Electric power communication system scheduling method and device

Abstract

The embodiment of the invention provides an electric power communication system scheduling method and device. The method comprises the following steps: acquiring system parameters from text information, image information and voice information; performing multi-modal fusion processing on the system parameters to obtain multi-modal fusion features; constructing a reward function based on a preset objective function by taking the resource state of the base station as a state and taking the power allocated by the base station for the equipment and the service quality weight of the equipment as actions according to the multi-modal fusion features, and determining an optimal power allocation strategy based on a reinforcement learning method; wherein the objective function is constructed based on service quality requirements of the equipment and allocable resources of the base station. According to the method, multi-modal system parameters can be effectively fused, the accuracy and efficiency of resource scheduling decision making are improved, system resources and service quality requirements are integrated, resource scheduling of the electric power communication system is optimized, and system performance and service quality are improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of power communication systems, and in particular to a method and device for scheduling a power communication system. Background Art

[0002] As an important support for the power system, the power communication system is directly related to the safe and stable operation of the entire power network. The traditional power communication system dispatching relies on manual experience operation, which is difficult to cope with complex and changing systems and prone to dispatching errors, affecting the stability and economic benefits of power supply. With the development of technologies such as big data and artificial intelligence, how to apply new technologies to the power communication system and optimize system dispatching is an urgent problem to be solved in the current power communication field. Summary of the invention

[0003] In view of this, an object of the embodiments of the present application is to propose a method and device for scheduling a power communication system to solve the scheduling problem of the power communication system.

[0004] Based on the above purpose, the embodiment of the present application provides a power communication system scheduling method, including:

[0005] Obtain system parameters from text information, image information and voice information;

[0006] Performing multimodal fusion processing on the system parameters to obtain multimodal fusion features;

[0007] According to the multimodal fusion characteristics, taking the resource status of the base station as the state, the power allocated by the base station to the device and the service quality weight of the device as the action, a reward function is constructed based on a preset objective function, and the optimal power allocation strategy is determined based on a reinforcement learning method; wherein, the objective function is constructed based on the service quality requirements of the device and the allocatable resources of the base station.

[0008] Optionally, the resource status of the base station is expressed as:

[0009]

[0010] Among them, s k is the state of base station k, T k is the throughput of base station k, N k is the number of devices connected to base station k, is the device N connected to the base station k i The service quality requirements are determined according to the service quality weight and service quality satisfaction of the equipment. is the device N connected to the base station k i The average satisfaction, θ μ is the satisfaction threshold of the equipment, Ka kIndicates that all base stations in the base station set K take their own actions a k The throughput obtained.

[0011] Optionally, the objective function is expressed as:

[0012]

[0013] ikB k ≤RB max

[0014]

[0015] Among them, RB k is the number of resource blocks that can be allocated to base station k, RB max is the total number of resource blocks; P is the power allocated by base station k to the device, W is the service quality weight of the device, and T(P, W) is the total throughput T of the base station under the control of parameters P and W.

[0016] Optionally, the reward function is:

[0017]

[0018] in, For state s k Take action a k The reward value obtained, A1 and A2 are set constants.

[0019] Optionally, the service quality satisfaction of the device is expressed as:

[0020]

[0021] Among them, P RSRP is the received reference signal power, D is the delay, N devices connected to the base station Δ i The throughput of the received reference signal is ω1, ω2, and ω3, respectively, are the weights of the received reference signal power, delay, and throughput.

[0022] Optionally, the calculation method of the received reference signal power is:

[0023]

[0024] Among them, P EPRE is the downlink reference signal power, which is a multimodal fusion feature obtained by processing the power parameters of the system text, system voice and system line image; G BSThe antenna power of the base station is based on the multi-modal fusion features obtained by processing the antenna power parameters in the performance monitoring text, antenna operation voice instructions, and system monitoring images; The antenna gain received by the device is based on the multimodal fusion features obtained by processing the antenna parameters in the antenna technology text, antenna technology voice command, and antenna image; L BS is the feeder loss of the base station, which is based on the multimodal fusion features obtained by processing the feeder parameters in the antenna monitoring text, antenna technical voice instructions, and antenna images; M shadowing is the shadow fading loss based on the threshold of the downlink reference signal power, L obstacle is the signal loss caused by obstacles, L mopdel is the transmit and receive path loss of the propagation model of the pre-built wireless communication process, L OTA M is the loss caused by the transmission distance between the base station and the device. shadowing , L obstacle , L mopdel , L OTA They are respectively multimodal fusion features obtained by processing environmental assessment text, environmental monitoring voice information, shadow fading loss parameters, signal loss parameters, path loss parameters, and propagation loss parameters in signal propagation images.

[0025] Optionally, the delay is calculated as follows:

[0026] D=D e +0.8ΔD1+0.2×D2(18)

[0027] Among them, D e is a fixed delay overhead, D1 is the time required from the sender sending data to the receiver receiving the data for the first time, D2 is the retransmission delay, and D e , D1, and D2 are multimodal fusion features obtained by processing the fixed delay parameters, initial transmission delay parameters, and retransmission delay parameters in the transmission performance detection text, transmission monitoring voice information, and transmission monitoring images, respectively.

[0028] Optionally, the throughput is calculated as follows:

[0029]

[0030] in, From base station Δ to device N i bandwidth, From base station Δ to device N i The signal-to-interference-noise ratio, They are multimodal fusion features obtained by processing traffic analysis files, service quality feedback voice information, bandwidth parameters in system coverage images, and signal-to-interference-noise ratio parameters.

[0031] Optionally, performing multimodal fusion processing on the system parameters to obtain multimodal fusion features includes:

[0032] Performing feature extraction on the text information to obtain text features of system parameters;

[0033] Extracting features from the image information to obtain image features of system parameters;

[0034] Extracting features from the speech information to obtain speech features of system parameters;

[0035] Performing dimension alignment processing on the text features, image features and voice features to obtain text features, image features and / or voice features with consistent dimensions;

[0036] Connect text features, image features, and speech features with consistent dimensions to obtain connected features;

[0037] The connection features are fused using a preset multi-head attention module to obtain fused multimodal fusion features.

[0038] The embodiment of the present application also provides a power communication system scheduling device, including:

[0039] An acquisition module, used for acquiring system parameters from text information, image information and / or voice information;

[0040] A feature fusion module, used for performing multimodal fusion processing on the system parameters to obtain multimodal fusion features;

[0041] The scheduling module is used to construct a reward function based on a preset objective function according to the multimodal fusion characteristics, with the resource status of the base station as the state, the power allocated by the base station to the device and the service quality weight of the device as the action, and determine the optimal power allocation strategy based on a reinforcement learning method; wherein the objective function is constructed based on the service quality requirements of the device and the allocatable resources of the base station.

[0042] From the above, it can be seen that the power communication system scheduling method and device provided in the embodiment of the present application obtains system parameters from text information, image information and voice information, performs multimodal fusion processing on the system parameters, obtains multimodal fusion features, and according to the multimodal fusion features, takes the resource status of the base station as the state, takes the power allocated by the base station to the device and the service quality weight of the device as the action, constructs a reward function based on the preset objective function, and determines the optimal power allocation strategy based on the reinforcement learning method. The present application can effectively integrate multimodal system parameters, improve the accuracy and efficiency of resource scheduling decisions, integrate system resources and service quality requirements, optimize resource scheduling of the power communication system, and improve system performance and service quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0044] Figure 1 A schematic diagram of a method flow of an embodiment of the present application;

[0045] Figure 2 This is a block diagram of the device structure of an embodiment of the present application;

[0046] Figure 3 This is a structural block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present disclosure more clearly understood, the present disclosure is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0048] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present application should be the usual meanings understood by people with ordinary skills in the field to which the present disclosure belongs. The "first", "second" and similar words used in the embodiments of the present application do not represent any order, quantity or importance, but are only used to distinguish different components. "Including" or "comprising" and similar words mean that the elements or objects appearing in front of the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connecting" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to represent relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0049] like Figure 1 As shown, an embodiment of the present application provides a power communication system scheduling method, including:

[0050] S101: Acquire system parameters from text information, image information and / or voice information;

[0051] In this embodiment, the system parameters of the power communication system include signal power, antenna power, antenna gain, loss, delay, bandwidth, etc. Each system parameter can be obtained from file information such as system files in text form, monitoring data files, maintenance record files, relevant technical files, and environmental impact assessment files, from image information such as system monitoring images, system layout images, and satellite monitoring images in image form, and from voice information such as dispatching voice instructions, system operation voice instructions, maintenance operation voice instructions, system feedback voice instructions, and environmental assessment voice instructions of the power dispatching center. By fusing the multimodal features of text information, image information, and voice information for any system parameter, it is possible to effectively integrate information from different sources, improve the semantic understanding ability of data processing, and perform dispatching of the power communication system based on multimodal fusion features, which can improve dispatching accuracy and efficiency.

[0052] S102: Perform multimodal fusion processing on system parameters to obtain multimodal fusion features;

[0053] In this embodiment, the system parameters obtained from data sources such as text information, image information and voice information can be multimodal data such as text data, image data and voice data. The multimodal data is processed using a multimodal fusion method to fully explore and utilize the hidden information in the data and improve the accuracy and efficiency of information processing.

[0054] In some embodiments, multimodal fusion processing is performed on system parameters to obtain multimodal fusion features, including:

[0055] Extract features from text information to obtain text features of system parameters;

[0056] Extract features from image information to obtain image features of system parameters;

[0057] Extract features from speech information to obtain speech features of system parameters;

[0058] Perform dimension alignment processing on text features, image features and speech features to obtain text features, image features and speech features with consistent dimensions;

[0059] Connect text features, image features, and speech features with consistent dimensions to obtain connected features;

[0060] The preset multi-head attention module is used to perform feature fusion processing on the connection features to obtain the fused multimodal fusion features.

[0061] In this embodiment, feature extraction is performed on text information, image information and voice information including system parameters to obtain text features, image features and voice features of the system parameters. After dimension alignment of the text features, image features and voice features, connection features are obtained by feature connection. The connection features are input into a multi-head attention module, and the multi-head attention module is used to perform feature fusion on multimodal connection features to obtain multimodal fusion features.

[0062] In some implementations, feature extraction is performed on the text information, image information, and voice information including the system parameters to obtain text features, image features, and voice features of the system parameters, specifically including:

[0063] For text information, use word embedding to convert it into word vector Among them, n T is the number of words in the text, and the dimension of each word vector is d model For image information Use the convolution network to perform convolution operations to obtain image vectors Among them, (H, W) represents the image resolution, C represents the number of channels, and the convolution kernel size is S×S, so n I =HW / S 2 image blocks, and the feature map obtained by the convolution operation of each image block has an expanded dimension of d model For speech information, the audio waveform is preprocessed using a filter bank, and the frequency is f s The Hamming window divides the original audio wave into n V =(t / t s ),t s =2π / f s intervals, and use the convolution kernel to split the audio waveform spectrogram with the same time and frequency in each interval into n overlapping spectrograms. V After passing through a convolution layer with a convolution kernel size of S×S, the speech vector is obtained.

[0064] In some methods, the convolution kernel size S and dimension d model Determined during model training. The initial parameter d model It can be set to 1024, the number of convolution kernels is 1024, the size of the convolution kernel is 16×16, and f s The text information, image information, and speech information are processed by semantic embedding to obtain n T 、n I 、nV d model Embedding vector of

[0065] After the text information, image information and voice information are converted into word vectors, image vectors and voice vectors respectively, the word vectors, image vectors and voice vectors are positionally embedded to obtain n T 、n I 、n V d model Dimensional word vectors, image vectors, and speech vectors with positional markers Among them, sine function and cosine function can be used for position embedding, which can be expressed as:

[0066]

[0067] Where pos represents the position and i represents the dimension. The position embedding identifier is combined with the embedding vector The corresponding pos is superimposed to obtain the embedded vector with position mark

[0068] The word vector with position mark, image vector and speech vector are weighted with the pre-generated weight matrix and input into the two-layer feedforward neural network model, and the feedforward neural network model outputs the corresponding text semantic feature s T , image semantic features I and phonetic semantic features V . Use sine function and cosine function to analyze the text semantic features s T , image semantic features I and phonetic semantic features V Perform position embedding to obtain text semantic features with position identifiers Image semantic features and phonetic and semantic features

[0069] In some embodiments, dimension alignment is performed on text features, image features, and / or voice features to obtain text features, image features, and / or voice features with consistent dimensions, including:

[0070] The text semantic features, image semantic features and speech semantic features with position marks are dimensionally aligned to obtain text semantic features, image semantic features and speech semantic features with consistent dimensions.

[0071] In this embodiment, after obtaining text semantic features, image semantic features and speech semantic features with position identifiers, it is necessary to format and normalize the feature data of different modalities, and align the dimensions of the feature data of different modalities to ensure the consistency of the data in subsequent processing.

[0072] Among them, for text semantic features, the co-occurrence matrix GloVe (Global Vectors for Word Representation) is used to map the text semantic features with position marks. Perform feature conversion to obtain the converted text semantic features Expressed as The converted text semantic features are input into the bidirectional long short-term memory network model, and the model is used to extract the key text features in the text semantic features, which is expressed as θ T is the model parameter.

[0073] For image semantic features with location labels The converted image semantic features are obtained through standardization Expressed as Among them, μ and σ are the mean and standard deviation of pixel values ​​respectively. Input the convolutional neural network model and use the model to extract key image features from the image semantic features, expressed as θ I is the model parameter.

[0074] For speech semantic features with positional markers Mel-Frequency Cepstrum (MFCC) conversion and voice activity detection (VAD) are used to filter the non-speech part to ensure data purity and obtain the converted speech semantic features. Expressed as The converted speech semantic features are input into the recurrent neural network model, and the model is used to extract the key speech features in the speech semantic features, which is expressed as θ V is the model parameter.

[0075] Key text features after feature conversion and extraction Key image features Key voice features The dimensionally aligned features are connected to obtain the connected features. The connection features are input into the multi-head attention module, and the multi-head attention module is used to perform feature fusion on the multimodal features.

[0076] In some embodiments, each self-attention head of the multi-head attention module converts the input connection feature F into the corresponding query vector Q, key vector K and value vector V, calculates the dot product of the query vector and the key vector to obtain an attention score, and normalizes it through the softmax function to obtain an attention weight, which is expressed as:

[0077]

[0078] Among them, W Q , W K , W v They represent the weights of the query vector Q, key vector K, and value vector V respectively, and H represents the vector dimension.

[0079] Connect the attention weights of each self-attention head and then add them to the cross attention weight W F Perform point multiplication to get the output of the multi-head attention layer, expressed as:

[0080]

[0081] Where n is the number of sub-attention heads, represents the output of i attention operations performed on the self-attention head SA.

[0082] The output of the multi-head attention layer and the output of the i-1 attention operation on the self-attention head SA are combined Add and normalize, expressed as:

[0083]

[0084] Among them, LN represents the normalization operation, i represents the number of iterations, and θ i Represents the model parameters of the multi-head attention layer. The attention iteration operation is continuous, and the output of the previous attention needs to be used as the input of the next attention operation.

[0085] Normalized features Input the feedforward neural network model, and output the features and characteristics of the feedforward neural network model After normalization, it is expressed as:

[0086]

[0087] Among them, FNN represents the feedforward neural network model, θ j Represents the model parameters of the feedforward neural network model.

[0088] The multimodal fusion feature after the final multimodal fusion is expressed as S F , the calculation formula can be expressed as:

[0089] ...

[0091]

[0092] Among them, l represents the set number of attention layers, which can be adjusted according to the quality of the semantic fusion results.

[0093] In some embodiments, the weights of the self-attention heads (i.e., W Q , W K , W v The specific method includes: defining the semantic feature extraction task as MDP: M=(S,A,P,R,P0,γ,J), where S represents the state space, A represents the action space, P:S×A×S→[0,∞) represents the state transition, R::S×A→[-R max ,R max ] represents bounded reward, P0:S→[0,∞) represents the initial state distribution, γ∈[0,1] represents the discount factor, and J represents the objective function.

[0094] Training a meta-knowledge through few sample data includes: setting the state space S: taking the input multimodal fusion features as the state representation S, and the initial state as the embedding vector Set up the action space A: define the weight matrix as action a, where the action is the attention weight A={a} assigned by the agent to each embedding vector in the state, Initialize action a0 randomly. Set reward function R: The state s' obtained by executing action a in state s of iteration t in the next iteration cycle t+1 is s' = (s t+1 |s t ,a t ) corresponding to the reward; considering the impact of the output on the task performance after the input is weighted and other operations, the cross entropy loss function is used to construct the reward function. The reward function of a single time step is expressed as:

[0095]

[0096] Among them, C represents the number of categories, y j is the probability that the model predicts category j, is the true category label. In the application scenario of power communication system, the semantic meaning of data represents a category. For example, the semantic categories of multimodal fusion features include failure, normal operation, etc.

[0097] Create a global agent and multiple sub-agents. The multiple sub-agents interact with the environment through asynchronous multi-threaded operations to explore their own strategies, calculate the advantage function of the model's predicted probability, and feed back their respective strategies to the global agent after the exploration. The global agent selects the best strategy from the strategies fed back by each sub-agent, trains according to the best strategy, and iterates the above process until the objective function is reached.

[0098] In some approaches, the advantage function is:

[0099]

[0100] Among them, θ v are network parameters, is the target value function considering k-step rewards in the Deep Q-Network (DQN), γ i is the reward discount factor for the i-th iteration, r t+i is the reward for the i+1th iteration, V(s t θ v ) represents the deep Q network in state s t The value calculated below, θ v represents the parameters of the Q network.

[0101] In some embodiments, the action model of the agent adopts the actor network model, and the strategy model of the agent adopts the critic network model. The actor network model and the critic network model share model parameters except the parameters of the output layer. During the training process, the parameter gradient of the actor network is calculated according to the advantage function, which is expressed as:

[0102]

[0103] The calculation method of the network parameters of the critic network model is:

[0104]

[0105] After each sub-agent updates the parameter gradients and network parameters of the actor network model and the critic network model according to the above formula, the global agent adds the parameter gradients of each sub-agent and updates the model parameters θ of the global agent v and θ.

[0106] The state update method of the agent is:

[0107] s t ′=MLP(MSA(s t ;θ2)+s t ) (12)

[0108] Among them, MSA(s t) represents the state s at time t t Input to the Q network and adopt the attention matrix obtained by the strategy. MLP includes two fully connected layers. The input dimension of the first fully connected layer is n×d model , the output of the first layer is hidden_size, s t ′ is the state s t The state after the next iteration.

[0109] The output of the first fully connected layer is input into the second fully connected layer. The output of the second fully connected layer is nonlinearly mapped and then normalized to obtain the output of the encoder layer, that is, the state s at time t+1. t+1 , with dimensions of n×d model , for text vector, image vector and semantic vector, n is T 、n I 、n V .

[0110] The updated status is represented as:

[0111] s t+1 =LN(ReLu(MLP(s t ';θ3));θ4) (13)

[0112] Among them, LN represents the layer normalization operation, which is used to stabilize the output within a reasonable range, expressed as:

[0113]

[0114] Among them, x represents the input feature, μ and σ 2 They represent the mean and variance of feature x, respectively, and are used to map the input to a standard distribution with a mean of 0 and a variance of 1. α and b are learnable parameters.

[0115] In some embodiments, the training process of the global agent includes: initializing the parameters of the actor network model and the critic network model, using the actor network model sampling trajectory and the state transfer function shown in formula (13), using the loss function to update the model parameters in formula (13), and obtaining the support set D support ={s1,a1,+1,s2,…,s k ,a k ,r k ,s k+1}, and the value function of the last state in the support set, expressed as:

[0116]

[0117] Calculate the advantage function of k time steps according to formula (9), calculate the parameter gradient of the actor network according to formula (10), accumulate the parameter gradients of each sub-agent to get dθ', calculate the mean square error of the advantage function, and calculate the network parameter θ of the critic network model v ' parameter gradient, accumulate the parameter gradients of each sub-agent to get dθ v '.

[0118] According to the updated parameter gradient, update the base learner parameters: θ′←θ′-η θ′ dθ′, Among them, η θ′ is the learning rate of the base learner, θ v ′ is the sub-learner parameter, is the learning rate of the sub-learner, θ1′ is the parameter for training a specific task after training the base learner, η1 is the learning rate of the specific task learner, represents the gradient of the loss function.

[0119] Using the updated actor network model and state transition network, the sampled trajectories are used to obtain the query set D query ={s1,a1,r1,s2,...,s N ,a N ,r N ,s N+1}. Compute the advantage function of the global agent in the query set and update the meta-learner parameters:

[0120] Use the trained global agent to output the optimal self-attention weight W Q , W K , W V and W F .

[0121] In this embodiment, with few-sample learning and multimodal feature fusion as the core, multimodal semantic fusion based on the few-sample cross-attention mechanism is realized. By capturing information from different data sources such as text, images and voice, and using the few-sample learning method to extract key features under different modalities, effective learning can be achieved even when the amount of data is small, which greatly expands the model's ability to process data in different fields; the cross-attention mechanism is used to strengthen the semantic association between different modal data, which can identify and highlight the modal features that have the greatest impact on the final task performance, and improve the accuracy and robustness of semantic fusion. The cross-attention layer can be superimposed in multiple layers, so as to deeply explore and optimize the mutual information between different modalities, improve the efficiency and accuracy of the model in processing multimodal data, and enhance the adaptability and interpretability of the model in complex semantic tasks.

[0122] S103: According to the multimodal fusion characteristics, taking the resource status of the base station as the status, taking the power allocated by the base station to the device and the service quality weight of the device as the action, building a reward function based on a preset objective function, and determining the optimal power allocation strategy based on a reinforcement learning method; wherein the objective function is built based on the service quality requirements of the device and the allocatable resources of the base station.

[0123] In this embodiment, after multimodal fusion processing is performed on the system parameters of text, image and voice modes to obtain multimodal fusion features, a model is constructed with the resource status of the base station as the state, the power allocated by the base station to the device and the service quality weight of the device as the action to achieve the preset objective function. The power allocation strategy of the optimal device is determined based on the reinforcement learning method to achieve the purpose of optimizing resource scheduling.

[0124] In some embodiments, in a 5G heterogeneous communication scenario, the power communication system includes a macro base station and a micro base station. The macro base station U obtains multi-modal system parameters, performs multi-modal fusion processing to obtain multi-modal fusion features, and distributes the multi-modal fusion features of the system parameters to K micro base stations M = {M1, ..., M K}, where the coverage radius of the macro base station is R, the coverage radius of the micro base station is r, the base station set is Δ = M ∪ U, and there are N devices in the network. The devices covered by the macro base station but not by the micro base station are represented as N M , the number of devices covered by the micro base station is represented by N K , N=N M ∪N K .

[0125] In order to meet the service quality requirements of the equipment, the service quality satisfaction of the equipment is constructed, which is expressed as:

[0126]

[0127] Among them, P EPRE is the received reference signal power, D is the delay, N devices connected to the base station Δ i The throughput of the received reference signal is ω1, ω2, and ω3, respectively, are the weights of the received reference signal power, delay, and throughput.

[0128] In some embodiments, the received reference signal power is calculated as follows:

[0129]

[0130] Among them, P EPRErepresents the downlink reference signal power, which is a multimodal fusion feature obtained by processing power parameters in system text (such as operation log files, fault maintenance log files, etc.), system voice (such as voice instructions from the dispatch center), and system line images (such as substation and power line monitoring images); G BS Indicates the antenna power of the base station, which is based on the multi-modal fusion features obtained by processing the antenna power parameters in the performance monitoring text (e.g., base station performance monitoring text), the antenna operation voice command (e.g., voice commands related to base station operation and antenna adjustment), and the system monitoring image (e.g., base station antenna monitoring image); Indicates the antenna gain received by the device, which is a multi-modal fusion feature obtained by processing antenna parameters from antenna technical text, antenna technical voice instructions (such as voice feedback instructions from field maintenance personnel), and antenna images (such as antenna physical layout and location images); L BS Indicates the feeder loss of the base station, which is based on the multi-modal fusion features obtained by processing the feeder parameters in the antenna monitoring text (e.g., base station feeder performance monitoring text), antenna technical voice instructions, and antenna images; M shadowing represents the shadow fading loss based on the threshold of the downlink reference signal power; L obstacle Indicates the signal loss caused by obstacles; L mopdel represents the transmit and receive path loss of the propagation model of the pre-built wireless communication process; L OTA Indicates the loss caused by the propagation distance between the base station and the device, M shadowing , L obstacle , L mopdel , L OTA They are multimodal fusion features obtained by processing shadow fading loss parameters, signal loss parameters, path loss parameters, and propagation damage parameters from environmental assessment texts, environmental monitoring voice information, and signal propagation images (such as satellite monitoring images and signal images taken by drones).

[0131] In some methods, the delay is calculated as:

[0132] D=D e +0.8×D1+0.2×D2(18)

[0133] Among them, D e represents the fixed delay overhead, D1 represents the initial transmission delay, that is, the time required from the sender sending data to the receiver receiving the data for the first time, D2 represents the retransmission delay, D e, D1, and D2 are multimodal fusion features obtained by processing the fixed delay parameters, initial transmission delay parameters, and retransmission delay parameters in the transmission performance detection text (such as performance monitoring log files, fault analysis files), the transmission monitoring voice information (such as voice instructions from the dispatch center), and the transmission monitoring image, respectively.

[0134] In some ways, the throughput is calculated as:

[0135]

[0136] in, Indicates base station Δ to device N i bandwidth, Indicates base station Δ to device N i The signal-to-interference-noise ratio, They are multimodal fusion features obtained by processing traffic analysis files, service quality feedback voice information, bandwidth parameters in system coverage images, and signal-to-interference-noise ratio parameters.

[0137] In some embodiments, the base station schedules resources based on maximum throughput. Each device can only connect to one base station. The base station divides its resources into multiple resource blocks, and can allocate resource blocks to devices. Specifically, the power in the resource blocks is allocated to the devices. The overall network performance is optimized by finding the optimal resource block allocation method.

[0138] In some embodiments, the throughput of micro base station k can be expressed as:

[0139]

[0140] Among them, T i,k represents the throughput of device i connected to micro base station k, N k Represents the number of devices connected to micro base station k.

[0141] The throughput of all devices covered by the macro base station U is:

[0142]

[0143] Under the condition of meeting certain constraints, the system performance is optimized by adjusting the service quality weight of the equipment and the power allocation of the micro base station. The objective function of resource scheduling is constructed as follows:

[0144]

[0145] ikB k ≤RB max

[0146]

[0147] Among them, RBk Indicates the number of resource blocks that can be allocated by base station k (a micro base station in this application scenario), RB max Indicates the total number of resource blocks; Represents the device N connected to base station k i The service quality requirement, α i Indicates device N i The service quality weight of Represents the device N connected to base station k i The average satisfaction of μ represents the satisfaction threshold of the device, and T(P, W) is the total throughput T of the base station controlled by the allocated power P and the service quality weight W of the device.

[0148] The constraints of the above objective function are as follows: the resource blocks that can be allocated by the base station are limited by the total number of resource blocks; the resources provided to the device cannot exceed the satisfaction threshold of the device; the resources provided to the device must be greater than zero.

[0149] In some embodiments, based on the meta-reinforcement learning method, the resource state of the base station is taken as the state, the power P allocated by the base station to the device and the service quality weight W of the device are taken as the action a k ,A reward function is constructed based on the objective function, and each base station is regarded as an intelligent agent. The power communication system is used as the environment, and the optimal power allocation strategy of the base station is determined through the reinforcement learning process.

[0150] Specifically, the resource state of base station k, that is, the state can be expressed as:

[0151]

[0152] Among them, when the state s k =1, it means that the capacity of base station k is overloaded. k =2, it means that there are available resource blocks in base station k. k =3, it means that the capacity of base station k is not overloaded and the resources are not excessive (excessive resources means that the number of allocated resource blocks is too large); K is the set of base stations, a k is the action of base station k, Ka k Indicates that all base stations in the base station set K take their own actions a k The throughput obtained.

[0153] The reward function is constructed based on the objective function:

[0154]

[0155] Among them, in state s k Take action a k The reward value is Whenk =2, the reward value is A1 and A2 are both set constants.

[0156] In some embodiments, the training process of the meta-reinforcement learning algorithm includes: initializing model parameters, setting a maximum number of iterations, performing task-specific adaptation for the selected task in each iteration, and performing gradient updates based on the initial model parameters θ, which can be expressed as:

[0157]

[0158] Where α is the intra-task learning rate used to adjust the fast adaptation step on a specific task, f(θ) is the model parameterized by the model parameters θ, It is task T i The loss function is calculated using the negative value of the reward function. represents the derivative of parameter θ.

[0159] Get the task-specific tuned parameters θ i ', we implement task-specific adaptation. We use the adjusted parameters to perform the same task T i Or calculate the loss on the relevant validation task and find the gradient about the original model parameter θ:

[0160]

[0161] Wherein, n is the number of iterations. Through the above training process, a reinforcement learning model suitable for resource scheduling of the power communication system is obtained, and the resource scheduling of the power communication system is realized by using the trained reinforcement learning model.

[0162] The power communication system resource scheduling method provided in the embodiment of the present application obtains multi-modal system parameters such as text, image and voice, and uses a cross-attention mechanism to perform multi-modal feature fusion processing on the multi-modal system parameters, which can improve the efficiency and accuracy of multi-modal data fusion and the accuracy and efficiency of scheduling decisions. Taking into account the resource status of the base station and the service quality requirements of the equipment, the optimal system resource allocation strategy is determined based on the reinforcement learning method, which improves the overall performance of the power communication system, improves the service quality level, and improves user satisfaction.

[0163] It should be noted that the method of the embodiment of the present application can be performed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario and completed by multiple devices cooperating with each other. In the case of such a distributed scenario, one of the multiple devices can only perform one or more steps in the method of the embodiment of the present application, and the multiple devices will interact with each other to complete the described method.

[0164] It should be noted that the above is a description of a specific embodiment of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0165] like Figure 2 As shown, the embodiment of the present application also provides a power communication system scheduling device, including:

[0166] An acquisition module, used for acquiring system parameters from text information, image information and / or voice information;

[0167] A feature fusion module, used for performing multimodal fusion processing on the system parameters to obtain multimodal fusion features;

[0168] The scheduling module is used to construct a reward function based on a preset objective function according to the multimodal fusion characteristics, with the resource status of the base station as the state, the power allocated by the base station to the device and the service quality weight of the device as the action, and determine the optimal power allocation strategy based on the reinforcement learning method; wherein the objective function is constructed based on the service quality requirements of the device and the allocatable resources of the base station.

[0169] For the convenience of description, the above devices are described in terms of functions divided into various modules. Of course, when implementing the embodiments of the present application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0170] The apparatus of the above-mentioned embodiment is used to implement the corresponding method in the above-mentioned embodiment, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0171] Figure 3 A more specific schematic diagram of the hardware structure of an electronic device provided in this embodiment is shown, and the device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are connected to each other through the bus 1050 in the device.

[0172] The processor 1010 can be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0173] The memory 1020 may be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 may store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program codes are stored in the memory 1020 and are called and executed by the processor 1010.

[0174] The input / output interface 1030 is used to connect the input / output module to realize information input and output. The input / output module can be configured in the device as a component (not shown in the figure), or it can be externally connected to the device to provide corresponding functions. The input device may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device may include a display, a speaker, a vibrator, an indicator light, etc.

[0175] The communication interface 1040 is used to connect a communication module (not shown) to realize communication interaction between the device and other devices. The communication module can realize communication through a wired mode (such as USB, network cable, etc.) or a wireless mode (such as mobile network, WIFI, Bluetooth, etc.).

[0176] The bus 1050 includes a path that transmits information between the various components of the device (eg, the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040).

[0177] It should be noted that, although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040 and the bus 1050, in the specific implementation process, the device may also include other components necessary for normal operation. In addition, it can be understood by those skilled in the art that the above device may also only include the components necessary for implementing the embodiments of the present specification, and does not necessarily include all the components shown in the figure.

[0178] The electronic device of the above embodiment is used to implement the corresponding method in the above embodiment, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0179] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0180] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present disclosure (including the claims) is limited to these examples. Based on the concept of the present disclosure, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0181] In addition, to simplify the description and discussion, and in order not to make the embodiments of the present application difficult to understand, the known power supply / ground connection with the integrated circuit (IC) chip and other components may or may not be shown in the provided drawings. In addition, the device can be shown in the form of a block diagram to avoid making the embodiments of the present application difficult to understand, and this also takes into account the fact that the details of the implementation of these block diagram devices are highly dependent on the platform to be implemented in the embodiments of the present application (that is, these details should be fully within the scope of understanding of those skilled in the art). In the case of elaborating specific details (e.g., circuits) to describe exemplary embodiments of the present disclosure, it is obvious to those skilled in the art that the embodiments of the present application can be implemented without these specific details or when these specific details are changed. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0182] Although the present disclosure has been described in conjunction with specific embodiments of the present disclosure, many replacements, modifications and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may use the embodiments discussed.

[0183] The embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application should be included in the scope of protection of the present disclosure.

Claims

1. A power communication system scheduling method, characterized in that: include: Obtain system parameters from text information, image information and voice information; Performing multimodal fusion processing on the system parameters to obtain multimodal fusion features; According to the multimodal fusion characteristics, taking the resource status of the base station as the state, the power allocated by the base station to the device and the service quality weight of the device as the action, a reward function is constructed based on a preset objective function, and the optimal power allocation strategy is determined based on a reinforcement learning method; wherein, the objective function is constructed based on the service quality requirements of the device and the allocatable resources of the base station.

2. The method according to claim 1, characterized in that The resource status of the base station is expressed as: Among them, s k is the state of base station k, T k is the throughput of base station k, N k is the number of devices connected to base station k, is the device N connected to the base station k i The service quality requirements are determined according to the service quality weight and service quality satisfaction of the equipment. is the device N connected to the base station k i The average satisfaction, θ μ is the satisfaction threshold of the equipment, Ka k Indicates that all base stations in the base station set K take their own actions a k The throughput obtained.

3. The method according to claim 2, characterized in that The objective function is expressed as: s.t.RB k ≤RB max Among them, RB k is the number of resource blocks that can be allocated to base station k, RB max is the total number of resource blocks; P is the power allocated by base station k to the device, W is the service quality weight of the device, and T(P, W) is the total throughput T of the base station under the control of parameters P and W.

4. The method according to claim 3, characterized in that The reward function is: in, For state s k Take action a k The reward value obtained, A1 and A2 are set constants.

5. The method according to claim 3, characterized in that: The service quality satisfaction of the equipment is expressed as: Among them, P RSRP is the received reference signal power, D is the delay, N devices connected to the base station Δ i The throughput of the received reference signal is ω1, ω2, and ω3, respectively, are the weights of the received reference signal power, delay, and throughput.

6. The method according to claim 5, characterized in that The calculation method of the received reference signal power is: Among them, P EPRE is the downlink reference signal power, which is a multimodal fusion feature obtained by processing the power parameters of the system text, system voice and system line image; G BS The antenna power of the base station is based on the multi-modal fusion features obtained by processing the antenna power parameters in the performance monitoring text, antenna operation voice instructions, and system monitoring images; The antenna gain received by the device is based on the multimodal fusion features obtained by processing the antenna parameters in the antenna technology text, antenna technology voice command, and antenna image; L BS is the feeder loss of the base station, which is based on the multimodal fusion features obtained by processing the feeder parameters in the antenna monitoring text, antenna technical voice instructions, and antenna images; M shadowing is the shadow fading loss based on the threshold of the downlink reference signal power, L obstacle is the signal loss caused by obstacles, L mopdel is the transmit and receive path loss of the propagation model of the pre-built wireless communication process, L OTA M is the loss caused by the transmission distance between the base station and the device. shadowing , L obstacle , L mopdel , L OTA They are respectively multimodal fusion features obtained by processing environmental assessment text, environmental monitoring voice information, shadow fading loss parameters, signal loss parameters, path loss parameters, and propagation loss parameters in signal propagation images.

7. The method according to claim 5, characterized in that The calculation method of the time delay is: D=D e +0.8×D1+0.2×D2 (18) Among them, D e is a fixed delay overhead, D1 is the time required from the sender sending data to the receiver receiving the data for the first time, D2 is the retransmission delay, and D e , D1, and D2 are multimodal fusion features obtained by processing the fixed delay parameters, initial transmission delay parameters, and retransmission delay parameters in the transmission performance detection text, transmission monitoring voice information, and transmission monitoring images, respectively.

8. The method according to claim 5, characterized in that The throughput is calculated as follows: in, From base station Δ to device N i bandwidth, From base station Δ to device N i The signal-to-interference-noise ratio, They are multimodal fusion features obtained by processing traffic analysis files, service quality feedback voice information, bandwidth parameters in system coverage images, and signal-to-interference-noise ratio parameters.

9. The method according to claim 1, characterized in that: The system parameters are subjected to multimodal fusion processing to obtain multimodal fusion features, including: Performing feature extraction on the text information to obtain text features of system parameters; Extracting features from the image information to obtain image features of system parameters; Extracting features from the speech information to obtain speech features of system parameters; Performing dimension alignment processing on the text features, image features and voice features to obtain text features, image features and / or voice features with consistent dimensions; Connect text features, image features, and speech features with consistent dimensions to obtain connected features; The connection features are fused using a preset multi-head attention module to obtain fused multimodal fusion features.

10. A dispatching device for a power communication system, characterized in that: include: An acquisition module, used for acquiring system parameters from text information, image information and / or voice information; A feature fusion module, used for performing multimodal fusion processing on the system parameters to obtain multimodal fusion features; The scheduling module is used to construct a reward function based on a preset objective function according to the multimodal fusion characteristics, with the resource status of the base station as the state, the power allocated by the base station to the device and the service quality weight of the device as the action, and determine the optimal power allocation strategy based on a reinforcement learning method; wherein the objective function is constructed based on the service quality requirements of the device and the allocatable resources of the base station.

Images