has undergone drastic dynamic changes; Represents the logical OR operator, used to integrate event-driven triggering paths and steady-state upper limit triggering paths; The quiescent threshold is a preset floating-point hyperparameter whose value is set according to the normal fluctuation range of the system under stable operating conditions. It is used to confirm whether the system is in a stable operating state. This represents the logical AND operator; The current sequence length is represented by an integer counter, whose value ranges from 0 to... An integer, used to determine whether the upper limit of the length of the first data sequence S has been reached; The maximum sequence length is a preset integer parameter used as the length benchmark for the steady-state upper limit trigger path.

[0025] when The event-driven triggering path is defined as a high-priority triggering path. In this case, when the Energy Management System (EMS) experiences transient changes, such as sudden load changes leading to high-rate discharge or precursors of thermal runaway within the battery pack, the collected data vector exhibits a sharp, non-linear jump. This jump will cause fluctuations in data stream metrics. It surged and exceeded the preset alert threshold in a very short period of time. Once this condition is met, the system will immediately perform a segmentation operation, solidifying all data points accumulated in the buffer pool at this moment and outputting them as a sequence. In this case, regardless of the number of data points accumulated in the current buffer pool... How much, trigger signal All are immediately set to true to ensure the most timely capture and encapsulation of the entire transient process.

[0026] when The time is defined as the steady-state upper limit trigger path, which is a regular trigger path. Its activation must simultaneously satisfy two constraints. A segmentation operation is performed, solidifying all data points accumulated in the buffer pool at this moment and outputting them as a sequence. Specifically: the first condition is... When the Energy Management System (EMS) is in steady-state operation, such as at the end of constant current charging or in a static state, the collected data vector changes gradually, resulting in data stream fluctuation indicators. The value remained below the tranquility threshold. .

[0027] The second condition is The number of data points currently accumulated in the buffer pool Reaching or exceeding the preset maximum sequence length If both conditions one and two are met, a data split is performed, triggering a signal. All are immediately set to true; when the trigger signal is triggered When the condition is met, the buffer pool contains data from the start time. up to the current time All data vectors are organized into a single two-dimensional matrix, namely the first data sequence S, as follows: ;in, The first data sequence is a two-dimensional numerical matrix that encapsulates a complete dynamic event or steady-state period, and its values ​​are... A real matrix of dimension ; The concatenation operator is used to combine discrete data vectors into a single structured entity; T represents the transpose operator. Representing the start timestamp of the sequence, it is the first data vector in the current sequence to be formed. The sampling timestamp, which takes a positive integer value, is used to identify the time start point of the sequence; Representing the end timestamp of the sequence, it is the last data vector in the current sequence to be formed. The sampling timestamp, whose value is not less than A positive integer used to identify the time endpoint of the sequence.

[0028] Step S12: Map the first data sequence to the second data sequence. Parallel generation of absolute position code (PE) and relative timestamp code. Combine the two codes with the second data sequence Perform weighted fusion to generate the first feature sequence. In the specific implementation process, when the first data sequence When directly input into a downstream model, the lack of temporal sequence and interval representation makes it easy for the model to lose its awareness of the time dimension. Therefore, a feature engineering strategy is proposed to enrich the temporal context information of the first data sequence S. Further, in step S12, the first feature sequence... The generation method includes the following steps: Step 121, Input projection: Project the first data sequence The feature matrix is ​​linearly transformed to a higher dimension, the dimension of which can be set; specifically, the first data sequence is transformed through a learnable projection layer. Linear transformation to a higher-dimensional feature matrix, i.e., the second data sequence. The second data sequence It is a dimension A real matrix; This represents the feature dimension of the model, and its value is usually greater than that of the first data sequence. The dimension D of each data vector.

[0029] Step 122: Multi-scale temporal feature generation: Generate positional codes based on the sequential order of the data sequence, and generate relative timestamp codes based on the time difference between the current sequence and the starting point of the data. Encode the location and the relative timestamp. As a multi-scale temporal feature, this stage generates two different scale temporal codes in parallel: absolute position encoding (PE) and relative timestamp encoding. (Normalized Relative Timestamp Encoding, NRTE).

[0030] The generation of absolute positional encoding aims to endow data sequences with order-awareness. Optionally, the absolute positional encoding method employs deterministic sine and cosine functions for the second data sequence. Each position in, i.e., 0 to Generate a unique location identifier vector that is independent of the data content.

[0031] Absolute positional encoding (PE) enables the model to understand the relative positional relationships of different data points within the sequence structure, i.e., the even-numbered dimension components of the positional encoding vector. odd-dimensional components of the positional encoding vector The absolute position encoding formula is as follows: Where pos represents the position index, and its value ranges from 0 to... non-negative integers; This represents the encoded dimension index, with values ​​ranging from 0 to... The integer is used to control the frequency of the sine and cosine functions; The even-numbered dimension component of the positional encoding vector corresponds to the first element in the sequence. The position, the first position in its encoded vector Each dimension has a scalar value, and its value is... Real numbers within the interval; The odd-dimensional component representing the positional encoding vector corresponds to the first element in the sequence. The position, the first position in its encoded vector Each dimension has a scalar value, and its value is... Real numbers within the interval; 10000 represents the standard sine and cosine trigonometric functions, which are the basic mathematical functions for generating periodic position-coded signals; 10000 represents the wavelength control base, which is used to stretch the wavelength of the trigonometric functions to a sufficiently large range to ensure that the codes at different positions have good distinguishability. Indicates the first The wavelength used for the dimension.

[0032] Optionally, relative timestamp encoding is used to calculate each data point relative to the second data sequence. The normalized time difference at the starting point is calculated using the following formula: ;in, Representing a relative timestamp scalar, it corresponds to the second data sequence. The data point at position pos in the middle has a value that is a... Floating-point real numbers within the interval; The timestamp representing the current position pos; Indicates the start timestamp of the sequence, which is used as a reference point for calculating all relative times; Indicates the end timestamp of the sequence, used in conjunction with the start timestamp. Together, they determine the total physical time span of the sequence; Represents a very small positive number. It is a preset positive floating-point number close to zero, used to avoid division errors with a denominator of zero when dealing with sequences of length 1.

[0033] The generation of relative timestamp encoding aims to give data sequences a sense of duration; this relative timestamp encoding This enables the model to perceive the sequence over real physical time spans, such as distinguishing between a transient shock sequence lasting 0.5 seconds and a steady-state change sequence lasting 60 seconds.

[0034] Subsequently, the relative timestamp scalar is expanded using a dimension expansion function. Mapped to a high-dimensional relative timestamp vector .

[0035] Step 123, Feature Fusion: The second data sequence... The first feature sequence is generated by weighted fusion with multi-scale temporal features. Specifically, the second data sequence Absolute position encoding and relative timestamp encoding Weighted fusion is performed to generate the final first feature sequence. The specific process formula is as follows: ;in, This represents the first feature sequence, whose value is a dimensionless sequence. A real matrix; This represents learnable absolute position encoding fusion weights, used to automatically balance the importance of absolute position information during training; Indicates absolute position encoding; This represents learnable relative temporal encoding fusion weights, used to automatically balance the importance of relative temporal information during training; This indicates a relative timestamp encoding.

[0036] In the above embodiments, in order to resolve the first data sequence output in step S11 Due to the lack of explicit temporal relationships, downstream models struggle to understand the data effectively. Step S1200 addresses this by introducing and fusing multi-scale spatiotemporal features into the first data sequence to generate a first feature sequence rich in temporal semantics. .

[0037] Step S13: Perform global semantic compression on the first feature sequence to generate the first hidden state sequence: [The first feature sequence is then compressed...] The input is fed into a lightweight Transformer encoder for global context capture and multi-layer information distillation, transforming the first feature sequence... Semantic compression results in the first hidden state sequence Z.

[0038] Specifically, this step aims to refine the first feature sequence generated in step S12. The system performs semantic compression, transforming the information-rich but redundant explicit data representation into a highly condensed, low-dimensional implicit data representation, namely the first hidden state sequence Z. This first hidden state sequence Z not only significantly reduces the data dimensionality, but more importantly, each vector within it encodes the global context information of the entire input sequence, thus providing structural support against packet loss and out-of-order decoding at the cloud receiving side.

[0039] In the specific implementation process, this step is the core of realizing the transformation from data transmission to information transmission, and it involves the first feature sequence. The input is fed into a lightweight Transformer encoder optimized for edge computing scenarios. The lightweight design refers to the fact that this Transformer encoder significantly reduces computational and storage overhead while maintaining global modeling capabilities through techniques such as parameter pruning, knowledge distillation, or efficient self-attention variants, ensuring deployability on resource-constrained edge devices.

[0040] In step S13, global semantic compression is performed based on the lightweight Transformer encoder to generate the first hidden state sequence, including the following steps: Step 131, Global Context Capture: For the input first feature sequence The Transformer encoder learns the dependency weights between two time steps based on an attention mechanism to obtain global contextual information. The core of the Transformer encoder lies in its global context capture mechanism, namely, a multi-head self-attention mechanism, which is a sequence-to-sequence information processing unit. Unlike traditional models that capture proximity relationships, this mechanism can compute the dependency weights between any two time steps in the sequence in parallel, constructing a global contextual graph. For example, this global context capture mechanism can autonomously learn the strong correlation between "the temperature of cell A rises sharply at second 5" and "the voltage of cell B drops abnormally at second 15".

[0041] Step 132, Information Distillation and Semantic Compression: Perform information distillation and semantic compression on the obtained global context information to obtain the first hidden state sequence Z; specifically, when the first feature sequence After being processed by stacking multiple layers of the aforementioned encoders, the global context information contained within is distilled and semantically compressed, ultimately outputting a first hidden state sequence Z of the same length as the input but with lower dimensionality. Here, the first hidden state sequence Z is a sequence with dimension... The matrix, The dimension representing the first hidden state is used to determine the compression ratio of semantic compression.

[0042] S2000, the semantic change rate based on the first hidden state sequence Z. and network channel quality The dual perception, combined with a dynamic decision function, generates the optimal sending strategy command. and according to the optimal sending strategy instruction Perform independent encapsulation or buffer aggregation operations on the first hidden state sequence Z, and configure the metadata header. It is encapsulated as an application layer data stream.

[0043] Specifically, this step aims to address the technical problem of mismatch between transmission efficiency and service timeliness when transmitting semantically compressed data between servers using the Transport Control Protocol (TCP), due to dynamic changes in data semantics and network state fluctuations. This step creatively proposes a proactive and adaptive transmission optimization strategy. It dynamically senses the semantic value of the first hidden state sequence Z and the channel transmission capability of TCP in real time, dynamically adjusting the aggregation strength and transmission timing of application layer data. This transforms the first hidden state sequence Z output in step S1300 into an application layer data stream with context awareness and dynamic scheduling characteristics, significantly improving network bandwidth utilization and ensuring that high-value critical information is transmitted in the network link with minimal end-to-end latency. This surpasses traditional fixed transmission strategies, achieving a balance between high reliability, low latency, and efficient transmission.

[0044] Step S1 of this embodiment adaptively segments the original multidimensional data vector and injects multi-scale temporal features, and uses a Transformer encoder to perform global semantic compression, condensing and distributively encoding the context information of the entire temporal event into each output hidden state vector. This improves the problem that in a weak network environment, the receiving end lacks sufficient context clues due to receiving some out-of-order or lost data packets, making it difficult to actively and accurately reconstruct the original complete data sequence.

[0045] Further, in step S2, the method of encapsulating the first hidden state sequence into an application layer data stream based on the semantic change rate of the first hidden state sequence and the network channel quality, combined with the dynamic decision function to generate the optimal transmission strategy instruction, includes: step S21, collaboratively sensing the semantic change rate of the first hidden state sequence Z. and real-time network channel quality The optimal sending strategy command is generated by combining the dynamic decision function. In the specific implementation process, determining the application layer's sending strategy is a crucial step before sending the continuous first hidden state sequence Z to the cloud server via the Transmission Control Protocol (TCP) connection. Traditional fixed-length segmentation strategies, such as sending at fixed time periods or with fixed data sizes, have significant drawbacks: First, they may delay the sending of sequences containing key semantic information while waiting for data aggregation, missing the optimal opportunity for immediate response; second, they lack the ability to perceive the rapidly changing TCP channel state, attempting to send large aggregated data blocks when the channel is congested, exacerbating network latency, and failing to fully utilize available bandwidth when the channel is open, thus reducing overall data throughput due to an overly conservative sending strategy.

[0046] To overcome the aforementioned shortcomings, this step proposes a dual-perception adaptive aggregation and transmission decision mechanism. This mechanism is based on a dynamic decision function and considers the following two dimensions of input: the first dimension is content transformation perception, i.e., source perception. The semantic change rate of the first hidden state sequence Z is calculated in real time. This is used to quantify the degree of change in the operating state of the Energy Management System (EMS). In regions with smooth semantic transitions, such as when the EMS is running stably, the semantic change rate is... A lower rate indicates that the timeliness requirement for data is not high, and the system tends to aggregate the current first hidden state sequence Z at the application layer to build a larger data load and improve the effective data payload ratio of a single TCP transmission. Conversely, when semantic inflection points or abrupt change regions are detected, such as before and after a fault, the semantic change rate is higher. A pulsed rise will occur, indicating that the first hidden state sequence Z contains high-value key information. The system will immediately send the sequence to ensure that the key semantic event can reach the cloud analysis server with minimal application layer latency.

[0047] The second dimension is Transmission Control Protocol (TCP) channel state awareness, or channel awareness. This involves monitoring relevant parameters of the local TCP protocol stack on the edge server to obtain network quality indicators in real time. For example, Round-Trip Time (RTT) and Congestion Window (cwnd) size are used to evaluate the real-time carrying capacity of inter-server links. When these metrics perform well—low RTT and a large congestion window—it indicates that the network links are unobstructed and can efficiently transmit large data blocks. The system will then be more inclined to adopt an aggressive aggregation strategy to maximize the utilization of network bandwidth resources. Conversely, when network quality metrics are poor... When the situation deteriorates, i.e., round-trip latency increases and congestion window decreases, the system will proactively switch to a conservative strategy, such as reducing the aggregation size or switching to immediate transmission, in order to avoid exacerbating network congestion by sending excessively large application layer data streams and to ensure smooth data transmission.

[0048] To implement the aforementioned adaptive aggregation and transmission decision mechanism, the optimal transmission strategy is determined by a dynamic decision function, which incorporates the semantic change rate. and network quality indicators The perception results from the two dimensions are used as input and mapped to a predefined optimal sending strategy command. : ;in, This represents the optimal sending strategy instruction output after decision-making, and its value can be one of the following instructions: "Send immediately", "Conservative aggregation", or "Active aggregation".

[0049] Ultimately, the decision function outputs the optimal transmission strategy command. This will directly guide the subsequent steps in how to process the current first hidden state sequence Z, i.e., whether to encapsulate it independently or aggregate it, rather than physically splitting it, thereby enabling intelligent and refined control of the data flow at the application layer.

[0050] In the implementation of step S21 above, addressing the limitations of traditional fixed-transmission strategies in terms of efficiency and latency when data is reliably transmitted between servers, this embodiment proposes an adaptive aggregation and transmission mechanism based on dual perception. This mechanism dynamically balances semantic timeliness and channel transmission efficiency to determine the optimal transmission strategy instruction. According to the instruction The first hidden state sequence Z is processed. This mechanism can simultaneously consider the semantic timeliness of the data content and the transmission efficiency of the TCP physical channel at the decision-making time of data transmission, realizing the transformation from data transmission to intelligent dynamic information control.

[0051] Step S22: Dynamically schedule the first hidden state sequence Z using a dynamic transmission buffer driven by a strategy and combined with composite triggering conditions, according to the optimal transmission strategy instruction. Perform independent encapsulation or buffered aggregation operations to build application-layer data payloads. Specifically, this step aims to implement the optimal transmission strategy instruction generated in step S21. The first hidden state sequence Z output in step S13 is differentiated, and the data payload is constructed through a dynamic send buffer to output the application layer data payload. It is the execution link connecting intelligent decision-making and efficient transmission, and the physical basis for application-layer traffic regulation of reliable connections such as Transmission Control Protocol (TCP).

[0052] In the actual implementation process, since the underlying transport protocol already uses the Transmission Control Protocol (TCP) to ensure data reliability, this step no longer needs to add a fault tolerance mechanism. Instead, the focus is on how to make more intelligent use of reliable channel resources.

[0053] This step proposes a policy-driven application-layer buffering and scheduling mechanism. Unlike the traditional approach of directly submitting data to the TCP protocol stack, this mechanism introduces optimal sending policy instructions at the application layer. A strongly correlated dynamic transmission buffer intelligently schedules each hidden state sequence Z to be transmitted, including: according to the optimal transmission strategy instruction. Perform independent encapsulation, specifically: independent encapsulation for the immediate transmission strategy. When the optimal transmission strategy instruction is received... If the sequence is to be sent immediately, it indicates that the current first hidden state sequence Z has high timeliness value. In this case, the sequence will bypass or immediately clear the send buffer and be directly constructed as an independent application layer data payload. This involves encapsulating the data payload of a single sequence. This operation ensures minimal processing latency for high-value data, allowing it to be prioritized and submitted to the next step.

[0054] Based on the optimal sending strategy command Perform buffer aggregation operations, specifically: buffer aggregation for conservative or aggressive aggregation strategies. This occurs when the optimal transmission strategy instruction is received. When it is an aggregation type, it indicates that the current data is suitable for a high-efficiency transmission method. In this case, the current first hidden state sequence Z will be sent to the application layer send buffer and concatenated with one or more existing sequences in the buffer to form a more information-dense and aggregated application layer data payload. That is, encapsulating data payloads of multiple sequences.

[0055] To ensure the effectiveness and timeliness of buffered aggregation operations, the aggregation process is set with explicit termination conditions to avoid indefinite continuation.

[0056] A further technical solution involves controlling the submission of data payloads in the dynamic transmission buffer using the following triggering conditions: First, policy change triggering: when a sequence with an immediate transmission policy arrives, the aggregated data in the buffer will be prioritized for construction into application layer data payloads and submitted to subsequent steps; Second, capacity limit triggering: when the accumulated data volume in the buffer reaches a preset maximum capacity threshold, the construction and submission of data payloads will be automatically triggered to avoid excessive application layer data flow and maintain the efficiency of the Transmission Control Protocol (TCP); Third, when the waiting time of the first sequence in the buffer exceeds a preset maximum delay, the construction and submission of data payloads will be forcibly triggered to ensure data timeliness and avoid excessive delays due to prolonged failure to reach the capacity threshold; Finally, the optimal transmission policy instruction is output. Scheduled structured application layer data load This achieves a dynamic adaptive balance between semantic timeliness and channel state.

[0057] Step S23: Application layer data load Configure metadata headers It is encapsulated as an application layer data stream.

[0058] Specifically, this step aims to load the application layer data generated in step S22. This is encapsulated into a complete application-layer data stream with application-layer intelligence. By designing a structured metadata header containing forward-looking processing instructions, the contextual intelligence formed by the edge transmitter during semantic awareness, channel state assessment, and dynamic scheduling is encapsulated. This enables the cloud server to anticipate the importance, structural complexity, and recommended processing methods of the received data, thereby performing proactive and differentiated resource allocation and task scheduling to solve the technical problem of critical information processing delays caused by unified data stream processing.

[0059] In the specific implementation process, since this step uses the Transmission Control Protocol TCP as the transport layer protocol, its inherent reliability mechanism ensures the integrity and order of data. Therefore, the innovation of this step is not to provide a guarantee for the data reconstruction and decoding mechanism, but to build a processing coordination mechanism between the sending end and the receiving end at the application layer by designing a structured metadata header.

[0060] This metadata header enables each application layer protocol data unit to have self-description and self-scheduling capabilities. The receiving server does not need to rely on additional out-of-band signaling; it can obtain the complete context of a single data unit and perform optimal processing based on that data unit alone. Each application layer protocol data unit From a metadata header and an application layer data load It is pieced together. Among them, The application layer protocol data unit represents the basic information unit transmitted between servers via the Transmission Control Protocol (TCP) connection. This indicates a concatenation operation, used to concatenate metadata headers. and application layer data load Logically assembled into a single data unit.

[0061] To assign data units to the application layer It has complete self-description and self-scheduling capabilities, enabling intelligent collaboration between the sender and receiver at the application layer.

[0062] One specific implementation method, metadata header It includes the following key fields: Global Sequence Identifier (Sequence_ID), Semantic Importance Level Load characteristics description Forward-looking processing instructions and sending timestamp Global sequence identifier Sequence_ID: will be determined by the start identifier. and end identifier The resulting tuple serves as a global sequence identifier, used to uniquely identify application layer data units. The start and end range of one or more first hidden state sequences Z contained therein, when the tuple is encapsulated independently, the start identifier and end identifier of the tuple are the same; semantic importance level The semantic change rate of the first hidden state sequence in step S21 The quantized value obtained by mapping is used as the semantic importance level. This is used to define the business value of the current data load and is the core basis for cloud servers to perform differentiated processing; load characteristic description It is a description of application layer data load Composite fields that combine internal structure and data characteristics, specifically including: aggregate counts. and average compression ratio

[0063] Aggregate count This represents the number of first hidden state sequences Z aggregated in the data load; average compression ratio. The quantitative metric representing the average semantic compression level of the data in this data payload is determined by the sender based on the collected original multidimensional data vectors. The ratio of the total size to the total size of the first hidden state sequence Z generated by semantic compression in step S13 is used to help the cloud server estimate the computing resources required for decoding or reconstruction; prospective processing instructions : is determined by the edge sender based on semantic importance level and network quality indicators The enumerated values ​​generated by the comprehensive decision-making process are not simply descriptions of the data, but rather explicit processing and scheduling instructions sent to the cloud server. For example, PROC_URGENT instructs the cloud receiving server to route this data unit to a high-priority real-time analysis queue; PROC_NORMAL instructs the cloud receiving server to process it according to the regular first-in-first-out queue; and PROC_BATCH instructs the cloud receiving server to route this data unit to a low-priority batch processing queue or to archive it directly. The sending timestamp... It is a high-precision timestamp, generated by the edge transmitter when it is about to send the application layer protocol data unit. The timestamp recorded when the data is sent to the TCP send buffer.

[0064] Application layer data load Configure metadata headers Then, an application layer protocol data unit carrying the corresponding metadata header will be generated. These data are sequentially sent to the TCP send buffer. Over time, N application layer protocol data units are generated consecutively. Together, they form a logical and intelligent application layer data flow, endowing the data flow with rich application layer context and scheduling information, thereby achieving end-to-end performance improvement.

[0065] Step S2 of this embodiment adaptively segments the original multidimensional data vector and injects multi-scale temporal features, and uses a Transformer encoder to perform global semantic compression. This condenses and distributes the context information of the entire temporal event into each output hidden state vector, thereby improving the problem that in a weak network environment, the receiving end lacks sufficient context clues due to receiving some out-of-order or lost data packets, making it difficult to actively and accurately reconstruct the original complete data sequence.

[0066] Step S3: Perform streaming parsing and multi-dimensional cross-validation on the application layer data stream sequentially. If the validation passes, generate the second hidden state sequence. If the verification fails, an abnormal hidden state sequence is generated. Transferred to the guided semantic repair module to generate the second hidden state sequence This step aims to perform deep logical consistency and semantic rationality checks on the application layer data stream output from step S2300, and output a high-confidence, logically consistent second hidden state sequence. .

[0067] In step S3, the data receiving end sequentially performs streaming parsing and multi-dimensional cross-validation on the received application layer data stream. If the validation passes, a second hidden state sequence is generated; if the validation fails, an abnormal hidden state sequence is generated. Guided semantic repair is then performed to generate the second hidden state sequence, including the following steps: Step S31, processing application layer protocol data units in the application layer data stream. Perform streaming parsing and decoding to the first temporary hidden state sequence Based on the application layer protocol data unit metadata header For the first temporary hidden state sequence Perform multi-dimensional cross-validation; if the validation passes, generate a second hidden state sequence. If any verification fails, it is determined that there is an application layer data anomaly, and an anomaly hidden state sequence is generated. .

[0068] This step, serving as the first line of defense in cloud server data processing, aims to identify and capture logical or semantic data anomalies at the edge due to sensor malfunctions or application anomalies, rather than repairing the data integrity and order guaranteed by the TCP transmission control protocol. The innovation of this step lies in capturing the metadata header from the application layer data stream output in step S23 through multi-dimensional cross-validation. Application layer data anomalies that do not match the description should be addressed to ensure semantic consistency and business reliability of the data flow.

[0069] In some embodiments, the received application layer data stream is sequentially parsed and cross-validated using streaming, as follows: Step S311: The application layer data stream is segmented and reassembled into each application layer protocol data unit using a length prefix message delimitation method, and the data payload is extracted and decoded into a first temporary hidden state sequence; specifically, the cloud server receives the application layer data stream transmitted via the TCP transmission control protocol connection in step S23, which is represented as a continuous byte stream at the physical layer. To accurately segment and reassemble each application layer protocol data unit from this byte stream... The message delimiting method uses a length-prefix approach. First, a fixed-length length-prefix field is read from the TCP receive buffer and parsed to determine the total length of the subsequent message body. Subsequently, based on the total length value... Continue reading the corresponding byte block from the TCP receive buffer; this byte block is a complete and independent application layer protocol data unit. .

[0070] After identifying each application layer protocol data unit Then, extract its application layer data payload. And decode it into the first temporary hidden state sequence. .

[0071] Step S312, Smart Cross-Validation: Using metadata headers As a reliable benchmark, the first hidden temporary state sequence decoded in the previous stage is... Multi-dimensional cross-validation is performed, including semantic value verification, structural feature verification, and instruction content verification; the first dimension, semantic value verification, involves calculating the temporary first hidden state sequence. The degree of internal semantic fluctuation, and its correlation with the metadata header. semantic importance ranking Compare them.

[0072] If a piece of data marked as low importance (i.e., with a low semantic importance level) exhibits drastic semantic fluctuations, it is determined that there is a logical contradiction, which may indicate an unknown or abnormal event pattern.

[0073] The second dimension, structural feature verification: examining the temporary first hidden state sequence. The actual structure and metadata header Do the relevant characteristics match? For example, does the number of encapsulated sequences match the load characteristics described? For example, aggregate counting It is strictly consistent.

[0074] The third dimension is instruction content verification. This involves analyzing the temporary first hidden state sequence. The pattern presented by the data content is related to the metadata header it carries. Forward-looking processing instructions in The intent must match; for example, data given the PROC_URGENT instruction should present a pattern that matches the characteristics of a fault, and if it does not, it is considered abnormal.

[0075] Step S313: If all verifications pass, the data is determined to be normal, and a complete second hidden state sequence is generated. Its value is the first temporary hidden state sequence. If any verification fails, the current application layer protocol data unit is determined to be invalid. If an application-layer data anomaly exists, generate an anomaly hidden state sequence marked with the anomaly location and type. and its corresponding metadata header Transferred to the next level semantic repair engine module for guided semantic repair; Step S32, Semantic Repair Engine Module: Transfers the abnormal hidden state sequence The positions of the hidden state vectors marked as abnormal are replaced with mask tokens, and the sequence of abnormal hidden states is changed. Corresponding metadata header The input is fed into the Transformer decoder, which then collaboratively utilizes the sequence of anomalous hidden states. Internal global semantic information and metadata header Based on semantic prior knowledge, generate the second hidden state sequence. .

[0076] Specifically, this step, as the core of the application-layer self-healing system, aims to utilize the contextual reasoning capabilities of deep learning models to analyze the abnormal hidden state sequence generated in step S310. Perform guided semantic repair to ensure the generated second hidden state sequence Complete, continuous, and consistent with business logic.

[0077] In the specific implementation process, this step uses a Transformer encoder that is symmetrical to the encoder at the transmitting end in step S13 to process the abnormal hidden state sequence. Perform semantic repair; the repair process of guided semantic repair includes the following steps: Step S321, Mask sequence construction: construct the abnormal hidden state sequence All hidden state vector positions marked as anomalous or logically inconsistent are replaced with predefined, learnable [MASK] mask tokens to construct a structurally complete mask input sequence, but with the positions to be repaired identified; Step S322, Semantic Reasoning and Condition Generation: The mask input sequence is transmitted to the Transformer decoder, with metadata header... semantic importance ranking Guided by forward-looking processing instructions, the second hidden state sequence is generated after decoding. Input the constructed mask sequence and its corresponding metadata header. The input is shared with a Transformer decoder deployed in the cloud. This decoder, through its multi-head self-attention mechanism, can comprehensively consider all normal hidden state vectors in the sequence when predicting the vector at any [MASK] token position. Since the sender has already distilled and encoded the global context of the entire sequence into each hidden state vector through its self-attention mechanism during semantic compression in step S13, the abnormal hidden state sequence... Each non-anomaly hidden state vector carries global semantic information. Based on this global semantic information, the decoder further introduces a metadata header. semantic importance ranking and forward processing instructions This creates global semantic prior constraints, thereby setting boundaries and guidance for the semantic repair of the decoder, and achieving goal-oriented guided generation. For example, when processing instructions in advance... When the value is PROC_URGENT, the decoder is guided to generate a vector sequence that better matches the fault characteristics to fill the mask positions, rather than smooth steady-state data, in order to prioritize responding to abnormal scenarios; when the semantic importance level is... At lower levels, the decoder tends to generate vector sequences with less volatility that are more in line with normal operating patterns.

[0078] By jointly driving the solution with contextual semantic information and prior knowledge of metadata, the optimal solution is maintained not only in a probabilistic sense but also in a high degree of consistency with the original intent of the sender in terms of business logic. This improves the accuracy and rationality of the repair results and generates a high-confidence, semantically coherent second hidden state sequence. .

[0079] This embodiment cross-validates the decoded data content using the metadata header as a trusted benchmark and initiates guided repair for data that fails to pass verification. This solves the problem that traditional reliable transmission can only guarantee the physical integrity of the data but cannot verify the logical rationality of the content, which may lead to the cloud system making incorrect decisions based on abnormal data.

[0080] In S4, specifically, based on the second hidden state sequence Calculate end-to-end transmission delay Semantic reconstruction confidence and posterior abnormality score Attribution analysis is performed based on a preset decision tree logic to output attribution labels, and dynamic control signals are generated based on the attribution labels. The dynamic control signal The data is transmitted back to the edge transmitter via a high-priority control channel to complete the adaptive update of the transmission strategy.

[0081] This step aims to construct an end-to-end intelligent closed-loop control mechanism by performing in-depth evaluation and attribution of the quality of data reconstruction in the cloud, i.e., the second hidden state sequence output in step S3. It gains the ability to self-aware of the operational status of the entire end-to-end link and generates dynamic control signals. This information is then fed back to the edge to guide the optimization and adjustment of the transmission strategy, thereby evolving the system from a static data transmission pipeline into a cognitive communication system that proactively adapts to environmental changes.

[0082] Further, in step S4, the method for adaptively updating the transmission strategy includes: Step S41, calculating three indicators based on the second hidden state sequence, including end-to-end transmission delay, semantic reconstruction confidence, and posterior anomaly score; comprehensively judging the three indicators based on decision tree logic; and attributing the root cause of the problem to one of the following based on whether each indicator exceeds a preset threshold: transmission bottleneck, repair uncertainty, or novel anomaly event, to output an attribution label; specifically, based on the second hidden state sequence... Calculate end-to-end transmission delay Semantic reconstruction confidence and posterior anomaly scores Based on decision tree logic, the three indicators are comprehensively evaluated. Depending on whether each indicator exceeds the preset threshold, the root cause of the problem is determined and the attribution label is obtained. The attribution label can include transmission bottlenecks, repair uncertainties, and new abnormal events to identify the operating status of the data transmission link.

[0083] In the above implementation, the second hidden state sequence generated in step S3 A multi-dimensional and in-depth quantitative evaluation is conducted to analyze the root causes of performance bottlenecks or data anomalies, and an attribution label is output to provide a decision-making basis for the subsequent generation of adaptive dynamic control signals.

[0084] This step begins with multi-dimensional quality quantification. Quantitative indicators are extracted from three dimensions: transmission efficiency, repair quality, and the content itself, enabling a comprehensive understanding of attribution judgment: Regarding transmission efficiency, end-to-end transmission latency is calculated. The end-to-end transmission delay is defined as the time interval between the time the receiver receives data and the time of transmission recorded in the metadata header of the application layer data stream. This indicator reflects the overall efficiency of the network link and the transmission strategy. Specifically, in the application scenario of this embodiment, it is achieved through the second hidden state sequence. The sending timestamp is encapsulated in the corresponding metadata header. Compare the timestamp received by the current cloud server to calculate the data unit of each application layer protocol. End-to-end transmission delay .

[0085] To address the quality of repair, the confidence level of semantic reconstruction is calculated. To generate the second hidden state sequence The information entropy is calculated by analyzing the probability distribution of the mask token positions during the process to obtain the semantic reconstruction confidence. The specific semantic reconstruction confidence index is used to measure the model's confidence in its own repair results. If step S32 is triggered, then the second hidden state sequence is generated. During the process, a probability distribution is output for each [MASK] mask token position. The semantic reconstruction confidence is obtained by calculating the information entropy of this probability distribution and normalizing it. If the probability distribution is highly concentrated, it indicates that the model has a high degree of certainty about the generated results, corresponding to a high confidence level; conversely, if the probability distribution tends to be flat, it indicates that the model has high uncertainty, corresponding to a low confidence level, suggesting that the context of the input sequence may contain abnormal conditions.

[0086] At the content level, calculate the posterior anomaly score. The repaired second hidden state sequence is calculated using a deep learning model. The delayed anomaly score; this metric is used to assess the degree of anomaly in the final generated data content itself. The second hidden state sequence... Input a pre-trained anomaly detection model specifically designed for representing patterns in normal operating condition data, such as an autoencoder, and output a posterior anomaly score. If the posterior abnormal score High, even the second hidden state sequence While it has been structurally and semantically restored to be coherent, the overall event pattern it describes still deviates significantly from the normal state.

[0087] Furthermore, attribution analysis using decision trees yields attribution labels: Based on the pre-defined decision tree logic, a comprehensive assessment of the three quantitative indicators is conducted, attributing the root cause of the problem to one of three more refined core root causes: First, transmission bottleneck attribution. If end-to-end transmission delay... If the transmission degradation threshold is exceeded, the root cause of the problem is determined to be transmission link congestion or an inappropriate current transmission strategy, and the attribution label is marked as "TRANSMISSION_BOTTLENECK".

[0088] Second, address the uncertainty attribution. If the semantic reconstruction confidence level... If the repair confidence threshold is lower than the preset threshold, it is inferred that the transmission process itself is unimpeded. The problem lies in the repair process in step S32 encountering a rare damage mode outside the model's generalization ability, resulting in high uncertainty in the repair result. At this time, the attribution label is marked as "REPAIR_UNCERTAINTY".

[0089] Third, attribution of novel anomalous events. If the posterior anomalous score... The event exceeds the preset event threshold, and the end-to-end transmission delay is... and semantic reconstruction confidence If all data is within the normal range, the transmission process is considered efficient and the repair process has high confidence. However, if the final generated data reveals a high-value, previously unseen novel anomaly, the attribution label will be marked as "NOVEL_EVENT_DETECTED".

[0090] If all three indicators are within the normal range, the current recovery quality is considered acceptable, and the Attribution label is marked as "NOMINAL".

[0091] Step S42: Based on the attribution label, query the predefined decision rule base to generate the corresponding dynamic control signal. This serves as the updated data transmission strategy; specifically, this step aims to transform the Attribution label output in step S41 into specific, executable parameter adjustment instructions, forming a structured dynamic control signal. This enables targeted control over the behavior of edge transmitters.

[0092] In the specific implementation process, this step, based on a predefined decision rule base, maps the input attribution labels to a set of specific control instructions. This ensures that the system's response behavior is targeted and applies the most appropriate adjustment strategy to problems with different root causes, avoiding the blindness and inefficiency of traditional single adjustment mechanisms. Different strategy branches are executed according to the different attribution labels: Step 421, based on the attribution label, when the problem is identified as a transmission efficiency issue, output adjustment instructions to reduce network load and alleviate link congestion; if the attribution is a transmission bottleneck issue, i.e., the attribution label is marked as "TRANSMISSION_BOTTLENECK", then the current performance bottleneck is determined to be transmission efficiency. Therefore, the generated dynamic control signal... It will contain a set of instructions aimed at reducing network load and alleviating link congestion; specifically, the control instructions for reducing network load and alleviating link congestion include: (1) improving semantic compression rate: the instruction edge reduces the dimension of the first hidden state sequence Z in step S1300. (2) Implement a conservative aggregation strategy: The instruction edge lowers the maximum capacity threshold of the buffer in step S2200 to generate application layer protocol data units that are smaller in size and sent more frequently. This allows for adaptation to lower channel capacity.

[0093] Step 422: Based on the attribution label, when data corruption is identified, the semantic compression rate of the data sent at the data transmission end is reduced. Specifically, if the attribution is due to repair uncertainty, i.e., the attribution label is marked as "REPAIR_UNCERTAINTY", then the problem is determined to be that the data corruption pattern is rare, leading to high uncertainty in the repair model. Generated dynamic control signals. This will include a set of instructions designed to provide higher quality input to the repair engine. Preferably, the tuning instructions are to appropriately reduce the semantic compression ratio: the semantic compression ratio in step S13 is reduced at the instruction edges, for example, by increasing the dimension of the first hidden state sequence. This preserves richer original details for subsequent data transmission, provides more contextual clues for the cloud-based repair model, and improves the confidence of the next repair.

[0094] Step 423: Based on attribution labels, when a novel anomalous event problem is identified, the semantic compression rate of the data at the data sending end is reduced, and an alarm is issued; if the novel anomalous event problem, i.e., the attribution label is marked as "NOVEL_EVENT_DETECTED", then the system is determined to have discovered a high-value unknown event. At this point, the goal is no longer transmission efficiency, but rather the fidelity and integrity of the information. Therefore, the generated dynamic control signal... This will include a set of instructions designed to achieve the highest fidelity capture. Preferred instructions include: First, disabling or minimizing semantic compression: the instruction edge significantly reduces the compression intensity in step S1, or even completely suspends semantic compression in extreme cases, to ensure data integrity for the novel event, allowing for subsequent offline analysis and model iteration. Second, triggering high-priority alarms and sample collection: a high-priority alarm for an unknown operating condition event is triggered at the cloud application layer, and the current data is automatically labeled as high-value learning samples for online or offline retraining of the cloud model.

[0095] If the Attribution label is marked as "NOMINAL", the system is in a stable state and can selectively execute mild optimization instructions or maintain the current parameters unchanged.

[0096] Step S43: The dynamic control signal is transmitted back to the edge transmitter through the high-priority control channel. The edge transmitter then parses the dynamic control signal and updates the transmission strategy online.

[0097] An updated data transmission strategy is generated, and the dynamic control signal of the updated data transmission strategy is transmitted through a high-priority control channel. The data is transmitted back to the edge transmitter, where the dynamic control signal is processed. Perform analysis and update the transmission strategy online.

[0098] Specifically, this step aims to process the dynamic control signal generated in step S42. It reliably delivers data to the edge execution layer and completes policy updates, enabling the static transmission process to evolve into an intelligent system that dynamically interacts with the environment and continuously optimizes itself.

[0099] In practice, this step is the final execution stage for adaptive control, which involves processing the dynamic control signals from the cloud. Translated into actual edge operations, this includes two core actions: First, the return of control signals: The cloud server encapsulates dynamic control signals containing specific instructions. The data stream is reliably sent back to the designated edge sender that generated it via a high-priority, logically independent Transmission Control Protocol (TCP) control channel. To ensure reliable and low-latency delivery of control commands, this return channel is separated from the main data transmission link channel. This dynamic control signal... The data volume is extremely small and the transmission frequency is low, so the bandwidth overhead it occupies is negligible and will not have a significant impact on the main data transmission link.

[0100] Second, dynamic policy updates at the edge: A policy management module is deployed at the edge transmitter, which continuously monitors the control channel from the cloud. Once a dynamic control signal is received... The module immediately parses it and dynamically and online updates the core operating parameters used in semantic compression in step S1000 and data transmission in step S2000, thereby completing the update loop of the entire adaptive transmission strategy.

[0101] This invention improves upon the limitations of traditional adaptive transmission mechanisms, which are unable to identify the root causes of performance bottlenecks and can only adopt single, non-targeted control strategies, by conducting multi-dimensional evaluation and root cause attribution of transmission efficiency, repair quality, and data content, and accurately mapping the root causes of problems to differentiated dynamic control commands.

[0102] Example 2 is based on Example 1. This example provides a data transmission system based on semantic compression and dynamic decision control, as shown in Figure 2. It includes a data transmission end and a data receiving end. The data transmission end is used to process the received data and transmit it to the data receiving end. It includes: an encoding module, configured to adaptively segment the acquired multi-dimensional data to be transmitted and fuse multi-scale spatiotemporal features to generate a first feature sequence, and to perform global semantic compression on the first feature sequence to generate a first hidden state sequence; and an adaptive transmission module, configured to generate an optimal transmission strategy instruction based on the semantic change rate of the first hidden state sequence and the network channel quality, combined with a dynamic decision function, and transmit the first hidden state sequence to the data receiving end. The hidden state sequence is encapsulated as an application layer data stream for data transmission. The data receiving end includes: a data self-healing module, configured to sequentially perform streaming parsing and multi-dimensional cross-validation on the received application layer data stream. If the validation passes, a second hidden state sequence is generated; if the validation fails, an abnormal hidden state sequence is generated. Guided semantic repair is then performed to generate the second hidden state sequence. The transmission strategy adaptive update module is configured to calculate the end-to-end transmission delay, semantic reconstruction confidence, and posterior anomaly score based on the second hidden state sequence, identify the link operation status of data transmission, and generate a corresponding data transmission strategy to be transmitted to the data sending end to achieve adaptive update of the transmission strategy.

[0103] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0104] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

[0105] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A data transmission method based on semantic compression and dynamic decision control, characterized in that, The process includes the following steps: The data sender adaptively segments the acquired multi-dimensional data to be transmitted and fuses multi-scale spatiotemporal features to generate a first feature sequence. Global semantic compression is then applied to the first feature sequence to generate a first hidden state sequence. Based on the semantic change rate of the first hidden state sequence and network channel quality, an optimal transmission strategy instruction is generated using a dynamic decision function. The first hidden state sequence is then encapsulated into an application layer data stream for data transmission. The data receiver sequentially performs streaming parsing and multi-dimensional cross-validation on the received application layer data stream. If the validation passes, a second hidden state sequence is generated; if the validation fails, an abnormal hidden state sequence is generated. Guided semantic repair is then performed to generate the second hidden state sequence. Based on the second hidden state sequence, end-to-end transmission delay, semantic reconstruction confidence, and posterior anomaly score are calculated to identify the link operation status of data transmission and generate a corresponding data transmission strategy, which is then transmitted to the data sender to achieve adaptive updating of the transmission strategy.

2. The data transmission method based on semantic compression and dynamic decision control as described in claim 1, characterized in that: The method for generating the first hidden state sequence Z includes the following steps: calculating the data flow fluctuation index in real time based on the original multidimensional data vector; performing adaptive segmentation on the original multidimensional data vector according to the composite logical condition composed of the data flow fluctuation index and two preset thresholds to generate a first data sequence; mapping the first data sequence to a second data sequence; generating absolute position encoding and relative timestamp encoding in parallel; weighting and fusing the two encodings with the second data sequence to generate a first feature sequence; inputting the first feature sequence into a lightweight Transformer encoder for global context capture and multi-layer information distillation; and semantically compressing the first feature sequence into a first hidden state sequence.

3. The data transmission method based on semantic compression and dynamic decision control as described in claim 2, characterized in that: The data flow fluctuation index is defined as the moving average of the Euclidean distance increment of the data vector over a set period.

4. The data transmission method based on semantic compression and dynamic decision control as described in claim 2, characterized in that: The method for generating the first feature sequence includes the following steps: generating the first data sequence... A linear transformation is performed to obtain a higher-dimensional feature matrix, resulting in a second data sequence. Position codes are generated based on the sequential order of the data sequence, and relative timestamp codes are generated based on the time difference between the current sequence and the starting point of the data. The position codes and relative timestamp codes are then combined. As a multi-scale time series feature, the second data sequence is weighted and fused with the multi-scale time series feature to generate the final first feature sequence.

5. The data transmission method based on semantic compression and dynamic decision control as described in claim 1, characterized in that: A method for encapsulating the first hidden state sequence into an application layer data stream based on the semantic change rate of the first hidden state sequence and the network channel quality, combined with a dynamic decision function to generate an optimal transmission strategy instruction, includes: collaboratively sensing the semantic change rate of the first hidden state sequence and the real-time acquired network channel quality, and generating an optimal transmission strategy instruction by combining a dynamic decision function; dynamically scheduling the first hidden state sequence through a dynamic transmission buffer driven by a strategy and combined with composite triggering conditions, and performing independent encapsulation or buffer aggregation operations according to the optimal transmission strategy instruction to construct the application layer data payload; configuring metadata headers for the application layer data payload and encapsulating it into an application layer data stream.

6. The data transmission method based on semantic compression and dynamic decision control as described in claim 5, characterized in that: Metadata Header It includes the following key fields: Global Sequence Identifier (Sequence_ID), Semantic Importance Level Load characteristics description Forward-looking processing instructions and sending timestamp Global sequence identifier Sequence_ID: will be determined by the start identifier. and end identifier The constructed tuple serves as a global sequence identifier, with a semantic importance level. : The semantic change rate of the first hidden state sequence The quantized value obtained by mapping is used as the semantic importance level. ； Load characteristic description It is a description of application layer data load Composite fields combining internal structure and data characteristics; forward-looking processing instructions : is determined by the edge sender based on semantic importance level and network quality indicators The enumerated values ​​generated by the comprehensive decision-making process.

7. The data transmission method based on semantic compression and dynamic decision control as described in claim 1, characterized in that: The received application layer data stream is sequentially parsed and cross-validated in multiple dimensions, as follows: For the application layer data stream, each application layer protocol data unit is segmented and recombined using a length prefix message delimitation method. The data payload is extracted and decoded into a first temporary hidden state sequence. Using the metadata header of the application layer data stream as a trusted benchmark, the first hidden temporary state sequence decoded in the previous stage is cross-validated in multiple dimensions, including semantic value verification, structural feature verification, and instruction content verification. If all verifications pass, the data is determined to be normal, a complete second hidden state sequence is generated, and its value is the first temporary hidden state sequence, and the process ends. If any verification fails, the current application layer protocol data unit is determined to have an application layer data anomaly, and an anomaly hidden state sequence marked with the anomaly location and type is generated. Alternatively, guided semantic repair can be performed to generate the second hidden state sequence, including replacing the hidden state vector positions marked as anomaly in the anomaly hidden state sequence with mask tokens, and jointly inputting the metadata header corresponding to the anomaly hidden state sequence into the Transformer decoder. The Transformer decoder collaboratively utilizes the global semantic information within the anomaly hidden state sequence and the semantic prior knowledge of the metadata header to generate the second hidden state sequence.

8. The data transmission method based on semantic compression and dynamic decision control as described in claim 1, characterized in that: The adaptive update method for the transmission strategy includes: calculating three indicators based on the second hidden state sequence, including end-to-end transmission delay, semantic reconstruction confidence, and posterior anomaly score; comprehensively evaluating the three indicators based on decision tree logic; attributing the root cause of the problem to one of the following based on whether each indicator exceeds a preset threshold: transmission bottleneck, repair uncertainty, or novel anomaly event, and outputting an attribution label; querying a predefined decision rule base based on the attribution label to generate a corresponding dynamic control signal; and transmitting the dynamic control signal back to the edge transmitter through a high-priority control channel, whereby the edge transmitter parses the dynamic control signal and updates the transmission strategy online.

9. A data transmission system based on semantic compression and dynamic decision control, comprising a data transmission end and a data receiving end, characterized in that: The data transmission end, used to process the received data and transmit it to the data receiving end, includes: an encoding module configured to adaptively segment the acquired multi-dimensional data to be transmitted and then fuse multi-scale spatiotemporal features to generate a first feature sequence, and to perform global semantic compression on the first feature sequence to generate a first hidden state sequence; and an adaptive transmission module configured to generate an optimal transmission strategy instruction based on the semantic change rate of the first hidden state sequence and the network channel quality, combined with a dynamic decision function, and to encapsulate the first hidden state sequence into an application layer data stream for data transmission.

10. The data transmission system based on semantic compression and dynamic decision control as described in claim 9, characterized in that: The data receiving end includes: a data self-healing module, configured to sequentially perform streaming parsing and multi-dimensional cross-validation on the received application layer data stream; if the validation passes, a second hidden state sequence is generated; if the validation fails, an abnormal hidden state sequence is generated, and guided semantic repair is performed to generate the second hidden state sequence; and a transmission strategy adaptive update module, configured to calculate end-to-end transmission latency, semantic reconstruction confidence, and posterior anomaly score based on the second hidden state sequence, identify the link operation status of data transmission, and generate a corresponding data transmission strategy to transmit to the data sending end to achieve adaptive update of the transmission strategy.

Images