Method, device, equipment, medium and product for completion and prediction of wireless data

Abstract

This invention provides a method, apparatus, device, medium, and product for wireless data completion and prediction, relating to the field of data processing technology. The method includes: determining the phase space vector corresponding to standard variable data; inputting the phase space vector and its corresponding allowed interactive neighborhood into a knowledge graph relation network to obtain micro-update values; inputting the global context sequence into a time-series network to obtain node-level macro-update values; and determining the data completion value and future time-step prediction value corresponding to the input wireless data based on the decoder, all micro-update values, and all node-level macro-update values. This invention is compatible with various wireless data types and constructs a unified physical constraint and domain knowledge framework through a knowledge graph, thereby improving robustness in complex scenarios such as systemic blind spots and cross-scenario missing data, and achieving accurate completion and prediction of wireless data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, device, medium, and product for completing and predicting wireless data. Background Technology

[0002] As wireless communication technology evolves towards 5G and 6G, the dimensions and application scenarios of wireless data continue to expand, encompassing key performance indicators at the network layer, transmission status data at the link layer, sensor data from IoT terminals (such as temperature, humidity, and motion status), and dynamic link quality data from vehicle-to-everything (V2X) networks. This data forms the core foundation for intelligent network scheduling, collaborative device control, and early fault warning, demanding extremely high levels of data integrity and timeliness.

[0003] However, wireless data acquisition faces a variety of complex missing scenarios, including: (1) Systemic blind spots: target data and strongly related proxy data are missing simultaneously in a continuous time block (e.g., the multi-layer protocol stack counter of 6G network fails simultaneously due to link congestion); (2) Intermittent missing: data acquisition is interrupted due to low power sleep or signal blockage of IoT devices; (3) Cross-scenario missing: in the scenarios of vehicle networking and mobile edge computing, data transmission is interrupted or distribution characteristics change suddenly due to high-speed movement of devices or scene switching (indoor / outdoor, dense urban / suburban areas); (4) Heterogeneous missing: when multiple types of wireless data (numerical, discrete, binary) are collected in a mixed manner, some types of data are missing due to differences in acquisition protocols or hardware limitations.

[0004] Existing wireless data completion and prediction technologies are mainly divided into three categories: (1) traditional statistical methods (interpolation / filtering, low-rank / tensor decomposition); (2) deep learning methods (spatiotemporal graph neural networks, sequence prediction models); and (3) domain adaptation methods.

[0005] The aforementioned methods in the prior art have the following common defects: they are difficult to simultaneously accommodate multiple types of wireless data such as numerical, discrete, and time-series data; they lack robustness across scenarios (static / dynamic, indoor / outdoor); and the prediction results may violate wireless communication principles (such as negative power or throughput exceeding the theoretical upper limit) or data association rules (such as the physical mapping relationship between sensing data and link status), that is, the accuracy of the prediction results is low. Summary of the Invention

[0006] This invention provides a method, apparatus, device, medium, and product for wireless data completion and prediction, addressing the shortcomings of existing technologies. The technical solution of this invention is compatible with various types of wireless data. It constructs a unified physical constraint and domain knowledge framework through a knowledge graph. The combination of knowledge graph relationship networks and temporal networks improves robustness in complex scenarios such as systemic blind spots and cross-scenario missing data. Furthermore, by fusing micro-level update quantities and node-level macro-level update quantities, it achieves accurate completion and prediction of wireless data.

[0007] This invention provides a method for completing and predicting wireless data, comprising the following steps.

[0008] The input wireless data is standardized to obtain the standardized wireless data corresponding to the input wireless data. For each standard variable in the standardized wireless data, determine the phase space vector corresponding to the standard variable data; Construct a knowledge graph corresponding to the standardized wireless data and a structure encoding corresponding to the knowledge graph, and determine the allowed interaction neighborhoods corresponding to each of the standardized variable data based on the knowledge graph; For each phase space vector, the micro-update amount corresponding to the phase space vector is determined based on the knowledge graph relation network, the phase space vector, and the allowed interaction neighborhood corresponding to the phase space vector; the knowledge graph relation network is constructed based on the knowledge graph and the structure encoding; A global context sequence is determined based on all the phase space vectors, and the node-level macroscopic update amount corresponding to each of the standard variable data is determined based on the global context sequence and the time series network. Based on the decoder, all the micro-updates, and all the node-level macro-updates, the data completion value and future time step prediction value corresponding to the input wireless data are determined; wherein, the knowledge graph relationship network, the time series network, and the decoder are jointly trained based on the sample wireless data and the missing indication information corresponding to the sample wireless data.

[0009] According to the present invention, a method for completing and predicting wireless data is provided, wherein the knowledge graph relationship network is used for: For each phase space vector, the phase space vector is mapped to the positive latent space to obtain a non-negative latent state vector; Normalize the non-negative potential state vector on the positive half-unit hypersphere to obtain the direction components located on the positive half-unit hypersphere; The query vector, key vector, and value vector are determined based on the structural encoding, the allowed interactive neighborhood corresponding to the phase space vector, and the directional components. The micro-update amount corresponding to the phase space vector is determined based on the query vector, the key vector, and the value vector.

[0010] According to a wireless data completion and prediction method provided by the present invention, the step of determining a query vector, a key vector, and a value vector based on the structure encoding, the allowed interaction neighborhood corresponding to the phase space vector, and the direction component includes: Within the allowed interaction neighborhood corresponding to the phase space vector, an initial query vector and an initial key vector are generated based on the directional components; Based on the structure encoding, the initial query vector, and the initial key vector, the query vector and the key vector are determined; Local interaction weights are determined based on the angular similarity between the directional component and the directional components of the neighbors in the allowed interaction neighborhood. The value vector is determined based on the local interaction weights and the initial value vectors corresponding to the neighbor directional components in the allowed interaction neighborhood.

[0011] According to a wireless data completion and prediction method provided by the present invention, the time-series network is used for: The global context sequence is mapped to the positive latent space to obtain the global positive representation corresponding to the global context sequence; The macroscopic inertia update amount is determined based on the global positive representation and the temporal network; The macroscopic inertia update is broadcast or mapped to the nodes corresponding to each of the standard variable data to obtain the node-level macroscopic update for each of the standard variable data.

[0012] According to a wireless data completion and prediction method provided by the present invention, the step of determining the data completion value and future time step prediction value corresponding to the input wireless data based on the decoder, all the micro-updates, and all the node-level macro-updates includes: For each of the standard variable data, based on the gating network and the corresponding micro-update and node-level macro-update, the positive latent state corresponding to the standard variable data is determined; the gating network is jointly trained based on the sample wireless data and the missing information corresponding to the sample wireless data. Based on all the positive latent states and the decoder, the data completion value and future time step prediction value corresponding to the input wireless data are determined.

[0013] According to a wireless data completion and prediction method provided by the present invention, the step of determining the positive latent state corresponding to the standard variable data based on the gated network and the micro-update quantity and node-level macro-update quantity corresponding to the standard variable data includes: The fusion threshold is determined based on the gating network, the micro-update amount corresponding to the standard variable data, and the node-level macro-update amount. Based on the fusion threshold, the micro-update amount and the node-level macro-update amount are fused to obtain the total tangent vector update amount; The positive latent state is obtained by back-mapping the total tangent vector update with the current latent state.

[0014] According to a method for completing and predicting wireless data provided by the present invention, the step of constructing a knowledge graph corresponding to the standardized wireless data and a structure encoding corresponding to the knowledge graph includes: The node set is determined based on the standard variable data in the standardized wireless data. The edge set is determined based on the dependencies between the standard variable data; the dependencies include one or more of the following: process / hierarchy dependency, physical / mechanism dependency, control and feedback dependency, and scenario and entity association; The knowledge graph is constructed based on the set of nodes and the set of edges; Determine the structural encoding corresponding to the knowledge graph; the structural encoding includes one or more of topological location encoding, type identifier encoding, and scene / context encoding.

[0015] According to a method for completing and predicting wireless data provided by the present invention, the step of standardizing the input wireless data to obtain standardized wireless data corresponding to the input wireless data includes: For input wireless data of continuous numerical type, one of the following processing methods is performed: standardization, normalization, and robust scaling, to obtain the corresponding first sub-standardized data. For discrete / categorical data in the input wireless data, embedding or one-hot coding is performed to obtain the corresponding second sub-normalized data; For input wireless data of binary / symbolic type, one of the following methods is used: embedding, one-hot coding, and threshold mapping to obtain the corresponding third sub-normalized data; The standardized wireless data is determined based on the first sub-standardized data, the second sub-standardized data, and the third sub-standardized data.

[0016] The present invention also provides a wireless data completion and prediction device, comprising the following modules: The standardization module is used to standardize the input wireless data to obtain the standardized wireless data corresponding to the input wireless data. The vector module is used to determine the phase space vector corresponding to each standard variable data in the standardized wireless data. The knowledge graph module is used to construct a knowledge graph corresponding to the standardized wireless data and a structure encoding corresponding to the knowledge graph, and to determine the allowed interaction neighborhoods corresponding to each of the standard variable data based on the knowledge graph. The micro-module is used to determine the micro-update amount corresponding to each phase space vector based on the knowledge graph relation network, the phase space vector, and the allowed interaction neighborhood corresponding to the phase space vector; the knowledge graph relation network is constructed based on the knowledge graph and the structure encoding; The macro module is used to determine the global context sequence based on all the phase space vectors, and to determine the node-level macro update amount corresponding to each of the standard variable data based on the global context sequence and the time series network. The completion prediction module is used to determine the data completion value and future time step prediction value corresponding to the input wireless data based on the decoder, all the micro-updates, and all the node-level macro-updates; wherein, the knowledge graph relationship network, the time series network, and the decoder are jointly trained based on the sample wireless data and the missing indication information corresponding to the sample wireless data.

[0017] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the wireless data completion and prediction method as described above.

[0018] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the wireless data completion and prediction method as described above.

[0019] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the wireless data completion and prediction method as described above.

[0020] The present invention provides a method, apparatus, device, medium, and product for wireless data completion and prediction. It standardizes input wireless data to obtain standardized wireless data. For each standard variable in the standardized wireless data, it determines the corresponding phase space vector. It constructs a knowledge graph and its corresponding structural encoding, and determines the allowed interaction neighborhood for each standard variable based on the knowledge graph. For each phase space vector, it inputs the phase space vector and its corresponding allowed interaction neighborhood into a knowledge graph relational network to obtain the micro-update amount corresponding to the phase space vector output by the knowledge graph relational network. The knowledge graph relational network is constructed based on the knowledge graph and structural encoding. It determines a global context sequence based on all phase space vectors and inputs the global context sequence into a time-series network to obtain the node-level macro-update amount corresponding to each standard variable output by the time-series network. Based on the decoder, all micro-update amounts, and all node-level macro-update amounts, it determines the data completion value and future time-step prediction value corresponding to the input wireless data. The knowledge graph relational network, time-series network, and decoder are jointly trained based on sample wireless data and the missing information corresponding to the sample wireless data. The technical solution of this invention is compatible with various wireless data. It constructs a unified physical constraint and domain knowledge framework through knowledge graphs. The combination of knowledge graph relationship networks and temporal networks improves robustness in complex scenarios such as systemic blind spots and cross-scenario missing data. Furthermore, by fusing micro-level update quantities and node-level macro-level update quantities, it achieves accurate completion and prediction of wireless data. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is one of the flowcharts illustrating the wireless data completion and prediction method provided by the present invention.

[0023] Figure 2 This is the second flowchart illustrating the wireless data completion and prediction method provided by the present invention.

[0024] Figure 3 This is a schematic diagram of the structure of the wireless data completion and prediction device provided by the present invention.

[0025] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] To address the aforementioned problems in the prior art, this invention provides a method for wireless data completion and prediction. Figure 1 This is one of the flowcharts illustrating the wireless data completion and prediction method provided by the present invention, such as... Figure 1 As shown, the method includes the following steps 110 to 160.

[0028] Step 110: Standardize the input wireless data to obtain the standardized wireless data corresponding to the input wireless data.

[0029] Specifically, since the input wireless data may be of multiple types, scales, and missing values, it is necessary to standardize the input wireless data to obtain the standardized wireless data corresponding to the input wireless data. The input wireless data includes, but is not limited to, network / wireless telemetry indicators, cross-layer counters, link quality / feedback, latency / throughput / resource usage statistics, equipment status, environmental / sensor measurements, industrial process parameters, traffic flow indicators, and other data.

[0030] Step 120: For each standard variable data in the standardized wireless data, determine the phase space vector corresponding to the standard variable data.

[0031] Specifically, for each standard variable data in standardized wireless data, a time delay embedding of the standard variable data can be constructed to obtain the phase space vector corresponding to the standard variable data, which contains local dynamic information.

[0032] For example, standard variable data At any moment phase space vectors It can be expressed by the following formula: in, Indicates time Standard variable data The standardized value, Indicates time Standard variable data The standardized value, Indicates time Standard variable data The standardized value, Indicates time Standard variable data The standardized value, Indicates the delay step size. Indicates the embedding dimension. This indicates transpose. When historical data is insufficient, zero-padding or boundary padding can be used.

[0033] It should be noted that single-moment observations are often insufficient to characterize system dynamics (e.g., control loop lag, inertial response, queue accumulation, gradual environmental changes, etc.). Time-delay embedding combines the "current value + several historical lag values" into a state vector, which is equivalent to constructing a "phase space coordinate" for the system. This allows subsequent networks to directly perceive trends, periods, inertia, and lag effects, thus better reflecting the evolution of real physics / processes.

[0034] Step 130: Construct the knowledge graph corresponding to the standardized wireless data and the structure encoding corresponding to the knowledge graph, and determine the allowed interaction neighborhoods corresponding to each of the standard variable data based on the knowledge graph.

[0035] Specifically, a knowledge graph corresponding to standardized wireless data and its corresponding structural encoding can be constructed to provide a stable dependency skeleton and interpretable constraints for subsequent computations. Furthermore, the permissible interaction neighborhoods corresponding to each standardized variable data can be determined based on the knowledge graph.

[0036] For example, a knowledge graph can be represented as ,in, This represents the set of nodes in a knowledge graph. This represents the set of edges in a knowledge graph. An interactive neighborhood can be represented as... Furthermore, it is possible to define that subsequent information transmissions only allow information to travel along... The structure prior is propagated through the edges, thereby injecting the structural prior into the subsequent knowledge graph relation network in the form of "hard mask / hard domain".

[0037] Furthermore, rule constraints can be injected into subsequent calculations in the form of hard or semi-hard constraints. For example, interaction terms that violate known monotonicity, range constraints, or coupling relationships can be masked, pruned, projected, or penalized to reduce noise propagation and enhance stable inference capabilities in the blind zone. Step 130 is equivalent to "allowing propagation only along defined causal / logical links" in engineering, prohibiting irrelevant variables from "randomly pulling" each other in the blind zone; and rule constraints can promptly remove inferences that clearly do not conform to the mechanism, making the completed trajectory closer to the physically feasible region.

[0038] Step 140: For each phase space vector, determine the micro-update amount corresponding to the phase space vector based on the knowledge graph relation network, the phase space vector, and the allowed interaction neighborhood corresponding to the phase space vector; the knowledge graph relation network is constructed based on the knowledge graph and the structure encoding.

[0039] Specifically, after obtaining the knowledge graph and structural encoding, an initial knowledge graph relation network can be constructed based on the knowledge graph and structural encoding, and then the knowledge graph relation network can be obtained through joint training with the temporal network, gating network and decoder.

[0040] Furthermore, for each spatial vector, the phase space vector and its corresponding allowed interactive neighborhood can be input into the knowledge graph relation network to obtain the micro-update quantity corresponding to the phase space vector output by the knowledge graph relation network.

[0041] Step 150: Determine the global context sequence based on all the phase space vectors, and determine the node-level macroscopic update amount corresponding to each of the standard variable data based on the global context sequence and the time series network.

[0042] Specifically, the global context sequence can be determined based on all phase space vectors. Specifically, all phase space vectors can be aggregated at each time step to form the global context sequence. Aggregation methods can include, but are not limited to, splicing and flattening, pooling aggregation, weighted aggregation, and attention aggregation.

[0043] Furthermore, the global context sequence can be input into the time series network to obtain the node-level macroscopic update quantities corresponding to each standard variable data output by the time series network.

[0044] Step 160: Based on the decoder, all the micro-updates and all the node-level macro-updates, determine the data completion value and future time step prediction value corresponding to the input wireless data; wherein, the knowledge graph relationship network, the time series network and the decoder are jointly trained based on the sample wireless data and the missing indication information corresponding to the sample wireless data.

[0045] Specifically, the data completion value and future time step prediction value corresponding to the input wireless data can be determined based on the decoder, all micro-updates, and all node-level macro-updates. The knowledge graph relation network, temporal network, and decoder are jointly trained based on the sample wireless data and the missing information corresponding to the sample wireless data.

[0046] During the joint training phase, missing information (e.g., a missing mask) corresponding to the sample wireless data can be identified. This missing information includes the missing location and missing pattern. Missing patterns may include, but are not limited to, random missing data, continuous time block missing data, and structured blind spots (simultaneous missing of target variable data and its proxy variable data within a continuous interval). The missing information can be used to identify invisible / unusable observations in the infinite sample input data. There may also be existing missing entries in the infinite sample input data. The values ​​of these existing missing entries can be filled with preset padding values, such as zero-padding with the normalized mean or other padding values ​​that do not introduce significant bias. The missing data is not "truly zero," but rather "invisible." By explicitly identifying the missing location and using neutral padding values, "observation interruption" can be distinguished from "system state change," preventing the model from treating missing data as real physical changes; it also provides reliability clues for subsequent gating fusion. (During the application phase, existing missing entries in the input wireless data can also be identified and filled). In addition, missing information can optionally be used as an additional feature to be concatenated with data features, but this is not a necessary limitation. Further, after obtaining the sample wireless data and the missing information corresponding to the sample wireless data, a set of supervision indexes can be determined based on the missing information, and the loss can be calculated only on the outputs on the set of supervision indexes. The loss function may include: (1) a scale error term: such as mean square error or its variants, used to constrain numerical accuracy; (2) a shape consistency term: such as a trend / shape alignment term based on cosine similarity, used to improve the consistency and stability of trajectory shape in long missing block scenarios; the weights of each loss term can be set as needed, and the embodiments of the present invention do not make specific limitations here. The advantage of this loss function setting is that the scale error constrains "whether the value is correct", and the shape consistency constrains "whether the trend / pattern is correct". In long missing blocks, simple point-to-point numerical fitting is prone to drift, while shape constraints are closer to the continuity requirements of system evolution.

[0047] The wireless data completion and prediction method provided by this invention standardizes the input wireless data to obtain standardized wireless data. For each standard variable in the standardized wireless data, the phase space vector corresponding to the standard variable is determined. A knowledge graph and its corresponding structural encoding are constructed, and the allowed interaction neighborhoods corresponding to each standard variable are determined based on the knowledge graph. For each phase space vector, the phase space vector and its corresponding allowed interaction neighborhood are input into a knowledge graph relation network to obtain the micro-update amount corresponding to the phase space vector output by the knowledge graph relation network. The knowledge graph relation network is constructed based on the knowledge graph and structural encoding. A global context sequence is determined based on all phase space vectors, and the global context sequence is input into a time-series network to obtain the node-level macro-update amount corresponding to each standard variable output by the time-series network. Based on the decoder, all micro-update amounts, and all node-level macro-update amounts, the data completion value and future time-step prediction value corresponding to the input wireless data are determined. The knowledge graph relation network, time-series network, and decoder are jointly trained based on sample wireless data and the missing information corresponding to the sample wireless data. The technical solution of this invention is compatible with various wireless data. It constructs a unified physical constraint and domain knowledge framework through knowledge graphs. The combination of knowledge graph relationship networks and temporal networks improves robustness in complex scenarios such as systemic blind spots and cross-scenario missing data. Furthermore, by fusing micro-level update quantities and node-level macro-level update quantities, it achieves accurate completion and prediction of wireless data.

[0048] In one embodiment, the knowledge graph relationship network is used for: For each phase space vector, the phase space vector is mapped to the positive latent space to obtain a non-negative latent state vector; Normalize the non-negative potential state vector on the positive half-unit hypersphere to obtain the direction components located on the positive half-unit hypersphere; The query vector, key vector, and value vector are determined based on the structural encoding, the allowed interactive neighborhood corresponding to the phase space vector, and the directional components. The micro-update amount corresponding to the phase space vector is determined based on the query vector, the key vector, and the value vector.

[0049] Specifically, the following processes exist within the relational network of a knowledge graph: For each phase space vector, the phase space vector is mapped to the positive latent space to obtain a non-negative latent state vector. Then, the non-negative latent state vector can be normalized on a positive semi-unit hypersphere to obtain the directional components located on the positive semi-unit hypersphere. The positive latent space is used to ensure geometric consistency and constraint compliance. The final output still needs to return to the original dimensions usable in engineering (such as rate, delay, quality indicators, state coding, etc.) for downstream system consumption.

[0050] Furthermore, the query vector, key vector, and value vector can be determined based on the structural encoding, the allowed interactive neighborhood, and the directional components corresponding to the phase space vector. After obtaining the query vector, key vector, and value vector, the micro-update amount of the micro-Riemannian information flow corresponding to the phase space vector can be determined based on the query vector, key vector, and value vector.

[0051] It's important to note that many phase space vectors are inherently non-negative in terms of "intensity / resource / count / capacity" (e.g., power, resource usage, counter, rate intensity, reliability intensity, etc.). Even if the original observations have differences in sign or dimensions, positive potential boosting can unify them into "non-negative intensity coordinates," ensuring that subsequent updates always occur within an interpretable non-negative domain, reducing the occurrence of physically incomprehensible negative values. In many scenarios, "amplitude / intensity" is easily affected by external factors and drifts (e.g., large-scale fading and power control adjustments in wireless, load changes and sensor gain changes in industry), but the system's "mode / trend / relative structure" often better reflects the internal mechanisms (e.g., interference coupling, control loops, process dependencies). Direction / mode length decoupling is equivalent to separating "shape" and "energy": first, use direction to find "whose mode is more similar," then use mode length to retain "how strong the intensity is," making local propagation more consistent with physical semantics. The micro-update values ​​obtained in the above embodiments are not simply numerical interpolations, but more like "physical quantity corrections propagated along interpretable local paths", thus further improving the interpretability of the micro-update values.

[0052] For example, the positive latent space can be represented as Non-negative latent state vector It can be expressed by the following formula: in, This represents a nonlinear function that maintains positiveness. This nonlinear function can be a soft-plus function, an exponential map, or another monotonic positiveizing function.

[0053] This directional component It can be expressed by the following formula: in, The strength / amplitude (magnitude information) of the non-negative latent state vector is represented to achieve decoupling between scale and pattern. Optionally, the angular similarity of the directional components can form a consistent or monotonically corresponding relationship with the Fisher-Rao metric in information geometry, thereby making the similarity calculation insensitive to amplitude fluctuations.

[0054] In one embodiment, determining the query vector, key vector, and value vector based on the structure encoding, the allowed interaction neighborhood corresponding to the phase space vector, and the direction component includes: Within the allowed interaction neighborhood corresponding to the phase space vector, an initial query vector and an initial key vector are generated based on the directional components; Based on the structure encoding, the initial query vector, and the initial key vector, the query vector and the key vector are determined; Local interaction weights are determined based on the angular similarity between the directional component and the directional components of the neighbors in the allowed interaction neighborhood. The value vector is determined based on the local interaction weights and the initial value vectors corresponding to the neighbor directional components in the allowed interaction neighborhood.

[0055] Specifically, within the allowed interaction neighborhood corresponding to the phase space vector, an initial query vector and an initial key vector can be generated based on the direction components, and the structure encoding can be injected into the initial query vector and the initial key vector to obtain the query vector and the key vector. The injection of the structure encoding enhances the structural recognizability.

[0056] Furthermore, the angular similarity between the directional component and the directional components of its neighbors in the allowed interaction neighborhood can be calculated to obtain the local interaction weight corresponding to that directional component. The angular similarity can be calculated using dot product similarity or cosine similarity methods. This process can be viewed as "performing pattern matching within the allowed physical / logical domain."

[0057] Furthermore, the initial value vectors corresponding to the directional components of neighbors in the allowed interaction neighborhood can be weighted and aggregated based on the local interaction weights to obtain the value vector.

[0058] The initial value vector is generated from the non-negative latent state vector (or its linear mapping) to preserve intensity information. This allows for a scale-robust and intensity-consistent micro-update mechanism based on "matching neighbors by direction and aggregating contributions by intensity." For example, the non-negative latent state vector can be represented as... The micro-update quantity can be expressed as .

[0059] In the above embodiments, the determination process of local interaction weights only compares the direction angles among neighbors that are allowed by the prior structure, avoiding irrelevant nodes from being incorrectly aggregated due to accidental similarity in amplitude; at the same time, angle similarity is closer to "shape consistency", which is beneficial to maintaining a reasonable trajectory shape even when the amplitude is unstable or severely missing. Since whether a neighbor is "similar" is determined by direction (pattern / shape alignment), and how much influence a neighbor has is determined by intensity (energy / resource / intensity contribution), weighted aggregation based on the initial value vectors corresponding to the directional components of neighbors can accurately obtain the value vector.

[0060] In one embodiment, the time-series network is used for: The global context sequence is mapped to the positive latent space to obtain the global positive representation corresponding to the global context sequence; The macroscopic inertia update amount is determined based on the global positive representation and the temporal network; The macroscopic inertia update is broadcast or mapped to the nodes corresponding to each of the standard variable data to obtain the node-level macroscopic update for each of the standard variable data.

[0061] Specifically, Figure 2 This is the second flowchart illustrating the wireless data completion and prediction method provided by the present invention, as shown below. Figure 2 As shown, the global context sequence can be mapped to a positive latent space to obtain a global positive representation corresponding to the global context sequence, so that it can be fused with the micro-update quantity within the same latent space / tangent space framework. For example, the global context sequence can be represented as... The global positive representation can be expressed as .

[0062] Furthermore, the global positive representation can be input into the temporal modeling unit in the temporal network to obtain the macroscopic inertia update output by the temporal modeling unit. The temporal modeling unit can be a Long Short-Term Memory (LSTM) network, a Gated Recurrent Unit (GRU), a gated state-space model, or other sequential network structures. The macroscopic inertia update can take any real value in the tangent space, without being constrained by positivity, to express the inertia direction and correction trend.

[0063] Furthermore, the macroscopic inertia update can be broadcast or mapped to the nodes corresponding to each standard variable data, thus obtaining the node-level macroscopic update for each standard variable data. The node-level macroscopic update quantity allows the continuity and recoverability of state evolution to be maintained by global inertia even when evidence is insufficient in a local domain or the domain is disrupted by structural deficiencies. It should be noted that the macroscopic update quantity at each node level is the same.

[0064] In the above embodiments, the macroscopic context corresponds to the "overall system operating condition": even if local nodes are missing, other observable parts can still reflect the current load, environment, strategy, and control status, providing an "overall operating condition anchor point" for the blind zone. Macroscopic inertia is similar to "dead calculation / inertial navigation": during the blind zone, the system state often does not change instantaneously, but evolves continuously along existing trends and control inertia. Macroscopic flow characterizes this continuity, preventing the network from drifting or collapsing due to insufficient local evidence within the missing block. Broadcasting is equivalent to providing each node with the same set of "operating condition drivers and inertial trends." When a local link breaks, the global trend provides a safety net, ensuring that the completed trajectory does not experience non-physical abrupt changes.

[0065] In one embodiment, determining the data completion value and future time step prediction value corresponding to the input wireless data based on the decoder, all the micro-updates, and all the node-level macro-updates includes: For each of the standard variable data, based on the gating network and the corresponding micro-update and node-level macro-update, the positive latent state corresponding to the standard variable data is determined; the gating network is jointly trained based on the sample wireless data and the missing information corresponding to the sample wireless data. Based on all the positive latent states and the decoder, the data completion value and future time step prediction value corresponding to the input wireless data are determined.

[0066] Specifically, for each standard variable data, the positive latent state corresponding to the standard variable data can be determined based on the gating network and the corresponding micro-update and node-level macro-update quantities. The gating network can be jointly trained with the knowledge graph relation network, temporal network, and decoder based on the sample wireless data and the missing information corresponding to the sample wireless data.

[0067] Furthermore, all positive latent states are input into the decoder. The decoder maps all positive latent states back to the space corresponding to the input radio data, obtaining the data completion value and future time step prediction value corresponding to the input radio data output by the decoder. The data completion value is the data completion value for missing positions. The decoder can be a linear readout layer, a multilayer perceptron, or a multi-head decoding structure; a type-adaptive decoding head can be used for multiple types of data, but this is not a necessary limitation.

[0068] In the above embodiments, the gating network can reasonably integrate micro-level update quantities and node-level macro-level update quantities, thereby decoding the positive latent state back to the original data space and outputting accurate data completion values ​​and future time step prediction values.

[0069] In one embodiment, determining the positive latent state corresponding to the standard variable data based on the gating network and the micro-level and node-level macro-level update quantities corresponding to the standard variable data includes: The fusion threshold is determined based on the gating network, the micro-update amount corresponding to the standard variable data, and the node-level macro-update amount. Based on the fusion threshold, the micro-update amount and the node-level macro-update amount are fused to obtain the total tangent vector update amount; The positive latent state is obtained by back-mapping the total tangent vector update with the current latent state.

[0070] Specifically, such as Figure 2 As shown, for each standard variable data, the corresponding micro-level update quantity and node-level macro-level update quantity are input into the gating network to obtain the fusion threshold output by the gating network. The value of the fusion threshold ranges from 0 to 1 and can be a scalar or a dimension-wise vector. The fusion threshold is used to express the relative reliability of local structural information and global inertial information: when there is severe missing information or insufficient structural evidence, the proportion of node-level macro-level update quantity is increased; when local information is sufficient and structural constraints are effective, the proportion of micro-level update quantity is increased.

[0071] Furthermore, the micro-level update quantity and the node-level macro-level update quantity can be fused based on the fusion threshold (the fusion method can be a convex combination or its equivalent form, which is not specifically limited in this embodiment of the invention) to obtain the total tangent vector update quantity.

[0072] After obtaining the total tangent vector update, a shrinkage mapping can be performed between the total tangent vector update and the current latent state to obtain a positive latent state. The shrinkage mapping can use Softplus shrinkage, exponential shrinkage, or other mapping functions that maintain positivity and numerical stability, ensuring that the updated latent state remains within the positive latent space. This reduces potential geometric deviations (such as non-physical "shortcut" biases across missing blocks) that may occur during linear extrapolation in Euclidean space. The shrinkage update is equivalent to "predicting one step according to velocity first, then projecting back into the feasible region," ensuring that the state does not run into infeasible regions such as negative intensity or unexplainable energy, and suppressing "tunneling inconsistencies" caused by the Euclidean straight lines pulling the ends of the missing block.

[0073] For example, the micro-update quantity can be expressed as The node-level macro update quantity can be expressed as The fusion threshold can be expressed as A positive potential state can be represented as The total tangent vector update can be expressed by the following formula: in, This represents the total tangent vector update amount. This indicates element-wise multiplication.

[0074] In the above embodiments, the gated network can be regarded as a "reliability arbiter": when the local domain is masked or amplified by noise in the blind zone, the micro-propagation is unreliable, and it relies more on macro-inertia; when the local structure is still usable, it absorbs more micro-corrections to improve the precision. The fusion of micro-updates and node-level macro-updates in the "cut space" is equivalent to first treating both types of effects as "change rate / change direction", and then combining them by weighting according to reliability, avoiding inconsistencies caused by directly scrambling in the state space.

[0075] In one embodiment, constructing the knowledge graph corresponding to the standardized wireless data and the structural encoding corresponding to the knowledge graph includes: The node set is determined based on the standard variable data in the standardized wireless data. The edge set is determined based on the dependencies between the standard variable data; the dependencies include one or more of the following: process / hierarchy dependency, physical / mechanism dependency, control and feedback dependency, and scenario and entity association; The knowledge graph is constructed based on the set of nodes and the set of edges; Determine the structural encoding corresponding to the knowledge graph; the structural encoding includes one or more of topological location encoding, type identifier encoding, and scene / context encoding.

[0076] Specifically, the node set is determined based on the standard variable data in the standardized wireless data. This set of nodes can represent variables, entities, or states (such as metrics, counters, sensor values, control values, scene entities, etc.). The set of edges can also be determined based on the dependencies between standard variable data. Edge sets are used to represent dependencies between standard variable data. Dependencies include one or more of the following: process / hierarchical dependencies, physical / mechanistic dependencies, and control / feedback dependencies. Process / hierarchical dependencies include, for example, processing links, protocol stacks, pipeline sequences, and causal chains. Physical / mechanistic dependencies include, for example, monotonic relationships, conservation / constraints, and capacity / bandwidth / interference relationships. Control / feedback dependencies include, for example, closed-loop control, retransmission / feedback mechanisms, and scheduling and resource allocation relationships. Scenario-entity relationships include, for example, environmental state → parameter change → observation change.

[0077] Furthermore, a knowledge graph can be constructed based on the set of nodes and the set of edges. The knowledge graph can be obtained from expert rules, standard documents, business logic, system design documents or historical structure statistics; the knowledge graph can be regarded as a time-invariant or slowly changing structural prior, which can be pre-configured offline and remain available during observation interruptions or blind spots. The structural encoding corresponding to the knowledge graph can also be determined; the structural encoding includes one or more of topological position encoding, type identification encoding and scene / context encoding. Among them, (1) topological position encoding: construct a Laplace matrix (or its normalized form) based on the adjacency relationship of the knowledge graph, extract feature vectors as topological position encoding, and use them to characterize the structural role of nodes in the graph; (2) type identification encoding: attach type encoding to different types of nodes or different data types to distinguish their statistical attributes and semantic attributes; (3) scene / context encoding: map scene entities or system context (e.g., device type, environmental parameters, operating mode, etc.) into encoding and inject them into relevant nodes to enhance cross-scene adaptability. It should be noted that the structural encoding can also include other dimensions, which are not specifically limited in this embodiment of the invention.

[0078] In the above embodiments, the knowledge graph is equivalent to the system's "structural skeleton / circuit diagram": even if some sensors or counters fail, the system's processing chain and control logic still exist. It provides an interpretable channel for "which quantities should affect which quantities, and along what path information should be transmitted," avoiding inferences based solely on accidental correlations in blind spots. Structural encoding is equivalent to assigning coordinates and role labels to the "skeleton": different nodes occupy different positions in links, loops, and processes, and their available information and scope of influence differ. Through position / type / scenario encoding, the model can more easily learn that "the same numerical change represents different meanings under different roles."

[0079] In one embodiment, the standardization process of the input wireless data to obtain standardized wireless data corresponding to the input wireless data includes: For input wireless data of continuous numerical type, one of the following processing methods is performed: standardization, normalization, and robust scaling, to obtain the corresponding first sub-standardized data. For discrete / categorical data in the input wireless data, embedding or one-hot coding is performed to obtain the corresponding second sub-normalized data; For input wireless data of binary / symbolic type, one of the following methods is used: embedding, one-hot coding, and threshold mapping to obtain the corresponding third sub-normalized data; The standardized wireless data is determined based on the first sub-standardized data, the second sub-standardized data, and the third sub-standardized data.

[0080] Specifically, the input wireless data can include various data types, such as continuous numerical, discrete / categorical, and binary / symmetric. For continuous numerical data, one of the following processing methods is applied: standardization, normalization, or robust scaling, to obtain the corresponding first sub-standardized data. For discrete / categorical data, embedding or one-hot coding is performed to obtain the corresponding second sub-standardized data. For binary / symmetric data, one of the following processing methods is applied: embedding, one-hot coding, or threshold mapping, to obtain the corresponding third sub-standardized data.

[0081] Optionally, outliers in the first, second, and third sub-normalized data can be truncated, smoothed, or denoised to reduce the impact of extreme perturbations on subsequent phase space construction.

[0082] Furthermore, standardized wireless data can be determined based on the first sub-standardized data, the second sub-standardized data, and the third sub-standardized data.

[0083] In the above embodiments, it is equivalent to transforming "observations with different dimensions and different coding methods" into comparable measurements under the same coordinate system, avoiding the subsequent similarity and propagation being dominated by a certain type of dimension or coding form, and improving the usability across indicators and scenarios from an engineering perspective.

[0084] The technical solution of the present invention is illustrated below through a complete example: The operation and maintenance platform accesses the uplink key performance indicator data stream of a certain 5G base station cell, with a time sampling interval of, for example, 1 second. Several telemetry variables can be used as a set of nodes (or a set of variables), for example, 34 dimensions, including: (1) target variable: physical layer uplink throughput; (2) strongly correlated proxy variables: modulation and coding strategy index, physical resource block occupancy rate, etc.; (3) weakly correlated but physically semantic peripheral variables: path loss, power margin, reference signal received power, etc. Optionally, variables related to power control / reliability / feedback (e.g., transmit power, block error rate, acknowledgment (ACK) or negative acknowledgment (NACK)) can also be included to enhance physical observability in blind zones.

[0085] The variables are preprocessed by type and unified to a comparable numerical domain: continuous numerical variables are standardized / normalized / robustly scaled; discrete or categorical variables can be embedded or scaled after numericalization; outliers can be optionally pruned or smoothed to reduce the impact of extreme perturbations on subsequent phase space construction.

[0086] In actual operation, the operation and maintenance platform can detect the absence of consecutive time blocks through field visibility / reporting integrity, and determine the blind zone interval accordingly. For example, in a certain operation log, a continuous invisible interval may be observed from the 100th second to the 132nd second (inclusive). At this time, the target throughput and its strongly correlated proxy variable are missing at the same time, making it difficult for traditional imputation methods that rely on statistical correlation to work effectively.

[0087] Missing information (missing mask) is generated based on the missing location, and missing entries are filled with a preset neutral padding value (e.g., 0 for standardized data) to distinguish between "invisible" and "true zero". The missing mask is used for subsequent gating reliability judgment, output quality identification and consistency verification; in some implementations, the missing mask can also be used as an additional feature concatenation input, but this is not a necessary limitation.

[0088] To characterize dynamic features such as control loop lag, queue accumulation, and gradual changes in the wireless channel, a time delay embedding (phase space vector) is constructed for each standard variable data.

[0089] Knowledge graphs can also be constructed, which should contain at least the following types of dependencies (examples, not limitations): cross-layer processing chain dependencies; resource and link adaptive dependencies; power control and reliability feedback dependencies. It is important to emphasize that even if observations of nodes such as modulation and coding strategies / schemes and physical resource blocks are missing in the blind zone, the knowledge graph structure itself still exists and is usable, serving as a "dependency skeleton" during the blind zone, providing interpretable information pathways beyond statistical correlation.

[0090] The knowledge graph is generated by structural encoding, such as topological location encoding, type / semantic encoding, and optional scene / context encoding; the encoding dimensions and injection methods can be selected according to the implementation and are not limited.

[0091] Permissible interaction neighborhoods are determined based on the edge set of the knowledge graph. Subsequent micro-information transmission is only allowed to propagate along the edges in the edge set (hard mask / hard neighborhood), thereby avoiding the "accidental correlation propagation" between irrelevant variables during the blind zone. Optional rule constraints can be defined for consistency control (not constituting a necessary limitation).

[0092] Phase space vectors are mapped to positive latent space to obtain non-negative latent representations, and direction / mode length decoupling and spherical normalization are performed. Within the allowed interaction neighborhood, geometrically consistent angular similarity (corresponding to Fisher-Rao consistent angular similarity) is calculated based on spherical direction components to obtain local interaction weights. Local aggregation is then performed while preserving intensity information to generate micro-update quantities. Although throughput and strong proxies (Modulation and Coding Scheme (MCS) and Physical Resource Block (PRB)) are not visible, physical quantities such as path loss are usually still observable. The model can geometrically match the directional patterns of throughput nodes and these physical quantity nodes on the spherical manifold to obtain trend correction clues.

[0093] Phase space vectors are converged at each time step to form a global context sequence, which is then input into the temporal modeling unit in the temporal network to generate a macroscopic inertial update. This macroscopic update is then broadcast or mapped to each node to obtain a node-level macroscopic update. The macroscopic inertial component is mainly used to maintain evolutionary continuity during the blind zone: the temporal memory unit outputs the inertial update based on the global operating context before the start of the blind zone, so that the first half of the blind zone can still maintain a relatively smooth continuous evolution (avoiding unreasonable abrupt changes) even when local evidence is insufficient.

[0094] Microscopic and node-level macroscopic updates are input into a gating network, and a fusion threshold (0 to 1, which can be a scalar or a dimension-wise vector) is output to express the relative credibility of "local structural evidence" and "global inertial evidence". Then, convex combination fusion is performed in the tangent space, and the fused update is subjected to a positive retraction mapping to update the latent state, so that the updated positive latent state is kept in the positive latent space, thereby suppressing non-physical deviations across missing blocks and improving numerical stability.

[0095] The positive latent state is input into the decoder and mapped back to the target data space, outputting the throughput interpolation trajectory of the blind zone interval (e.g., from the 100th to the 132nd second); the decoder can adopt a linear readout layer or a multilayer perceptron structure; a type-adaptive decoding head can also be used for multiple types of data (not a necessary limitation).

[0096] After outputting the completed data values ​​and predicted values ​​for future time steps, a consistency check can be optionally performed: for example, checking the non-negativity and reasonable range of throughput, and pruning, marking, or alerting obvious anomalies; a quality identifier can also be output along with the interpolation results so that downstream operation and maintenance systems can display credibility or select strategies (this is an optional implementation and does not constitute a necessary limitation). In the systemic blind spot where the target throughput and its strongly correlated proxy variables (such as MCS, PRB) are not visible at the same time, the protocol / physical dependency skeleton, spherical shape matching, and macroscopic inertia support provided by the knowledge graph can still be relied upon to achieve continuous interpolation and recovery of the throughput trajectory, and provide usable data support for subsequent operation and maintenance analysis.

[0097] The wireless data completion and prediction apparatus provided by the present invention will be described below. The wireless data completion and prediction apparatus described below can be referred to in correspondence with the wireless data completion and prediction method described above.

[0098] Figure 3 This is a schematic diagram of the structure of the wireless data completion and prediction device provided by the present invention, as shown below. Figure 3 As shown, the wireless data completion and prediction device 300 includes the following modules: The standardization module 310 is used to standardize the input wireless data to obtain the standardized wireless data corresponding to the input wireless data. Vector module 320 is used to determine the phase space vector corresponding to each standard variable data in the standardized wireless data; The knowledge graph module 330 is used to construct a knowledge graph corresponding to the standardized wireless data and a structure encoding corresponding to the knowledge graph, and to determine the allowed interaction neighborhoods corresponding to each of the standard variable data based on the knowledge graph. Micro-module 340 is used to determine the micro-update amount corresponding to each phase space vector based on the knowledge graph relation network, the phase space vector, and the allowed interaction neighborhood corresponding to the phase space vector; the knowledge graph relation network is constructed based on the knowledge graph and the structure encoding; The macro module 350 is used to determine the global context sequence based on all the phase space vectors, and to determine the node-level macro update amount corresponding to each of the standard variable data based on the global context sequence and the time series network. The completion prediction module 360 ​​is used to determine the data completion value and future time step prediction value corresponding to the input wireless data based on the decoder, all the micro-updates and all the node-level macro-updates; wherein, the knowledge graph relationship network, the time series network and the decoder are jointly trained based on the sample wireless data and the missing indication information corresponding to the sample wireless data.

[0099] In one embodiment, the knowledge graph relationship network is used for: For each phase space vector, the phase space vector is mapped to the positive latent space to obtain a non-negative latent state vector; Normalize the non-negative potential state vector on the positive half-unit hypersphere to obtain the direction components located on the positive half-unit hypersphere; The query vector, key vector, and value vector are determined based on the structural encoding, the allowed interactive neighborhood corresponding to the phase space vector, and the directional components. The micro-update amount corresponding to the phase space vector is determined based on the query vector, the key vector, and the value vector.

[0100] In one embodiment, the knowledge graph relationship network is further used for: Within the allowed interaction neighborhood corresponding to the phase space vector, an initial query vector and an initial key vector are generated based on the directional components; Based on the structure encoding, the initial query vector, and the initial key vector, the query vector and the key vector are determined; Local interaction weights are determined based on the angular similarity between the directional component and the directional components of the neighbors in the allowed interaction neighborhood. The value vector is determined based on the local interaction weights and the initial value vectors corresponding to the neighbor directional components in the allowed interaction neighborhood.

[0101] In one embodiment, the time-series network is used for: The global context sequence is mapped to the positive latent space to obtain the global positive representation corresponding to the global context sequence; The macroscopic inertia update amount is determined based on the global positive representation and the temporal network; The macroscopic inertia update is broadcast or mapped to the nodes corresponding to each of the standard variable data to obtain the node-level macroscopic update for each of the standard variable data.

[0102] In one embodiment, the completion prediction module 360 ​​is specifically used for: For each of the standard variable data, based on the gating network and the corresponding micro-update and node-level macro-update, the positive latent state corresponding to the standard variable data is determined; the gating network is jointly trained based on the sample wireless data and the missing information corresponding to the sample wireless data. Based on all the positive latent states and the decoder, the data completion value and future time step prediction value corresponding to the input wireless data are determined.

[0103] In one embodiment, the completion prediction module 360 ​​is further configured to: The fusion threshold is determined based on the gating network, the micro-update amount corresponding to the standard variable data, and the node-level macro-update amount. Based on the fusion threshold, the micro-update amount and the node-level macro-update amount are fused to obtain the total tangent vector update amount; The positive latent state is obtained by back-mapping the total tangent vector update with the current latent state.

[0104] In one embodiment, the knowledge graph module 330 is specifically used for: The node set is determined based on the standard variable data in the standardized wireless data. The edge set is determined based on the dependencies between the standard variable data; the dependencies include one or more of the following: process / hierarchy dependency, physical / mechanism dependency, control and feedback dependency, and scenario and entity association; The knowledge graph is constructed based on the set of nodes and the set of edges; Determine the structural encoding corresponding to the knowledge graph; the structural encoding includes one or more of topological location encoding, type identifier encoding, and scene / context encoding.

[0105] In one embodiment, the standardization module 310 is specifically used for: For input wireless data of continuous numerical type, one of the following processing methods is performed: standardization, normalization, and robust scaling, to obtain the corresponding first sub-standardized data. For discrete / categorical data in the input wireless data, embedding or one-hot coding is performed to obtain the corresponding second sub-normalized data; For input wireless data of binary / symbolic type, one of the following methods is used: embedding, one-hot coding, and threshold mapping to obtain the corresponding third sub-normalized data; The standardized wireless data is determined based on the first sub-standardized data, the second sub-standardized data, and the third sub-standardized data.

[0106] The wireless data completion and prediction device provided by this invention standardizes the input wireless data to obtain standardized wireless data. For each standard variable data in the standardized wireless data, it determines the phase space vector corresponding to the standard variable data. It constructs a knowledge graph corresponding to the standardized wireless data and a structure encoding corresponding to the knowledge graph, and determines the allowed interaction neighborhoods corresponding to each standard variable data based on the knowledge graph. For each phase space vector, it inputs the phase space vector and its corresponding allowed interaction neighborhood into a knowledge graph relation network to obtain the micro-update amount corresponding to the phase space vector output by the knowledge graph relation network. The knowledge graph relation network is constructed based on the knowledge graph and structure encoding. It determines a global context sequence based on all phase space vectors and inputs the global context sequence into a time-series network to obtain the node-level macro-update amount corresponding to each standard variable data output by the time-series network. Based on the decoder, all micro-update amounts, and all node-level macro-update amounts, it determines the data completion value and the predicted value for the future time step corresponding to the input wireless data. The knowledge graph relation network, time-series network, and decoder are jointly trained based on sample wireless data and the missing information corresponding to the sample wireless data. The technical solution of this invention is compatible with various wireless data. It constructs a unified physical constraint and domain knowledge framework through knowledge graphs. The combination of knowledge graph relationship networks and temporal networks improves robustness in complex scenarios such as systemic blind spots and cross-scenario missing data. Furthermore, by fusing micro-level update quantities and node-level macro-level update quantities, it achieves accurate completion and prediction of wireless data.

[0107] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a wireless data completion and prediction method, which includes: The input wireless data is standardized to obtain the standardized wireless data corresponding to the input wireless data. For each standard variable in the standardized wireless data, determine the phase space vector corresponding to the standard variable data; Construct a knowledge graph corresponding to the standardized wireless data and a structure encoding corresponding to the knowledge graph, and determine the allowed interaction neighborhoods corresponding to each of the standardized variable data based on the knowledge graph; For each phase space vector, the micro-update amount corresponding to the phase space vector is determined based on the knowledge graph relation network, the phase space vector, and the allowed interaction neighborhood corresponding to the phase space vector; the knowledge graph relation network is constructed based on the knowledge graph and the structure encoding; A global context sequence is determined based on all the phase space vectors, and the node-level macroscopic update amount corresponding to each of the standard variable data is determined based on the global context sequence and the time series network. Based on the decoder, all the micro-updates, and all the node-level macro-updates, the data completion value and future time step prediction value corresponding to the input wireless data are determined; wherein, the knowledge graph relationship network, the time series network, and the decoder are jointly trained based on the sample wireless data and the missing indication information corresponding to the sample wireless data.

[0108] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0109] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the wireless data completion and prediction methods provided by the above methods, the method comprising: The input wireless data is standardized to obtain the standardized wireless data corresponding to the input wireless data. For each standard variable in the standardized wireless data, determine the phase space vector corresponding to the standard variable data; Construct a knowledge graph corresponding to the standardized wireless data and a structure encoding corresponding to the knowledge graph, and determine the allowed interaction neighborhoods corresponding to each of the standardized variable data based on the knowledge graph; For each phase space vector, the micro-update amount corresponding to the phase space vector is determined based on the knowledge graph relation network, the phase space vector, and the allowed interaction neighborhood corresponding to the phase space vector; the knowledge graph relation network is constructed based on the knowledge graph and the structure encoding; A global context sequence is determined based on all the phase space vectors, and the node-level macroscopic update amount corresponding to each of the standard variable data is determined based on the global context sequence and the time series network. Based on the decoder, all the micro-updates, and all the node-level macro-updates, the data completion value and future time step prediction value corresponding to the input wireless data are determined; wherein, the knowledge graph relationship network, the time series network, and the decoder are jointly trained based on the sample wireless data and the missing indication information corresponding to the sample wireless data.

[0110] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the wireless data completion and prediction methods provided by the methods described above, the method comprising: The input wireless data is standardized to obtain the standardized wireless data corresponding to the input wireless data. For each standard variable in the standardized wireless data, determine the phase space vector corresponding to the standard variable data; Construct a knowledge graph corresponding to the standardized wireless data and a structure encoding corresponding to the knowledge graph, and determine the allowed interaction neighborhoods corresponding to each of the standardized variable data based on the knowledge graph; For each phase space vector, the micro-update amount corresponding to the phase space vector is determined based on the knowledge graph relation network, the phase space vector, and the allowed interaction neighborhood corresponding to the phase space vector; the knowledge graph relation network is constructed based on the knowledge graph and the structure encoding; A global context sequence is determined based on all the phase space vectors, and the node-level macroscopic update amount corresponding to each of the standard variable data is determined based on the global context sequence and the time series network. Based on the decoder, all the micro-updates, and all the node-level macro-updates, the data completion value and future time step prediction value corresponding to the input wireless data are determined; wherein, the knowledge graph relationship network, the time series network, and the decoder are jointly trained based on the sample wireless data and the missing indication information corresponding to the sample wireless data.

[0111] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0112] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for completing and predicting wireless data, characterized in that, include: The input wireless data is standardized to obtain the standardized wireless data corresponding to the input wireless data. For each standard variable in the standardized wireless data, determine the phase space vector corresponding to the standard variable data; Construct a knowledge graph corresponding to the standardized wireless data and a structure encoding corresponding to the knowledge graph, and determine the allowed interaction neighborhoods corresponding to each of the standardized variable data based on the knowledge graph; For each phase space vector, the micro-update amount corresponding to the phase space vector is determined based on the knowledge graph relation network, the phase space vector, and the allowed interaction neighborhood corresponding to the phase space vector; the knowledge graph relation network is constructed based on the knowledge graph and the structure encoding; A global context sequence is determined based on all the phase space vectors, and the node-level macroscopic update amount corresponding to each of the standard variable data is determined based on the global context sequence and the time series network. Based on the decoder, all the micro-updates, and all the node-level macro-updates, the data completion value and future time step prediction value corresponding to the input wireless data are determined; wherein, the knowledge graph relationship network, the time series network, and the decoder are jointly trained based on the sample wireless data and the missing indication information corresponding to the sample wireless data.

2. The method for completing and predicting wireless data according to claim 1, characterized in that, The knowledge graph relational network is used for: For each phase space vector, the phase space vector is mapped to the positive latent space to obtain a non-negative latent state vector; Normalize the non-negative potential state vector on the positive half-unit hypersphere to obtain the direction components located on the positive half-unit hypersphere; The query vector, key vector, and value vector are determined based on the structural encoding, the allowed interactive neighborhood corresponding to the phase space vector, and the directional components. The micro-update amount corresponding to the phase space vector is determined based on the query vector, the key vector, and the value vector.

3. The method for completing and predicting wireless data according to claim 2, characterized in that, The process of determining the query vector, key vector, and value vector based on the structural encoding, the allowed interactive neighborhood corresponding to the phase space vector, and the direction component includes: Within the allowed interaction neighborhood corresponding to the phase space vector, an initial query vector and an initial key vector are generated based on the directional components; Based on the structure encoding, the initial query vector, and the initial key vector, the query vector and the key vector are determined; Local interaction weights are determined based on the angular similarity between the directional component and the directional components of the neighbors in the allowed interaction neighborhood. The value vector is determined based on the local interaction weights and the initial value vectors corresponding to the neighbor directional components in the allowed interaction neighborhood.

4. The method for completing and predicting wireless data according to claim 1, characterized in that, The time-series network is used for: The global context sequence is mapped to the positive latent space to obtain the global positive representation corresponding to the global context sequence; The macroscopic inertia update amount is determined based on the global positive representation and the temporal network; The macroscopic inertia update is broadcast or mapped to the nodes corresponding to each of the standard variable data to obtain the node-level macroscopic update for each of the standard variable data.

5. The method for completing and predicting wireless data according to claim 1, characterized in that, The process of determining the data completion value and future time step prediction value corresponding to the input wireless data based on the decoder, all the micro-updates, and all the node-level macro-updates includes: For each of the standard variable data, based on the gating network and the corresponding micro-update and node-level macro-update, the positive latent state corresponding to the standard variable data is determined; the gating network is jointly trained based on the sample wireless data and the missing information corresponding to the sample wireless data. Based on all the positive latent states and the decoder, the data completion value and future time step prediction value corresponding to the input wireless data are determined.

6. The method for completing and predicting wireless data according to claim 5, characterized in that, The determination of the positive latent state corresponding to the standard variable data based on the gated network and the micro-updates and node-level macro-updates corresponding to the standard variable data includes: The fusion threshold is determined based on the gating network, the micro-update amount corresponding to the standard variable data, and the node-level macro-update amount. Based on the fusion threshold, the micro-update amount and the node-level macro-update amount are fused to obtain the total tangent vector update amount; The positive latent state is obtained by back-mapping the total tangent vector update with the current latent state.

7. The method for completing and predicting wireless data according to claim 1, characterized in that, The construction of the knowledge graph corresponding to the standardized wireless data and the structural encoding corresponding to the knowledge graph includes: The node set is determined based on the standard variable data in the standardized wireless data. The edge set is determined based on the dependencies between the standard variable data; the dependencies include one or more of the following: process / hierarchy dependency, physical / mechanism dependency, control and feedback dependency, and scenario and entity association; The knowledge graph is constructed based on the set of nodes and the set of edges; Determine the structural encoding corresponding to the knowledge graph; the structural encoding includes one or more of topological location encoding, type identifier encoding, and scene / context encoding.

8. The method for completing and predicting wireless data according to any one of claims 1 to 7, characterized in that, The standardization process for the input wireless data to obtain standardized wireless data corresponding to the input wireless data includes: For input wireless data of continuous numerical type, one of the following processing methods is performed: standardization, normalization, and robust scaling, to obtain the corresponding first sub-standardized data. For discrete / categorical data in the input wireless data, embedding or one-hot coding is performed to obtain the corresponding second sub-normalized data; For input wireless data of binary / symbolic type, one of the following methods is used: embedding, one-hot coding, and threshold mapping to obtain the corresponding third sub-normalized data; The standardized wireless data is determined based on the first sub-standardized data, the second sub-standardized data, and the third sub-standardized data.

9. A device for completing and predicting wireless data, characterized in that, include: The standardization module is used to standardize the input wireless data to obtain the standardized wireless data corresponding to the input wireless data. The vector module is used to determine the phase space vector corresponding to each standard variable data in the standardized wireless data. The knowledge graph module is used to construct a knowledge graph corresponding to the standardized wireless data and a structure encoding corresponding to the knowledge graph, and to determine the allowed interaction neighborhoods corresponding to each of the standard variable data based on the knowledge graph. The micro-module is used to determine the micro-update amount corresponding to each phase space vector based on the knowledge graph relation network, the phase space vector, and the allowed interaction neighborhood corresponding to the phase space vector; the knowledge graph relation network is constructed based on the knowledge graph and the structure encoding; The macro module is used to determine the global context sequence based on all the phase space vectors, and to determine the node-level macro update amount corresponding to each of the standard variable data based on the global context sequence and the time series network. The completion prediction module is used to determine the data completion value and future time step prediction value corresponding to the input wireless data based on the decoder, all the micro-updates, and all the node-level macro-updates; wherein, the knowledge graph relationship network, the time series network, and the decoder are jointly trained based on the sample wireless data and the missing indication information corresponding to the sample wireless data.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the wireless data completion and prediction method as described in any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the wireless data completion and prediction method as described in any one of claims 1 to 8.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the wireless data completion and prediction method as described in any one of claims 1 to 8.

Images