Power Internet of Things Transmission Management Method and System Based on Intelligent Converged Terminal

Abstract

This invention discloses a power Internet of Things (IoT) transmission management method and system based on intelligent fusion terminals, belonging to the field of intelligent communication and scheduling technology for power IoT. The method includes: collecting multiple multi-source operating condition data, performing latent space mapping to generate multiple implicit preference vectors; performing instruction parsing to obtain heterogeneous constraint parameters; fusing the heterogeneous constraint parameters and multiple implicit preference vectors through cross-modal attention to generate a joint task intent embedding vector; inputting the joint task intent embedding vector into a diffusion model to generate a transmission strategy instance; and performing data transmission management in the target area. This application solves the technical problems in existing technologies where the power IoT cannot adapt to complex dynamic operating environments and lacks global collaborative scheduling for multi-terminal transmission, resulting in low communication efficiency, insufficient transmission reliability, and uneven resource allocation.

Description

Technical Field

[0001] This invention relates to the field of intelligent communication and dispatching technology for the power Internet of Things (IoT), specifically to a power IoT transmission management method and system based on intelligent converged terminals. Background Technology

[0002] The Internet of Things (IoT) for power systems is a core digital infrastructure for building new power systems. With the large-scale deployment of smart converged terminals on the power distribution and consumption side, the various services they support, such as distribution automation, electricity consumption information collection, and equipment status monitoring, generate massive amounts of heterogeneous data from multiple sources. These services place differentiated and dynamic demands on the latency, reliability, security, and energy consumption of data transmission, making traditional fixed transmission management models inadequate for the digital transformation needs of new power systems.

[0003] Existing transmission management technologies mostly employ strategies based on fixed rules or static models, such as preset modulation and coding schemes and resource allocation templates. These methods struggle to adapt to dynamically changing network conditions and diverse upper-layer task requirements in real time and with high precision. In particular, they fail to comprehensively utilize the complex patterns hidden within multi-source historical data, and their flexibility and overall optimization capabilities are insufficient when facing multiple constraints such as latency, bandwidth, and security. Therefore, there is an urgent need for an intelligent method that can deeply integrate environmental states and task intentions and dynamically generate optimal transmission strategies. Summary of the Invention

[0004] This application provides a power Internet of Things (IoT) transmission management method and system based on intelligent fusion terminals, aiming to solve the technical problems in the existing technology where the power IoT cannot adapt to complex dynamic operating environments and lacks global collaborative scheduling for multi-terminal transmission, resulting in low communication efficiency, insufficient transmission reliability, and uneven resource allocation.

[0005] In view of the above problems, this application provides a power Internet of Things transmission management method and system based on intelligent converged terminals.

[0006] The first aspect disclosed in this application provides a power Internet of Things (IoT) transmission management method based on intelligent fusion terminals. The method includes: collecting multiple multi-source operating condition data from multiple intelligent fusion terminals in a target area; performing latent space mapping to generate multiple implicit preference vectors; receiving transmission task instructions; performing instruction parsing to obtain heterogeneous constraint parameters; performing cross-modal attention fusion of the heterogeneous constraint parameters and the multiple implicit preference vectors to generate a joint task intent embedding vector; inputting the joint task intent embedding vector into a diffusion model; generating a transmission strategy instance based on the back-diffusion process of the diffusion model; wherein the transmission strategy instance includes dynamically generated encoding and modulation parameters, frame encapsulation format parameters, and communication resource mapping parameters; and performing data transmission management in the target area based on the transmission strategy instance.

[0007] Another aspect of this application discloses a power Internet of Things (IoT) transmission management system based on intelligent fusion terminals. This system includes: a working condition data acquisition module, used to acquire multiple multi-source working condition data from multiple intelligent fusion terminals in a target area, perform latent space mapping, and generate multiple implicit preference vectors; a constraint parameter acquisition module, used to receive transmission task instructions, perform instruction parsing, obtain heterogeneous constraint parameters, and perform cross-modal attention fusion of the heterogeneous constraint parameters and the multiple implicit preference vectors to generate a joint task intent embedding vector; a strategy instance generation module, used to input the joint task intent embedding vector into a diffusion model, and generate a transmission strategy instance based on the back diffusion process of the diffusion model, wherein the transmission strategy instance includes dynamically generated encoding and modulation parameters, frame encapsulation format parameters, and communication resource mapping parameters; and a data transmission management module, used to perform data transmission management in the target area based on the transmission strategy instance.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] A technical solution was adopted, which involves extracting the implicit transmission preference vector of intelligent fusion terminals, constructing heterogeneous multimodal constraint tensors, fusing cross-modal attention association of multi-terminal features, and dynamically generating adaptation strategies in conjunction with a diffusion model to coordinate regional data transmission management. This solution solves the technical problems of existing technologies being unable to adapt to the complex and dynamic operating environment of the power Internet of Things, lack of collaborative scheduling among multiple terminals, and rigid transmission methods with poor adaptability. It achieves the technical effects of global optimization of communication resource allocation, stable and reliable transmission, accurate matching of multiple constraints, and a significant improvement in the overall communication operation efficiency of the power Internet of Things.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0011] Figure 1 A flowchart illustrating the power Internet of Things transmission management method based on a smart fusion terminal is provided for embodiments of this application; Figure 2 A schematic diagram of the structure of a power Internet of Things transmission management system based on an intelligent fusion terminal is provided for the embodiments of this application.

[0012] Figure labeling: 11 Working condition data acquisition module, 12 Constraint parameter acquisition module, 13 Strategy instance generation module, 14 Data transmission management module. Detailed Implementation

[0013] The overall concept of the technical solution provided in this application is as follows: This application provides a method and system for power Internet of Things (IoT) transmission management based on intelligent fusion terminals. It extracts implicit transmission preference vectors from historical operating data of the intelligent fusion terminals, integrates multi-dimensional heterogeneous transmission constraints such as latency, bandwidth, security, and energy consumption, and then uses cross-modal attention fusion to combine global task requirements with the terminal's own characteristics. Finally, it intelligently generates adaptive transmission strategies through a diffusion model, thereby achieving refined control of data transmission across multiple terminals and adapting to the complex and dynamic communication environment of the power IoT.

[0014] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0015] Example 1 like Figure 1 As shown in the embodiments of this application, a power Internet of Things (IoT) transmission management method based on an intelligent converged terminal is provided. The method includes: Step S100: Collect multiple multi-source operating condition data from multiple intelligent fusion terminals in the target area, perform latent space mapping, and generate multiple implicit preference vectors.

[0016] Specifically, multi-source operational data refers to various types of operational data collected from intelligent fusion terminals, mainly including explicit channel parameters such as real-time signal-to-noise ratio and interference intensity, and transmission event logs such as historical modulation methods and packet loss records. Latent space mapping refers to using machine learning models, such as variational autoencoders, to compress and map high-dimensional, complex raw data into a low-dimensional, continuous latent space. Vectors in this space can capture the underlying essential patterns and abstract features in the data. Implicit preference vectors refer to a set of low-dimensional vectors obtained through latent space mapping. Each vector represents the implicit, difficult-to-observe regular preferences of a terminal device in its historical and current states, such as the time correlation of the channel, periodic patterns of interference, or potential trends in device performance degradation.

[0017] Specifically, explicit channel parameters and historical transmission event logs of each intelligent fusion terminal in the target area are collected and cached in real time, and arranged in chronological order to form a first time-series vector and a second time-series vector, respectively. The two sets of time-series vectors corresponding to each terminal are concatenated in the feature dimension to form a concatenated time-series vector that can simultaneously reflect instantaneous channel conditions and long-term behavioral patterns. The concatenated time-series vectors of all terminals are input one by one into a pre-trained variational autoencoder. The encoder calculates the probability distribution of each input in the latent space and performs random sampling accordingly, thereby generating a deterministic, low-dimensional implicit preference vector for each terminal.

[0018] This step transforms massive, high-dimensional, and heterogeneous raw operating data into dense, low-dimensional feature vectors that can profoundly characterize the internal state and behavioral patterns of the equipment, providing a computationally computable and fusionable high-quality information foundation for subsequent intelligent decision-making.

[0019] Step S200: Receive the transmission task instruction, execute instruction parsing, obtain heterogeneous constraint parameters, and perform cross-modal attention fusion of the heterogeneous constraint parameters and the multiple implicit preference vectors to generate a joint task intent embedding vector.

[0020] Specifically, the transmission task instruction is a control command issued by the power IoT cloud platform or edge gateway, containing data transmission requirements and constraints. Heterogeneous constraint parameters are a set of parameters composed of constraints of different data types and different physical meanings, including delay constraint parameters, bandwidth constraint parameters, security level parameters, and energy consumption budget parameters.

[0021] Specifically, the system receives transmission task instructions from the cloud platform or edge gateway, performs structured parsing on the instructions, and extracts four types of heterogeneous constraint parameters: latency constraint parameters, bandwidth constraint parameters, security level parameters, and energy consumption budget parameters. The latency constraint parameters, bandwidth constraint parameters, and energy consumption budget parameters are encoded into first, second, and third real tensors of their respective dimensions. The security level parameters are converted into discrete embedding tensors using an embedding matrix. These four tensors are stacked along the channel dimension to obtain a unified multimodal tensor as the heterogeneous constraint parameter. This heterogeneous constraint parameter is then copied into K copies equal to the number of implicit preference vectors. Each set of heterogeneous constraint parameters is concatenated with an implicit preference vector along the feature dimension to obtain K concatenated vectors representing the local state of the corresponding intelligent fusion terminal under global constraints. These K concatenated vectors are stacked into a two-dimensional feature matrix along the terminal dimension and input into a multi-head self-attention layer. By calculating the attention weight between every two concatenated vectors in the matrix, each concatenated vector adaptively fuses the state information of all other terminals, outputting K context enhancement vectors. Finally, these K context enhancement vectors are globally averaged to obtain a global feature vector, which is the joint task intent embedding vector.

[0022] This step achieves global collaborative fusion of different types of heterogeneous constraint parameters and individual transmission characteristics of multiple terminals, solving the technical problem in the existing technology that the task requirements and terminal status are disconnected and that global collaborative optimization of multiple terminals cannot be achieved. It provides accurate and unified condition input for the subsequent diffusion model to generate the optimal transmission strategy that satisfies global constraints.

[0023] Step S300: Embed the joint task intent into the vector input diffusion model, and generate a transmission strategy instance based on the reverse diffusion process of the diffusion model, wherein the transmission strategy instance includes dynamically generated coding and modulation parameters, frame encapsulation format parameters, and communication resource mapping parameters.

[0024] Specifically, the joint task intent embedding vector is a fixed-dimensional real-number vector, equal in dimension to the dimension of a single implicit preference vector. This vector encodes a synthesis of information: the global constraints of the current transmission task, the local operating characteristics of each intelligent fusion terminal, and the resource coordination relationships between terminals to satisfy the global constraints. This vector serves as the conditional input to the diffusion model, used to generate transmission strategy instances that simultaneously adapt to all terminal operating conditions and satisfy the global task constraints. The diffusion model is a generative artificial intelligence model that progressively adds Gaussian noise to real data through forward diffusion, and then progressively removes noise through back diffusion guided by the conditional vector, generating target data that meets the constraints. The forward diffusion process is the training process of the diffusion model, progressively adding Gaussian noise to historical optimal transmission strategy data until the data becomes pure noise, learning the distribution pattern of the noise. The back diffusion process is the inference process of the diffusion model, starting from pure noise and guided by the joint task intent embedding vector, progressively predicting and removing noise to generate a transmission strategy that meets the constraints. The transmission strategy instance refers to a dynamically generated complete data transmission scheme, including three core parameters: coding and modulation, frame encapsulation format, and communication resource mapping, which can be directly used to guide terminal transmission. The coding and modulation parameters determine the parameters of data coding and modulation, including the non-equal probability constellation diagram parameters and the sparse parity-check matrix parameters, which directly affect the transmission rate and anti-interference capability. The frame encapsulation format parameters determine the parameters of the data frame structure, including the message authentication code length and the frame header control field, which directly affect transmission security and compatibility. The communication resource mapping parameters determine the parameters of the communication resource allocation method, including the time slot allocation matrix, subcarrier frequency hopping sequence, and power allocation vector, which directly affect resource utilization and the probability of transmission collisions.

[0025] Specifically, the joint task intent embedding vector is used as a conditional vector input to the conditional diffusion model. An initial noise tensor with the same dimension and shape as the target transmission strategy instance is randomly sampled from a standard multivariate Gaussian distribution as the current noise data. Starting from the preset maximum time step T, T consecutive denoising steps are executed sequentially downwards. In each denoising step, a noise prediction network with a graph neural network as its backbone simultaneously receives the noise data at the current moment, the current time step identifier, and the joint task intent embedding vector as a condition. It predicts the noise components contained in the noise data at the current moment. The denoised data for this step is obtained by subtracting the predicted noise components from the current noise data, and this is used as the next denoising step. The input is repeated for the above denoising steps until all T steps are completed, and the final denoised data is used as the original data of the transmission strategy instance. The complex coordinate sequence and probability sequence are parsed from the first half of the vector of the original data and the non-equiprobability constellation diagram parameters are generated. The sparse parity-check matrix parameters are obtained by linear transformation from the second half of the vector of the original data. The two together constitute the coding and modulation parameters. At the same time, the first scalar value is extracted from the original data and rounded to obtain the message authentication code length. The binary string is extracted as the meta-control field to generate the frame encapsulation format parameters. The time slot allocation matrix, subcarrier frequency hopping sequence and power allocation vector are extracted to generate the communication resource mapping parameters, and a complete transmission strategy instance containing three types of core parameters is obtained.

[0026] The diffusion model consists of a time-step embedding module, a joint task intent conditional embedding module, and a noise prediction network with a graph neural network as its backbone. The time-step embedding uses sinusoidal positional encoding to convert discrete time steps into continuous vectors. The joint task intent conditional embedding maps the joint task intent embedding vector to the same feature dimension as the noise data through two ReLU-activated fully connected layers. The noise prediction network uses a stacked three-layer graph convolutional layer structure, with each layer followed by a normalization and ReLU activation function, effectively modeling the communication resource relationships between multiple intelligent fusion terminals. The input is the noise data at the current moment, the concatenation result of the time-step embedding vector and the conditional embedding vector, and the output is the predicted noise component. During training, a training set is constructed by collecting historically validated optimal transmission strategy instances from the target area's power IoT operation. For each training... The training samples undergo a forward diffusion process, starting with the original transmission strategy data from step 0. Gaussian noise is gradually added according to a preset linear noise scheduler until it becomes pure noise at step T. Simultaneously, the actual noise component added at each time step is recorded. The noise data, corresponding time step embedding, and joint task intent condition embedding for each time step are concatenated and input into the noise prediction network to predict the noise component at that time step. The mean squared error loss function is used to calculate the difference between the predicted noise and the actual noise as the total loss. The AdamW optimizer is used for iterative training with an initial learning rate of 1e-4, and a cosine annealing learning rate scheduler is used to dynamically adjust the learning rate. Training continues until the loss converges and the accuracy of the generated transmission strategy meets the constraints of latency, bandwidth, security, and energy consumption, reaching over 95%. The model weights are then frozen and can be used for the backward diffusion generation process in the inference stage.

[0027] This step employs a conditional diffusion model to achieve end-to-end dynamic generation of full-dimensional transmission strategies. It solves the technical problem in existing technologies where transmission strategies are pre-configured and fixed, and cannot simultaneously meet the global optimization of multiple constraints such as latency, bandwidth, security, and energy consumption. It can adaptively generate the optimal transmission scheme according to real-time task requirements and terminal status, significantly improving the reliability and resource utilization of data transmission in complex power scenarios.

[0028] Step S400: Based on the transmission strategy instance, perform data transmission management for the target area.

[0029] Specifically, the edge gateway receives and verifies the generated transmission policy instance, breaking it down into three parts: coding and modulation parameters, frame encapsulation format parameters, and communication resource mapping parameters. This information is then broadcast to all intelligent converged terminals within the target area via a dedicated control channel. Upon receiving the policy parameters, each terminal automatically configures its communication module parameters, writing the non-equiprobability constellation diagram and sparse parity-check matrix into the modulation and demodulation unit, the message authentication code length and meta-control field into the frame processing unit, and the time slot allocation matrix, subcarrier frequency hopping sequence, and power allocation vector into the resource scheduling unit. Finally, the edge gateway sends a time synchronization command to all terminals to calibrate each terminal. The local clock controls all terminals to start data transmission at the same time, occupy corresponding channel resources according to the time slot allocation matrix, switch communication frequencies according to the subcarrier frequency hopping sequence, and adjust the transmit power according to the power allocation vector. During transmission, the edge gateway receives transmission status data reported by each terminal in real time, continuously monitors key indicators such as transmission success rate, packet loss rate, and latency. When the transmission quality is detected to be lower than the preset threshold or a new transmission task instruction is received, the current transmission strategy is immediately terminated, triggering a new round of implicit preference vector extraction, cross-modal attention fusion, and transmission strategy generation process, thereby realizing closed-loop dynamic management of data transmission in the target area.

[0030] This step enables end-to-end closed-loop management of the transmission strategy from generation to execution, ensuring that the optimal transmission strategy can be accurately implemented. At the same time, through real-time status monitoring and dynamic adjustment mechanisms, it can quickly adapt to the time-varying characteristics of the power line carrier channel and changes in task requirements, significantly improving the stability, reliability, and resource utilization of data transmission in the target area.

[0031] Furthermore, the power Internet of Things transmission management method based on intelligent fusion terminals also includes: extracting multiple explicit channel parameter sets of multiple intelligent fusion terminals from multiple multi-source operating condition data, arranging them by time to obtain multiple first time-series vectors; extracting multiple transmission event log sets of multiple intelligent fusion terminals from multiple multi-source operating condition data, arranging them by time to obtain multiple second time-series vectors; mapping and concatenating multiple first time-series vectors and multiple second time-series vectors to obtain multiple concatenated time-series vectors, inputting them into the encoder of a variational autoencoder, outputting the mean and variance of multiple latent spaces, and sampling from multiple latent spaces to obtain multiple implicit preference vectors.

[0032] Specifically, the explicit channel parameter set refers to the set of parameters that can be directly measured or observed, reflecting the instantaneous state of the physical channel. It includes signal-to-noise ratio, interference intensity, and node online status at multiple consecutive time points. The implicit preference vector is the sampled point, a low-dimensional, dense real-number vector, which is a latent representation of the complex operating conditions of the terminal. The latent space is a low-dimensional continuous feature space obtained by variational autoencoders through unsupervised learning, where each point corresponds to an implicit semantic representation of the original high-dimensional data, capable of capturing the inherent statistical regularities of the data.

[0033] Specifically, explicit channel parameters for the most recent N time periods are extracted from the cache database of each intelligent fusion terminal, arranged in chronological order and normalized to form the first time-series vector of the terminal; at the same time, transmission event logs within the same time period are extracted, and the classification information is one-hot encoded and arranged in chronological order together with the numerical information to form the second time-series vector. For the k-th terminal, its first and second time-series vectors are concatenated end-to-end along the feature dimension to generate a longer concatenated time-series vector. Specifically, for each intelligent fusion terminal, a strict timestamp alignment operation is performed on the first and second time-series vectors within the same time window to ensure that the number of rows in the two time-series vectors is completely consistent and that each row corresponds to the parameters and logs at the same sampling time. All time steps within the time window are traversed, and for each time step, the single-step feature sub-vectors of the first and second time-series vectors are concatenated end-to-end along the feature dimension, with the single-step feature sub-vector of the second time-series vector directly appended to the end of the single-step feature sub-vector of the first time-series vector. All the concatenated single-step feature sub-vectors are rearranged in their original chronological order to form a two-dimensional concatenated time-series vector. The number of rows equals the total number of sampling time steps within the time window, and the number of columns equals the sum of the single-step feature dimensions of the first and second time-series vectors. This vector simultaneously integrates channel state and behavioral feedback. The concatenated temporal vectors of all terminals are sequentially input into the pre-trained variational autoencoder network. The encoder network calculates and outputs the mean vector in the latent space for each input vector. Sum of variance vectors Finally, using the reparameterization technique, a random noise sample is taken from the standard normal distribution. And through the formula: Calculate the terminal's final implicit preference vector. , where the operator It is the Hadamard product, which represents the element-wise multiplication of corresponding positions of two vectors of the same dimension. Therefore, in this formula... The operation utilizes the Hadamard product, combining σ, representing the intensity of uncertainty, with the random source ε. This is equivalent to injecting a random perturbation of corresponding intensity into each dimension of the latent variable, according to its degree of uncertainty. For example, the implicit preference vector of a terminal might be [0.12, -0.85, 1.34, 0.02, ...]. Its numerical value itself has no direct physical meaning, but the vector dimension may be implicitly associated with abstract characteristics such as channel memory duration, sensitivity to sudden interference, and signal-to-noise ratio attenuation rate.

[0034] The variational autoencoder used to extract the implicit preference vector adopts an end-to-end structure consisting of a temporal feature extraction layer, a fully connected hidden layer, and a dual-branch output layer. The input layer dimension is consistent with the dimension of the flattened concatenated temporal vector. The temporal feature extraction layer uses a single bidirectional GRU layer to extract the long-short-term temporal dependencies between channel state and transmission behavior. Subsequently, it connects to a two-layer ReLU-activated fully connected hidden layer for feature dimensionality reduction and nonlinear transformation. Two parallel fully connected output layers output the mean vector and log-variance vector of the latent space distribution, respectively. Note that the numerical instability problem caused by negative variance should be avoided. The latent space dimension is set according to the terminal feature complexity. 4-16 dimensional; During construction, more than 30 days of historical multi-source operating condition data from intelligent fusion terminals in the target area are collected to construct an unsupervised training set. After Z-score normalization preprocessing, the training data is input into the encoder. The mean and logarithmic variance of the encoder output are substituted into the reparameterization technique to generate implicit samples. The original input data is reconstructed through a symmetric decoder. The mean square error reconstruction loss and the KL divergence loss of the standard Gaussian distribution are jointly calculated as the total loss function. The Adam optimizer is used to iteratively train with an initial learning rate of 1e-3 until the loss converges. The encoder weights after training are frozen and can be used for real-time extraction of implicit preference vectors.

[0035] This step uses a variational autoencoder to extract implicit preference vectors from massive heterogeneous time-series data in an unsupervised manner. These vectors deeply characterize the unique operating modes and potential failure tendencies of each terminal. This transforms raw data, which is difficult to use directly for decision-making, into high-quality, low-dimensional situational representations, laying a precise data foundation for subsequent cross-terminal collaboration and dynamic strategy generation.

[0036] Furthermore, the explicit channel parameters include the signal-to-noise ratio (SNR), interference intensity, and node online status of high-speed power line carrier communication; the transmission event log includes the modulation and coding scheme used for transmission, the measured SNR, the distribution of packet loss locations, and the number of retransmissions; the implicit preference vector includes at least three of the following: channel memory kernel, interference periodic implicit phase, nonlinear attenuation exponent of SNR with distance, and precursor modes of equipment failure.

[0037] Specifically, the signal-to-noise ratio (SNR) of high-speed power line carrier communication is the ratio of carrier signal power to background noise power, measured in dB. A higher value indicates better channel quality. Interference intensity is the average power of electromagnetic interference from power equipment in the same frequency band, measured in dBm. A lower value indicates weaker interference. Node online status is a Boolean value representing the communication connection status between the intelligent fusion terminal and the master station; 1 indicates online, and 0 indicates offline. Modulation and coding scheme is the constellation modulation and channel coding combination used for data transmission, such as QPSK, 16QAM, and 64QAM, usually represented by integer codes. Packet loss location distribution refers to the set of sequence numbers of lost data frames in a single transmission, used to locate the time location of sudden channel interference. Channel memory kernel is a characteristic value representing the time correlation of channel state; a larger value indicates a more stable channel state. Interference periodicity latent phase represents the phase shift of power line frequency interference, which can be used to predict the peak time of interference, measured in radians. The nonlinear attenuation exponent of SNR with distance is a characteristic value representing the attenuation rate of high-speed power line carrier communication signal with transmission distance; a larger value indicates faster attenuation. The precursory modes of equipment failure characterize the implicit features of an impending failure in the terminal communication module; the higher the value, the higher the risk of failure.

[0038] Specifically, specific parameters are precisely parsed and structured from the raw data stream: real-time signal-to-noise ratio (SNR), interference intensity, and heartbeat signals are read from the physical layer driver to form explicit channel parameters; the modulation and coding scheme of each transaction record, the measured SNR at that time, the offset of packet loss within the frame, and the retransmission counter are parsed from the transport layer protocol stack log to form a transmission event log. After the variational autoencoder mapping process, these specific, multi-dimensional raw data generate implicit preference vectors, whose dimensions are no longer abstract values, but are interpretably associated with at least three deep physical modes mentioned above, such as channel memory kernel, periodic latent phase of interference, nonlinear attenuation exponent of SNR with distance, and precursory modes of equipment failure.

[0039] This step enhances the interpretability and relevance of subsequent strategy generation by explicitly associating the components of the implicit preference vector with potential patterns that have physical or engineering significance, providing accurate and traceable feature basis for adaptive matching of time-varying patterns in complex power communication scenarios.

[0040] Furthermore, the system receives a transmission task instruction, executes instruction parsing, and obtains heterogeneous constraint parameters, including: parsing delay constraint parameters, bandwidth constraint parameters, security level parameters, and energy consumption budget parameters from the transmission task instruction; encoding the delay constraint parameters as a first real tensor, encoding the bandwidth constraint parameters as a second real tensor, converting the security level parameters into discrete embedding tensors through an embedding matrix, and encoding the energy consumption budget parameters as a third real tensor; stacking the first real tensor, the second real tensor, the discrete embedding tensor, and the third real tensor in the channel dimension to obtain a multimodal tensor, and using the multimodal tensor as the heterogeneous constraint parameters.

[0041] Specifically, latency constraint parameters refer to the maximum end-to-end data transmission delay allowed by the task, typically measured in milliseconds; smaller values ​​indicate higher real-time requirements. Bandwidth constraint parameters refer to the minimum transmission bandwidth required by the task, typically measured in Mbps; larger values ​​indicate higher data transmission rate requirements. Security level parameters refer to the required communication security protection level for the task, which are discrete categorical variables; different levels correspond to different encryption strengths and authentication mechanisms. Energy consumption budget parameters refer to the maximum allowed terminal communication energy consumption for the task, typically measured in mW; smaller values ​​indicate stricter energy consumption restrictions. Real tensors are multidimensional arrays composed of continuous real numbers, used to represent the standardized form of numerical constraint parameters. Embedded matrices, obtained through pre-training, map discrete categorical variables to continuous low-dimensional vectors, addressing the problem that categorical variables cannot directly participate in neural network computations.

[0042] Specifically, the system receives standardized transmission task instructions from the cloud platform or edge gateway, performs field-level parsing on the instructions, and extracts four core heterogeneous constraint parameters: latency constraint parameters, bandwidth constraint parameters, security level parameters, and energy consumption budget parameters. The specific method involves: verifying the legality of the instructions, including checking the frame header identifier, CRC checksum, sender authentication information, and instruction version number, discarding illegal or incorrectly formatted instructions; parsing the header fields of the instructions to identify the task type, target terminal group, and instruction execution timestamp, locating the constraint parameter fields in the instruction body; sequentially extracting the maximum permissible end-to-end delay field value as the latency constraint parameter, the minimum transmission bandwidth requirement field value as the bandwidth constraint parameter, the communication security protection level field value as the security level parameter, and the maximum single-terminal communication energy consumption limit field value as the energy consumption budget parameter; validating the four types of extracted parameters, checking whether the parameters are within the system's preset reasonable value range, truncating or correcting parameters exceeding the range, or triggering an exception handling process, and outputting the four types of structured parameters that pass the verification. For example, the constraints of a certain power distribution automation remote signaling transmission task obtained by parsing are: latency ≤ 100ms, bandwidth ≥ 2Mbps, security level 3, and energy consumption ≤ 50mW. The extracted parameters are standardized and encoded. The latency constraint parameter 100ms is normalized to the range of 0-1000ms and encoded as a first real tensor with dimension 1 [0.1]. The bandwidth constraint parameter 2Mbps is normalized to the range of 0-10Mbps and encoded as a second real tensor with dimension 1 [0.2]. The security level parameter 3 is encoded using a pre-trained embedding matrix. The parameters are converted into discrete embedding tensors [0.3, 0.7, -0.2, 0.5], and the energy budget parameter 50mW is normalized to the range of 0-500mW and encoded into a third real tensor [0.1] with dimension 1. The first real tensor, the second real tensor, the discrete embedding tensor, and the third real tensor are stacked end-to-end in the channel dimension to obtain a multimodal tensor [0.1, 0.2, 0.3, 0.7, -0.2, 0.5, 0.1]. This multimodal tensor is used as the heterogeneous constraint parameter input for subsequent cross-modal attention fusion.

[0043] The embedding matrix adopts a two-dimensional dense matrix structure. The number of rows is equal to the total number of communication security levels preset in the power Internet of Things, usually four levels, corresponding to different levels from basic protection to core confidentiality protection. The number of columns is the preset embedding dimension, usually set to 4-8 dimensions to balance semantic expressive power and computational complexity. Each row vector of the matrix uniquely corresponds to a continuous low-dimensional embedding representation of a security level. During construction, the classification system and value range of security levels are determined according to the communication security specifications of the power industry. The embedding matrix is ​​randomly initialized using a Xavier uniform distribution and used as a trainable parameter for end-to-end joint training with a variational autoencoder, a cross-modal attention layer, and a diffusion model. During training, the parameters of the embedding matrix are continuously updated iteratively with the backpropagation of the total loss function, so that the row vectors corresponding to different security levels gradually learn the semantic features related to their protection strength, computational overhead, and transmission latency. After training, the weights of the embedding matrix are frozen, and then any discrete security level parameter can be quickly converted into the corresponding continuous discrete embedding tensor by looking up a table, realizing a unified representation of classification constraints and numerical constraints.

[0044] This step achieves a unified tensor representation of numerical and categorical heterogeneous constraint parameters, solving the technical problem that different types of constraints cannot be directly input into neural networks for fusion calculation. At the same time, the semantic information of classification constraints such as security level is preserved by embedding the matrix, and the numerical range of each constraint parameter is unified by normalization processing. This provides standardized and high-quality constraint input for subsequent cross-modal attention fusion with implicit preference vectors, significantly improving the constraint compliance of subsequent transmission strategy generation.

[0045] Furthermore, the heterogeneous constraint parameters and the multiple implicit preference vectors are fused across modal attention to generate a joint task intent embedding vector. This includes: copying the multimodal tensor corresponding to the heterogeneous constraint parameters into K copies, concatenating each copy with one of the multiple implicit preference vectors along the feature dimension to obtain K concatenated vectors, where each concatenated vector represents the local fusion state of the corresponding intelligent fusion terminal under the constraints of the heterogeneous constraint parameters, and K is the number of the multiple implicit preference vectors, where K is a positive integer; stacking the K concatenated vectors into a matrix and inputting it into a multi-head self-attention layer to obtain K context enhancement vectors, where each context enhancement vector represents the state of the corresponding intelligent fusion terminal after collaboration with other intelligent fusion terminals; averaging the K context enhancement vectors along the intelligent fusion terminal dimension to obtain a global vector, and using this global vector as the joint task intent embedding vector.

[0046] Furthermore, the multi-head self-attention layer calculates the attention weight between every two concatenation vectors in the K concatenation vectors, enabling each concatenation vector to fuse information from other concatenation vectors and outputting K context enhancement vectors.

[0047] Specifically, the process involves obtaining the multimodal tensor corresponding to the encoded heterogeneous constraint parameters and the implicit preference vectors of K intelligent fusion terminals. The multimodal tensor is then completely copied into K copies, each containing identical global task constraint information. Each copied multimodal tensor is concatenated with an implicit preference vector along its feature dimension, resulting in K concatenated vectors. For example, when K=3, the multimodal tensor is 7-dimensional, and each implicit preference vector is 4-dimensional, three 11-dimensional concatenated vectors are obtained, each representing the local state of the corresponding terminal under global constraints. These K concatenated vectors are stacked into a K-row two-dimensional feature matrix along the terminal dimension. This matrix is ​​input into a multi-head self-attention layer, which first linearly transforms the input matrix into a query matrix, a key matrix, and a value matrix, and then calculates the value of each of the K concatenated vectors. The attention weights of the dot product between two concatenated vectors are calculated. For example, the attention weight between the concatenated vector of terminal 1 and the concatenated vector of terminal 2 is 0.6, and the attention weight between the concatenated vector of terminal 1 and the concatenated vector of terminal 2 and terminal 3 is 0.4. The attention weight between the concatenated vector of terminal 2 and terminal 1 is 0.5, and the attention weight between the concatenated vector of terminal 2 and terminal 3 is 0.5. The attention weight between the concatenated vector of terminal 3 and terminal 1 is 0.4, and the attention weight between the concatenated vector of terminal 3 and terminal 2 is 0.6. The value matrix is ​​weighted and summed using these attention weights, so that each concatenated vector adaptively integrates the information of all other concatenated vectors, outputting K context enhancement vectors. Each context enhancement vector represents the global state of the corresponding terminal after collaboration with other terminals. The arithmetic mean of these K context enhancement vectors is performed on the dimension of the intelligent fusion terminal to obtain a global feature vector, which is the joint task intent embedding vector.

[0048] The weight calculation of the multi-head self-attention layer is a multi-branch parallel computation based on the dot product attention mechanism. The core is to quantify the state correlation between the concatenated vectors of K terminals, allowing each terminal to adaptively fuse useful information from other terminals. The specific calculation process is as follows: Step 1: The input is a two-dimensional matrix composed of K concatenated vectors stacked together. Where K is the number of terminals. This is the total dimension of a single concatenated vector. It is achieved through three independent trainable linear transformation matrices. The input matrix is ​​projected into a query matrix Q, a key matrix K, and a value matrix V, respectively: , , All three are shaped Where: each row of Q corresponds to the query feature of a terminal, used to find related information in other terminals; each row of K corresponds to the key feature of a terminal, used to be queried and matched by other terminals; each row of V corresponds to the value feature of a terminal, containing the actual state information that needs to be fused.

[0049] Step 2: Split Q, K, and V into h independent attention heads along the feature dimensions, where h is the number of heads, typically 2, 4, or 8. The feature dimensions of each head are... :

[0050] Each head , , Each head calculates its attention weight independently, without interfering with each other.

[0051] Step 3: For each attention head, calculate the dot product of the query matrix and the transpose of the key matrix to obtain the original attention score matrix. : The element in the m-th row and n-th column of the matrix , representing the original score of the state correlation between the m-th terminal and the n-th terminal, is the basis for the attention weight between every two concatenated vectors.

[0052] Step 4: Divide the original fraction by That is, scaling factor, to avoid When the value is too large, the dot product becomes too high, causing the softmax function to enter the gradient saturation region. .

[0053] Softmax normalization involves performing a softmax operation on each row of the scaled score matrix, transforming the scores in each row into a probability distribution between 0 and 1 that sums to 1, resulting in the final attention weight matrix. : in It refers to the weighting coefficient of the information of the nth terminal when the mth terminal merges information: the higher the weight, the greater the influence of the state of the nth terminal on the transmission strategy of the mth terminal.

[0054] Step 5: Align the attention weight matrix with the value matrix Perform a weighted summation to obtain the output matrix of the head. : , The m-th row is the feature vector of the m-th terminal after fusing all terminal information under this header.

[0055] Step 6: Concatenate the output matrices of the h attention heads along the feature dimensions to obtain the total output matrix. : Through a trainable linear transformation matrix Feature fusion is performed, ultimately outputting K context enhancement vectors, each corresponding to a global state after terminal collaboration: Context enhancement vector matrix = .

[0056] In the power Internet of Things (IoT) transmission scenario, attention weights automatically learn the actual relationships between terminals: for example, adjacent terminals with similar channel interference characteristics, terminals on the same power line, and terminals that experience packet loss at the same time will generate higher attention weights, thereby achieving global collaborative allocation of transmission resources and avoiding local resource conflicts.

[0057] This step achieves terminal-by-terminal binding of global task constraints and individual terminal characteristics through copying and splicing. It also realizes global collaborative information interaction among multiple terminals through a multi-head self-attention mechanism. This solves the technical problem in existing technologies where transmission strategies only consider the state of a single terminal and cannot achieve global resource collaborative optimization among multiple terminals. The generated joint task intent embedding vector integrates global task constraints and the transmission characteristics of all terminals, providing accurate and unified conditional input for the subsequent diffusion model to generate a globally optimal transmission strategy.

[0058] Furthermore, the joint task intent is embedded into the vector input diffusion model, and a transmission strategy instance is generated based on the back diffusion process of the diffusion model. This includes: randomly sampling an initial noise tensor with the same shape as the target output dimension from a standard Gaussian distribution, and using the initial noise tensor as the current noise data; using the joint task intent embedding vector as a condition vector, executing T denoising steps starting from step T, in each denoising step, a noise prediction network with a graph neural network as the backbone predicts the current noise component based on the noise data at the current time, the current time step, and the condition vector, where T is a positive integer; subtracting the predicted current noise component from the current noise data to obtain the denoised data, and using it as the input to the next denoising step; repeating the execution for T steps to obtain the final denoised data, and using it as the original data of the transmission strategy instance; extracting coding and modulation parameters from the original data of the transmission strategy instance; and extracting frame encapsulation format parameters and communication resource mapping parameters from the transmission strategy instance.

[0059] Specifically, the initial noise tensor refers to a pure noise tensor randomly sampled from a standard Gaussian distribution, completely consistent with the dimension of the target transmission strategy, serving as the starting point for the backdiffusion process. The target output dimension refers to the total dimension of the original data of the preset transmission strategy instance, determined by the sum of the dimensions of three types of parameters: coding / modulation, frame encapsulation format, and communication resource mapping. The conditional vector, i.e., the joint task intent embedding vector, is used to guide the diffusion model to generate a transmission strategy that conforms to global task constraints and terminal states. The time step T is the total number of steps in the backdiffusion process, typically set to 50-200 steps; more steps result in higher quality but also greater computational cost. The graph neural network backbone noise prediction network is a neural network with graph convolutional layers at its core, adept at modeling the communication resource relationships between multiple terminals, used to predict the noise component at each time step.

[0060] Specifically, based on the preset target output dimension, such as 128 dimensions, an initial noise tensor with a shape of 128 dimensions is randomly sampled from a standard Gaussian distribution as the current noise data. Using the previously generated joint task intent embedding vector as a conditional vector, 50 consecutive denoising steps are executed sequentially from the preset maximum time step T=50 downwards. In each denoising step, such as step 30, the noise prediction network with a graph neural network as its backbone simultaneously receives the 128-dimensional noise data at the current time, the embedding vector at step 30, and the joint task intent embedding vector as a condition, predicting the 128-dimensional noise component at the current time. The predicted noise component is subtracted element-wise from the current noise data to obtain the denoised data for this step, which is then used as the input for step 29. The above denoising steps are repeated until all 5 steps are completed. Step 0 yields the final 128-dimensional denoised data as the original data for the transmission strategy instance. Encoding and modulation parameters are extracted from the first 64 dimensions of the original data. The first 32 dimensions are parsed into complex coordinate sequences and probability sequences to generate non-equiprobability constellation diagram parameters, and the last 32 dimensions are linearly transformed to obtain sparse parity-check matrix parameters. Frame encapsulation format parameters are then extracted from the middle 16 dimensions of the original data. The scalar value of the 65th dimension is rounded to obtain the message authentication code length, and the binary strings from the 66th to 80th dimensions are used as the frame header control field. Communication resource mapping parameters are extracted from the last 48 dimensions of the original data. Dimensions 81 to 96 are parsed into an 8×2 time slot allocation matrix, dimensions 97 to 112 are parsed into a 16-bit subcarrier frequency hopping sequence, and dimensions 113 to 128 are parsed into a 16-dimensional power allocation vector, resulting in a complete transmission strategy instance containing three types of core parameters.

[0061] This step achieves progressive generation of transmission strategies through a step-by-step reverse diffusion process, ensuring the stability and diversity of the generated results. At the same time, it effectively models the communication resource relationships between multiple terminals using a graph neural network backbone, solving the technical problem in existing technologies where transmission strategies are pre-configured and fixed, and cannot simultaneously satisfy multi-constraint global optimization. It can generate end-to-end full-dimensional transmission strategies including coding and modulation, frame encapsulation, and resource allocation, significantly improving the reliability and resource utilization of data transmission in complex power scenarios.

[0062] Further, extracting coding and modulation parameters from the original data of the transmission strategy instance includes: extracting the first half of the vector from the original data of the transmission strategy instance, parsing the first half of the vector into a complex coordinate sequence and a probability sequence; assigning each complex coordinate in the complex coordinate sequence to a constellation point, and assigning each probability value in the probability sequence to a prior probability of a constellation point, to obtain a non-equiprobability constellation diagram, and using it as the constellation diagram parameter in the coding and modulation parameters; extracting the second half of the vector from the original data of the transmission strategy instance, performing a linear transformation on the second half of the vector to obtain a sparse parity-check matrix, and using the sparse parity-check matrix as the error correction code parameter in the coding and modulation parameters; and using the constellation diagram parameter and the error correction code as the coding and modulation parameters.

[0063] Specifically, the complex coordinate sequence is a sequence of multiple two-dimensional real number pairs, each corresponding to a point on the complex plane, used to represent the position of the signal point in the constellation diagram. The probability sequence is a sequence of multiple real numbers between 0 and 1, each real number corresponding to the prior probability of transmission for a constellation point, and the sum of all probability values ​​is 1. A constellation point refers to a discrete point on the complex plane, each point corresponding to a specific digital symbol, and is the basic unit of digital modulation. Unlike the traditional equal-probability constellation diagram, the transmission probabilities of each constellation point in a non-equal-probability constellation diagram are not equal. It can be optimized according to the channel interference characteristics, allocating higher probabilities to constellation points with stronger anti-interference capabilities. The constellation diagram parameters contain a set of parameters including the complex coordinates of all constellation points and the corresponding prior probabilities, determining the anti-interference capability and spectral efficiency of the modulation scheme. The error correction code parameters, i.e., the sparse parity-check matrix, determine the error detection and correction capability during data transmission. The coding modulation parameters are a set of parameters composed of constellation diagram parameters and error correction code parameters, which together determine the data transmission rate, reliability, and anti-interference capability.

[0064] Specifically, from the original data of the transmission strategy instance output by the diffusion model, the first half of the vector is extracted according to a predetermined dimensional division rule. This first half of the vector is then divided into two equal segments. The first segment is sequentially parsed into multiple complex coordinates, each represented by two consecutive real numbers representing the real and imaginary parts, respectively. The second segment is sequentially parsed into a corresponding number of initial probability values. Softmax normalization is performed on these initial probability values ​​to make the sum of all probability values ​​equal to 1, resulting in a probability sequence. Then, each complex coordinate is mapped to a constellation point on the complex plane, and each probability value corresponds to the prior transmission probability of that constellation point, generating a non-equiprobability constellation diagram adapted to the current channel interference characteristics, which is then used for coding and modulation. The parameters include constellation diagram parameters; then, the latter half of the vector is extracted from the original data of the transmission strategy instance, multiplied by the linear transformation matrix to obtain the intermediate feature vector of the corresponding dimension, and then the intermediate feature vector is reshaped into a two-dimensional matrix of a preset dimension. Binarization thresholding is performed on each element in the matrix, setting elements with absolute values ​​less than the threshold to 0 and elements greater than or equal to the threshold to 1, resulting in a sparse parity check matrix with most elements being 0, which is used as the error correction code parameter in the coding and modulation parameters; the generated non-equiprobability constellation diagram parameters and sparse parity check matrix parameters are combined to obtain the complete coding and modulation parameters, which can be directly distributed and configured to be used in the high-speed power line carrier communication modules of all intelligent fusion terminals in the target area.

[0065] The linear transformation matrix is ​​used to convert continuous vectors into sparse parity-check matrices. It is determined using end-to-end joint training combined with sparsity post-constraints. The specific process is as follows: The shape of the linear transformation matrix is ​​determined based on the input and output dimensions: Input dimension: equal to the length of the latter half of the original data vector in the transmission strategy instance; Output dimension: equal to the total number of elements in the target sparse parity-check matrix. Therefore, the shape of the linear transformation matrix is ​​input dimension × output dimension. A Xavier uniform distribution is used to randomly initialize the linear transformation matrix, ensuring consistent variance between the input and output in the initial state and avoiding gradient vanishing or exploding problems.

[0066] The linear transformation matrix is ​​used as a trainable parameter and is jointly trained end-to-end with the noise prediction network, time step embedding module, and conditional embedding module of the conditional diffusion model. During the training phase, for each historical best transmission strategy sample, its true sparse parity-check matrix is ​​first extracted and flattened into a one-dimensional vector. A forward diffusion process is performed to add noise to the original transmission strategy data to obtain noise data at each time step. The noise prediction network outputs predicted noise, which is then back-diffused to obtain the predicted value of the original transmission strategy data. The latter half of the predicted value is truncated and multiplied by the linear transformation matrix to obtain the flattened vector of the predicted sparse parity-check matrix. The binary cross-entropy loss between the predicted sparse parity-check matrix and the true sparse parity-check matrix is ​​calculated and added as an auxiliary loss to the total loss function of the diffusion model. The total loss = mean squared error noise loss + λ × binary cross-entropy loss, where λ is the weighting coefficient. All parameters of the linear transformation matrix and the diffusion model are updated simultaneously through backpropagation until the total loss converges. After training, based on the sparsity requirements of LDPC codes for power line carrier communication (typically ≥85%), the linear transformation matrix is ​​fine-tuned. This involves: calculating the average sparsity of the sparse parity-check matrix generated by the current linear transformation matrix; if the sparsity is lower than required, adding an L1 regularization term to the loss function to penalize non-zero elements in the matrix; and repeating the fine-tuning training with a few iterations until the generated sparse parity-check matrix meets the preset sparsity requirements. The final determined weights of the linear transformation matrix are then frozen and deployed to the edge gateway along with the diffusion model. This is used during the inference phase to convert the continuous vectors output by the diffusion model into a compliant sparse parity-check matrix.

[0067] This step enables precise parsing from the continuous abstract vector output by the diffusion model to directly deployable physical layer coding and modulation parameters. It supports the dynamic generation of optimal non-equiprobability constellation diagrams and adaptive sparse check codes based on real-time channel characteristics and multi-task constraints. It solves the technical problems that traditional fixed coding and modulation methods cannot adapt to the time-varying pulse interference characteristics of power line carriers and cannot simultaneously meet the multi-objective requirements of transmission rate, reliability and energy consumption. It significantly improves the anti-interference capability and spectrum resource utilization of data transmission in complex power distribution scenarios.

[0068] Further, extracting frame encapsulation format parameters and communication resource mapping parameters from the transmission strategy instance includes: extracting a first scalar value from the original data, rounding the first scalar value to obtain the message authentication code length; extracting a binary string from the original data as a meta-control field, and using the meta-control field as the frame header control field in the frame encapsulation format parameters; using the message authentication code length as the authentication code length field in the frame encapsulation format parameters, and using the meta-control field as the frame header control field in the frame encapsulation format parameters to obtain the frame encapsulation format parameters; extracting a time slot allocation matrix, a subcarrier frequency hopping sequence, and a power allocation vector from the original data; using the time slot allocation matrix as a time slot allocation scheme, the subcarrier frequency hopping sequence as a frequency hopping scheme, and the power allocation vector as a power allocation scheme, and outputting the communication resource mapping parameters.

[0069] Specifically, the first scalar value refers to a single consecutive value at a pre-specified position in the original data, specifically used to encode the message authentication code length. The message authentication code length refers to the number of bytes in the message authentication code at the end of the data frame, determining the integrity verification strength of data transmission; a larger value indicates higher security but also higher transmission overhead. The meta-control field refers to the set of fields in the frame header used to identify core control information such as frame type, transmission priority, target address range, and verification method; it is the key basis for the terminal to parse the data frame. The frame encapsulation format parameters are a set of parameters that determine the overall structure of the data frame, composed of the message authentication code length and the frame header meta-control fields, directly affecting transmission security, compatibility, and parsing efficiency. The time slot allocation matrix is ​​a two-dimensional binary matrix; rows correspond to intelligent fusion terminals, and columns correspond to time slots. An element of 1 indicates that the terminal occupies the channel in that time slot, and 0 indicates idle; it is used to implement time division multiple access (TDMA) channel resource allocation. The subcarrier frequency hopping sequence is a discrete sequence composed of subcarrier numbers. The terminal switches between different subcarriers in different time slots according to this sequence to combat frequency-selective fading and fixed-frequency interference in power line carrier communication. The power allocation vector is a one-dimensional continuous vector, where each element corresponds to the transmit power of a subcarrier or time slot. It is used to optimize transmission performance and reduce the energy consumption of terminal communication.

[0070] Specifically, from the original data of the transmission strategy instance output by the diffusion model, the first scalar value is extracted according to a predetermined dimensional offset. For example, when the total dimension of the original data is 128, the continuous value of the 65th dimension is extracted as the first scalar value. This value is linearly mapped to the legal value range of 4-16 bytes and then rounded to obtain the message authentication code length. The next 15 consecutive values ​​are extracted from the original data. Binarization thresholding is performed on each value, setting values ​​greater than or equal to 0.5 to 1 and values ​​less than 0.5 to 0, resulting in a 15-bit binary string as the meta-control field. For example, the binary string 010010110010101 corresponds to a data frame type, high transmission priority, a destination address of broadcast, and a CRC32 verification method. The obtained message authentication code length is then used as... The authentication code length field in the frame encapsulation format parameters is used as the frame header control field in the frame encapsulation format parameters. The complete frame encapsulation format parameters are then combined and output. Next, communication resource mapping parameters are extracted from the remaining dimensions of the original data. The next 16 values ​​are extracted and reshaped into an 8x2 two-dimensional matrix. The matrix elements are binarized to obtain the time slot allocation matrix. The next 16 values ​​are extracted and binarized to obtain a 16-bit subcarrier frequency hopping sequence. The last 16 values ​​are extracted and normalized to the legal power range of 0-20dBm to obtain the power allocation vector. The time slot allocation matrix, the subcarrier frequency hopping sequence, and the power allocation vector are used as the power allocation scheme. The complete communication resource mapping parameters are then combined and output.

[0071] The process of extracting the first scalar value from the original data of the transmission strategy instance and finally obtaining the message authentication code length is a deterministic process based on predefined dimension offset, continuous value range mapping, and integerization. There are no random operations throughout the process, ensuring the consistency of parameter extraction between the training and inference phases. The specific steps are as follows: Before training the diffusion model, the dimensions of the original data of the transmission strategy instance are divided, and the continuous vector intervals corresponding to each type of parameter are defined: the total dimension of the original data = the dimension of coding and modulation parameters + the dimension of frame encapsulation format parameters + the dimension of communication resource mapping parameters; the first scalar value occupies the first dimension of the frame encapsulation format parameter interval, that is, its position is: the dimension of coding and modulation parameters + 1; this dimension division rule is completely consistent in the training and inference phases, ensuring that the parameter encoding learned by the diffusion model matches the subsequent extraction rules.

[0072] The specific extraction steps are as follows: First, locate the dimensional offset position and calculate the index position of the first scalar value in the original data according to predefined dimensional partitioning rules. For example, if the encoding and modulation parameters occupy the first 64 dimensions, the first scalar value corresponds to the 65th element of the original data, with the index starting from 1. Second, extract the original continuous numerical value. From the corresponding index position in the original data, directly extract the single floating-point number at that position as the original scalar value. This value is a continuous value generated by the back-diffusion process of the diffusion model, following the distribution learned during training, typically falling within ±3σ of a standard Gaussian distribution, i.e., between -3 and 3. Third, linearly map to the legal value range. The original continuous value is mapped to the legal range of message authentication code length specified by the power IoT communication protocol through a linear transformation, typically 4-16 bytes. The mapping formula is:

[0073] in, : The extracted raw continuous values; , The default range for the original value is usually -3 and 3; , The minimum and maximum valid length of the message authentication code, typically 4 and 16; : The mapped floating-point length value; perform a rounding operation on the mapped floating-point length value to obtain the final integer message authentication code length. If the rounded result exceeds the valid range, it will be automatically truncated to the nearest boundary value, such as 4 if less than 4, and 16 if greater than 16.

[0074] This step enables precise parsing from the continuous abstract vectors output by the diffusion model to directly deployable link layer frame format parameters and physical layer resource allocation parameters. It supports dynamic adjustment of message authentication code length and resource allocation scheme based on task security level and channel status, solving the technical problem that traditional fixed frame format and static resource allocation methods cannot adapt to the time-varying interference characteristics of power line carriers and the differentiated needs of multiple tasks. It significantly improves the security, flexibility and spectrum resource utilization of power Internet of Things data transmission.

[0075] In summary, the power Internet of Things transmission management method based on intelligent converged terminals provided in this application has the following technical effects: 1. A technical solution is adopted to first extract explicit channel parameters and transmission event logs from intelligent fusion terminals and generate implicit preference vectors, then perform cross-modal attention fusion of heterogeneous constraint parameters and implicit preference vectors from multiple terminals, and finally generate transmission strategies through conditional diffusion models. This solution solves the technical problem that traditional fixed transmission strategies cannot adapt to the time-varying channel characteristics of power line carriers and the global optimization requirements of multiple constraints, and significantly improves the reliability and resource utilization of data transmission in complex power distribution scenarios.

[0076] 2. By employing a technique that encodes numerical constraint parameters into real tensors, converts categorized security level parameters into discrete embedded tensors via embedding matrices, and then stacks these tensors into multimodal tensors along the channel dimension, the technical problem of not being able to directly input heterogeneous constraints of different types into neural networks for fusion calculation is solved, providing standardized and high-quality constraint inputs for subsequent cross-modal attention fusion.

[0077] 3. By calculating the state-related attention weights between every two terminal splicing vectors through a multi-head self-attention layer, the feature vector of each terminal can adaptively fuse the state information of all other terminals. This solves the resource conflict and local optima problem caused by the transmission strategy only considering the state of a single terminal in the existing technology, and realizes the global collaborative optimization allocation of multi-terminal communication resources.

[0078] Example 2 Based on the same inventive concept as the power Internet of Things transmission management method based on intelligent fusion terminals in the foregoing embodiments, such as Figure 2 As shown in the embodiment of this application, a power Internet of Things transmission management system based on intelligent fusion terminals is provided. The system includes: a working condition data acquisition module 11, used to collect multiple multi-source working condition data from multiple intelligent fusion terminals in a target area, perform latent space mapping, and generate multiple implicit preference vectors; a constraint parameter acquisition module 12, used to receive transmission task instructions, perform instruction parsing, obtain heterogeneous constraint parameters, and perform cross-modal attention fusion of the heterogeneous constraint parameters and the multiple implicit preference vectors to generate a joint task intent embedding vector; a strategy instance generation module 13, used to input the joint task intent embedding vector into a diffusion model, and generate a transmission strategy instance based on the back diffusion process of the diffusion model, wherein the transmission strategy instance includes dynamically generated encoding and modulation parameters, frame encapsulation format parameters, and communication resource mapping parameters; and a data transmission management module 14, used to perform data transmission management in the target area based on the transmission strategy instance.

[0079] Furthermore, the operating condition data acquisition module 11 is also used to perform the following steps: extracting multiple explicit channel parameter sets of multiple intelligent fusion terminals from multiple multi-source operating condition data, and arranging them by time to obtain multiple first time-series vectors; extracting multiple transmission event log sets of multiple intelligent fusion terminals from multiple multi-source operating condition data, and arranging them by time to obtain multiple second time-series vectors; mapping and concatenating multiple first time-series vectors and multiple second time-series vectors to obtain multiple concatenated time-series vectors, inputting them into the encoder of the variational autoencoder, outputting the mean and variance of multiple latent spaces, and sampling from multiple latent spaces to obtain multiple implicit preference vectors.

[0080] Furthermore, the operating condition data acquisition module 11 is also used to perform the following steps: the explicit channel parameters include the signal-to-noise ratio, interference intensity, and node online status of high-speed power line carrier communication; the transmission event log includes the modulation and coding scheme used for transmission, the measured signal-to-noise ratio, the distribution of packet loss locations, and the number of retransmissions; the implicit preference vector includes at least three of the following: channel memory kernel, interference periodic implicit phase, nonlinear attenuation exponent of signal-to-noise ratio with distance, and precursor modes of equipment failure.

[0081] Furthermore, the constraint parameter acquisition module 12 is also used to perform the following steps: parsing the delay constraint parameter, bandwidth constraint parameter, security level parameter, and energy consumption budget parameter from the transmission task instruction; encoding the delay constraint parameter as a first real tensor, encoding the bandwidth constraint parameter as a second real tensor, converting the security level parameter into a discrete embedding tensor through an embedding matrix, and encoding the energy consumption budget parameter as a third real tensor; stacking the first real tensor, the second real tensor, the discrete embedding tensor, and the third real tensor in the channel dimension to obtain a multimodal tensor, and using the multimodal tensor as the heterogeneous constraint parameter.

[0082] Furthermore, the constraint parameter acquisition module 12 is also used to perform the following steps: copying the multimodal tensor corresponding to the heterogeneous constraint parameters into K copies, concatenating each copy with one of the multiple implicit preference vectors in the feature dimension to obtain K concatenated vectors, wherein each concatenated vector represents the local fusion state of the corresponding intelligent fusion terminal under the constraint of the heterogeneous constraint parameters, K is the number of the multiple implicit preference vectors, and K is a positive integer; stacking the K concatenated vectors into a matrix, inputting it into a multi-head self-attention layer to obtain K context enhancement vectors, wherein each context enhancement vector represents the state of the corresponding intelligent fusion terminal after collaboration with other intelligent fusion terminals; averaging the K context enhancement vectors in the intelligent fusion terminal dimension to obtain a global vector, and using the global vector as the joint task intent embedding vector.

[0083] Furthermore, the constraint parameter acquisition module 12 is also used to perform the following steps: the multi-head self-attention layer calculates the attention weight between every two splicing vectors in the K splicing vectors, so that each splicing vector fuses the information of other splicing vectors, and outputs K context enhancement vectors.

[0084] Furthermore, the strategy instance generation module 13 is also used to perform the following steps: randomly sample an initial noise tensor with the same shape as the target output dimension from a standard Gaussian distribution, and use the initial noise tensor as the current noise data; use the joint task intent embedding vector as the condition vector, and execute T denoising steps starting from step T, in each denoising step, a noise prediction network with a graph neural network as the backbone predicts the current noise component based on the noise data at the current time, the current time step, and the condition vector, where T is a positive integer; subtract the predicted current noise component from the current noise data to obtain the denoised data, and use it as the input of the next denoising step; repeat the execution of T steps to obtain the final denoised data, and use it as the original data of the transmission strategy instance; extract coding and modulation parameters from the original data of the transmission strategy instance; extract frame encapsulation format parameters and communication resource mapping parameters from the transmission strategy instance.

[0085] Furthermore, the strategy instance generation module 13 is also used to perform the following steps: extracting the first half of the vector from the original data of the transmission strategy instance, parsing the first half of the vector into a complex coordinate sequence and a probability sequence; assigning each complex coordinate in the complex coordinate sequence to a constellation point, and assigning each probability value in the probability sequence to a prior probability of a constellation point, to obtain a non-equiprobability constellation diagram, and using it as the constellation diagram parameter in the coding and modulation parameters; extracting the second half of the vector from the original data of the transmission strategy instance, performing a linear transformation on the second half of the vector to obtain a sparse parity-check matrix, and using the sparse parity-check matrix as the error correction code parameter in the coding and modulation parameters; and using the constellation diagram parameter and the error correction code as the coding and modulation parameters.

[0086] Furthermore, the policy instance generation module 13 is also used to perform the following steps: extracting a first scalar value from the original data, and rounding the first scalar value to obtain the message authentication code length; extracting a binary string from the original data as a meta-control field, and using the meta-control field as the frame header control field in the frame encapsulation format parameters; using the message authentication code length as the authentication code length field in the frame encapsulation format parameters, and using the meta-control field as the frame header control field in the frame encapsulation format parameters to obtain the frame encapsulation format parameters; extracting a time slot allocation matrix, a subcarrier frequency hopping sequence, and a power allocation vector from the original data; using the time slot allocation matrix as a time slot allocation scheme, the subcarrier frequency hopping sequence as a frequency hopping scheme, and the power allocation vector as a power allocation scheme, and outputting the communication resource mapping parameters.

[0087] In summary, any step of the method described above can be stored as a computer instruction or program in an unrestricted computer memory, and can be called and identified by an unrestricted computer processor to implement any method in the embodiments of this application, without any additional restrictions.

[0088] Furthermore, the "first" or "second" mentioned above may not only represent a sequential relationship, but may also represent a specific concept, and / or refer to the individual or collective selection of multiple elements. Clearly, those skilled in the art can make various modifications and variations to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A power Internet of Things transmission management method based on intelligent converged terminals, characterized in that, include: Collect multiple multi-source operating condition data from multiple intelligent fusion terminals in the target area, perform latent space mapping, and generate multiple implicit preference vectors; Receive transmission task instructions, execute instruction parsing, obtain heterogeneous constraint parameters, and perform cross-modal attention fusion of the heterogeneous constraint parameters and the multiple implicit preference vectors to generate a joint task intent embedding vector; The joint task intent is embedded into the vector input diffusion model, and a transmission strategy instance is generated based on the reverse diffusion process of the diffusion model. The transmission strategy instance includes dynamically generated coding and modulation parameters, frame encapsulation format parameters, and communication resource mapping parameters. Based on the aforementioned transmission strategy instance, data transmission management of the target area is performed; Multiple explicit channel parameter sets of multiple intelligent fusion terminals are extracted from multiple multi-source operating condition data, and arranged in time to obtain multiple first time-series vectors; Multiple transmission event log sets of multiple intelligent fusion terminals are extracted from multiple multi-source operating condition data, and arranged by time to obtain multiple second time-series vectors; By iterating through multiple first-time-series vectors and multiple second-time-series vectors and mapping and concatenating them, multiple concatenated time-series vectors are obtained. These concatenated time-series vectors are then input into the encoder of the variational autoencoder, which outputs the mean and variance of multiple latent spaces. Finally, multiple latent preference vectors are obtained by sampling from these latent spaces. The heterogeneous constraint parameters and the multiple implicit preference vectors are fused across modal attention to generate a joint task intent embedding vector, including: The multimodal tensor corresponding to the heterogeneous constraint parameters is copied into K copies. Each copy is concatenated with one of the multiple implicit preference vectors in the feature dimension to obtain K concatenated vectors. Each concatenated vector represents the local fusion state of the corresponding intelligent fusion terminal under the constraint of the heterogeneous constraint parameters. K is the number of the multiple implicit preference vectors and K is a positive integer. The K concatenated vectors are stacked into a matrix and input into a multi-head self-attention layer to obtain K context enhancement vectors, where each context enhancement vector represents the state of the corresponding intelligent fusion terminal after collaboration with other intelligent fusion terminals; The K context enhancement vectors are averaged over the intelligent fusion terminal dimension to obtain a global vector, which is then used as the joint task intent embedding vector. The joint task intent is embedded into a vector input diffusion model, and a transmission strategy instance is generated based on the back diffusion process of the diffusion model, including: Randomly sample an initial noise tensor with the same shape as the target output dimension from a standard Gaussian distribution, and use the initial noise tensor as the current noise data; The joint task intent embedding vector is used as a condition vector, and T denoising steps are executed from step T downwards. In each denoising step, the noise prediction network with graph neural network as the backbone predicts the current noise component based on the noise data at the current time, the current time step and the condition vector, where T is a positive integer. Subtract the predicted current noise component from the current noise data to obtain the denoised data, and use it as the input for the next denoising step. After repeating the process for T steps, the final denoised data is obtained and used as the original data for the transmission strategy instance. Extract coding and modulation parameters from the raw data of the transmission strategy instance; Extract the frame encapsulation format parameters and communication resource mapping parameters from the transmission strategy instance.

2. The power Internet of Things transmission management method based on intelligent converged terminals as described in claim 1, characterized in that, The explicit channel parameters include the signal-to-noise ratio, interference intensity, and node online status of high-speed power line carrier communication; The transmission event log includes the modulation and coding scheme used in the transmission, the measured signal-to-noise ratio, the distribution of packet loss locations, and the number of retransmissions; The implicit preference vector includes at least three of the following: channel memory kernel, interference periodic implicit phase, nonlinear attenuation exponent of signal-to-noise ratio with distance, and precursor mode of equipment failure. Among them, the channel memory kernel represents the characteristic value of the time correlation of the channel state. The larger the value, the more stable the channel state. The interference periodic implicit phase represents the phase offset of power line frequency interference and is used to predict the peak time of interference. The unit is radians.

3. The power Internet of Things transmission management method based on intelligent converged terminals as described in claim 1, characterized in that, Receive transmission task instructions, execute instruction parsing, and obtain heterogeneous constraint parameters, including: The latency constraint parameters, bandwidth constraint parameters, security level parameters, and energy consumption budget parameters are parsed from the transmission task instructions. The delay constraint parameter is encoded as a first real tensor, the bandwidth constraint parameter is encoded as a second real tensor, the security level parameter is converted into a discrete embedding tensor through an embedding matrix, and the energy consumption budget parameter is encoded as a third real tensor. The first real tensor, the second real tensor, the discrete embedding tensor, and the third real tensor are stacked in the channel dimension to obtain a multimodal tensor, which is then used as the heterogeneous constraint parameter.

4. The power Internet of Things transmission management method based on intelligent converged terminals as described in claim 1, characterized in that, The multi-head self-attention layer calculates the attention weight between every two concatenated vectors in the K concatenated vectors, so that each concatenated vector integrates the information of other concatenated vectors and outputs K context enhancement vectors.

5. The power Internet of Things transmission management method based on intelligent converged terminals as described in claim 1, characterized in that, Extracting coding and modulation parameters from the raw data of the transmission strategy instance includes: Extract the first half of the vector from the original data of the transmission strategy instance, and parse the first half of the vector into a complex coordinate sequence and a probability sequence; Each complex coordinate in the complex coordinate sequence corresponds to a constellation point, and each probability value in the probability sequence corresponds to the prior probability of a constellation point, thus obtaining a non-equal probability constellation diagram, which is then used as the constellation diagram parameter in the coding and modulation parameters. The latter half of the vector is extracted from the original data of the transmission strategy instance, and the latter half of the vector is linearly transformed to obtain a sparse parity check matrix. The sparse parity check matrix is ​​used as the error correction code parameter in the coding and modulation parameters. The constellation diagram parameters and error correction codes are used as the coding and modulation parameters.

6. The power Internet of Things transmission management method based on intelligent converged terminals as described in claim 1, characterized in that, Extracting frame encapsulation format parameters and communication resource mapping parameters from the transmission strategy instance includes: Extract the first scalar value from the original data, and then round the first scalar value to obtain the message authentication code length; Extract a binary string from the raw data as a meta-control field, and use the meta-control field as the frame header control field in the frame encapsulation format parameters; The message authentication code length is used as the authentication code length field in the frame encapsulation format parameters, and the meta control field is used as the frame header control field in the frame encapsulation format parameters to obtain the frame encapsulation format parameters. Extract the time slot allocation matrix, subcarrier frequency hopping sequence, and power allocation vector from the raw data; The communication resource mapping parameters are output by using the time slot allocation matrix as the time slot allocation scheme, the subcarrier frequency hopping sequence as the frequency hopping scheme, and the power allocation vector as the power allocation scheme.

7. A power Internet of Things transmission management system based on intelligent converged terminals, characterized in that, The system is used to execute the power Internet of Things transmission management method based on a smart fusion terminal as described in any one of claims 1 to 6, the system comprising: The working condition data acquisition module is used to collect multiple multi-source working condition data from multiple intelligent fusion terminals in the target area, perform implicit space mapping, and generate multiple implicit preference vectors. The constraint parameter acquisition module is used to receive transmission task instructions, perform instruction parsing, obtain heterogeneous constraint parameters, and perform cross-modal attention fusion of the heterogeneous constraint parameters and the multiple implicit preference vectors to generate a joint task intent embedding vector. The strategy instance generation module is used to embed the joint task intent into the vector input diffusion model and generate a transmission strategy instance based on the back diffusion process of the diffusion model. The transmission strategy instance includes dynamically generated coding and modulation parameters, frame encapsulation format parameters, and communication resource mapping parameters. The data transmission management module is used to manage data transmission in the target area based on the transmission strategy instance.

Images