An artificial intelligence-based communication chip signal interference prediction and optimization system

Abstract

The application discloses a communication chip signal interference prediction and optimization system based on artificial intelligence, comprising the following modules: a data acquisition and preprocessing module, which is used for multidimensional communication state data and preprocessing; a communication behavior atlas construction module, which is used for constructing a communication behavior atlas based on multidimensional communication state data; a causal structure learning module, which is used for generating a preliminary interference causal diagram based on an improved NOTEARS algorithm and constructing a causal potential energy tensor; an interference classification and risk identification module, which is used for outputting interference type labels and risk level labels through a Gumbel-Softmax variational autoencoder model; a control strategy generation module, which is used for dividing interference scene control modes and constructing a control action library; and a real-time regulation and control execution module, which is used for executing real-time regulation and control actions to output control execution results. The application fuses the improved NOTEARS algorithm and the Gumbel-Softmax variational autoencoder model, and realizes a communication chip interference prediction and adaptive optimization system.

Description

Technical Field

[0001] This invention relates to the field of communication signal optimization technology, and in particular to an artificial intelligence-based communication chip signal interference prediction and optimization system. Background Technology

[0002] In the rapidly developing field of wireless communication systems and on-chip integrated communication modules, the scarcity of communication spectrum resources and the continuous improvement of chip integration have led to increasingly severe multi-source interference problems for communication chips, including electromagnetic interference, burst noise, environmental reflections, and multipath effects. These interferences not only significantly reduce the stability and throughput of communication links but can also have unpredictable system-level performance impacts on the communication control layer, becoming a key bottleneck restricting the stable operation of communication chips and the optimization of communication quality. Therefore, how to achieve efficient prediction and dynamic optimization control of communication interference in complex environments has become a core challenge in the intelligent evolution of current communication systems.

[0003] Existing technologies typically employ statically preset anti-interference parameter tables or power control and spectrum switching strategies based on empirical rules, with feedback-based compensation adjustments made after interference occurs. These methods generally lack the ability to model the evolution of communication behavior, cannot perceive the causal relationships between different modules within the communication system, and cannot generate proactive control strategies for different interference types and propagation mechanisms. Furthermore, most existing methods rely on shallow features or low-dimensional indicators such as signal-to-noise ratio and bit error rate, ignoring the complex control dependencies and data flow relationships between different communication levels. This results in poor generalization ability, delayed response, and coarse control granularity, making them unable to effectively adapt to dynamically changing multi-source interference scenarios. In addition, traditional interference classification methods are often based on supervised learning, requiring a large number of manually labeled interference samples, and are difficult to generalize to unknown types of anomalous interference.

[0004] Therefore, how to provide an artificial intelligence-based communication chip signal interference prediction and optimization system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an artificial intelligence-based signal interference prediction and optimization system for communication chips. This invention integrates causal structure learning and intelligent control technologies to construct an interference prediction and adaptive optimization system for communication chips. It extracts interference causal relationships using an improved NOTEARS algorithm, combines this with a GumbelSoftmax variational autoencoder to identify interference types and risk levels, and generates an optimal control strategy based on multi-objective optimization, achieving dynamic adjustment of communication parameters and adaptive performance improvement.

[0006] An artificial intelligence-based communication chip signal interference prediction and optimization system according to an embodiment of the present invention includes the following modules:

[0007] The data acquisition and preprocessing module is used to acquire multi-dimensional communication status data from the communication chip and preprocess it to generate multi-dimensional communication status data with a unified structure.

[0008] The communication behavior graph construction module is used to construct a communication behavior graph based on multi-dimensional communication state data with a unified structure and the control dependencies, physical connection paths and data flow paths between various functional modules in the communication chip.

[0009] The causal structure learning module is used to extract the state data sequence of nodes in the communication behavior graph based on the improved NOTEARS algorithm, generate a preliminary interference causal graph, and construct a causal potential tensor.

[0010] The interference classification and risk identification module is used to input the causal potential energy tensor into the Gumbel-Softmax variational autoencoder model, generate a set of latent vector representations through the encoder, and output interference type labels and risk level labels by the decoder after discrete sampling.

[0011] The control strategy generation module is used to classify the control modes of the interference scenario based on the output interference type label and risk level label, build a control action library, and output the optimal set of control actions.

[0012] The real-time control execution module is used to input the optimal set of control actions to the communication chip control interface, execute real-time control actions according to the target address and parameter control items, and output the control execution results.

[0013] According to an embodiment of the present invention, a communication chip signal interference prediction and optimization method based on artificial intelligence includes the following steps:

[0014] Step 1: Collect multi-dimensional communication status data of the communication chip under different operating conditions, and preprocess the data to generate multi-dimensional communication status data with a unified structure;

[0015] Step 2: Construct a communication behavior graph based on multidimensional communication state data with a unified structure;

[0016] Step 3: Use the improved NOTEARS algorithm to learn the causal structure of the communication behavior graph, obtain the initial interference causal graph, and construct the causal potential tensor;

[0017] Step 4: Input the causal potential tensor into the Gumbel-Softmax variational autoencoder model, and output the disturbance type label and risk level label corresponding to the node;

[0018] Step 5: Generate the corresponding set of optimal control actions based on the interference type label and risk level label;

[0019] Step 6: Execute real-time control actions based on the optimal set of control actions and the current communication status, and output the control execution results.

[0020] Optionally, the communication status data includes physical layer signal characteristics, link layer performance indicators, and control layer command information; the physical layer signal characteristics include signal-to-noise ratio, spectral power density distribution, instantaneous frequency drift amplitude, and channel occupancy rate; the link layer performance indicators include bit error rate, packet loss rate, average retransmission count, average link delay, and data throughput; the control layer command information includes power control commands, frequency selection configuration, frequency hopping parameters, and time slot scheduling parameters; the preprocessing step includes outlier removal, missing value imputation, and normalization processing for different types of data to generate multidimensional communication status data with a unified structure.

[0021] Optionally, the step of constructing the communication behavior graph specifically includes:

[0022] Multidimensional communication state data in different communication layers are defined as communication behavior graph nodes, which include physical layer nodes, link layer nodes and control layer nodes.

[0023] The physical layer node represents the underlying signal state, the link layer node represents the transmission performance state, and the control layer node represents the command input state.

[0024] Based on the control dependencies, physical connection paths, and data flow paths among the functional modules in the communication chip, directed edges are established in the communication behavior graph.

[0025] The starting point of the directed edge is an upstream module or variable, and the ending point is a controlled module or variable. The types of directed edges include control relationship edges, physical connection edges, and data flow edges.

[0026] Optionally, the step of performing causal structure learning on the communication behavior graph using the improved NOTEARS algorithm specifically includes:

[0027] Extract the state data sequence of each node in the communication behavior map within a set time window to obtain a set of state data sequences. The state data sequence is the communication state data corresponding to each node, which is used to reflect the dynamic changes of the communication state over time.

[0028] Perform multi-scale processing on the set of state data sequences, the multi-scale processing steps being:

[0029] For each state data sequence, a sliding time window of different preset lengths is constructed sequentially;

[0030] Within each sliding time window, the statistical characteristics of each state data sequence are calculated, including the mean, range, magnitude of change, slope of change, standard deviation, and average increment between adjacent time steps.

[0031] The statistical features of all scales are combined and concatenated into a unified multi-scale feature representation, and the multi-scale feature representations of all nodes are summarized to obtain a multi-scale feature representation set;

[0032] Calculate the Pearson correlation coefficient between any two node multiscale feature representations in the multiscale feature representation set, and summarize the Pearson correlation coefficients of all node pairs to construct the initial feature correlation matrix;

[0033] Based on the feature correlation matrix and the temporal order of the nodes, the positive causal relationship between the nodes is determined;

[0034] The positive causal relationships between all nodes are combined with the Pearson correlation coefficient to form an initial causal adjacency matrix. The value of each element in the initial causal adjacency matrix represents the strength of the causal influence between nodes.

[0035] The initial causal adjacency matrix is ​​topologically sorted. If a loop structure is found in the causal path, the loop is broken by removing the edge with the smallest Pearson correlation coefficient in the loop, and the causal adjacency matrix is ​​obtained.

[0036] From the de-looped causal adjacency matrix, edges with Pearson correlation coefficients higher than a set threshold are selected as causal path retention edges, and a preliminary interference causal graph is output. The nodes of the preliminary interference causal graph are the corresponding multi-scale feature representations, the edges are the corresponding causal path retention edges, and the edge weights are the corresponding Pearson correlation coefficients.

[0037] The fluctuation amplitude, variance and number of anomalous jumps of the multi-scale feature representation corresponding to each node in the preliminary disturbance causal graph are calculated within a preset time window, and the initial potential value of the corresponding node is obtained by weighting according to the preset weights.

[0038] For each directed causal edge in the initial interference causal graph, from the source node to the target node, read the Pearson correlation coefficient of the corresponding edge and sum it with the initial potential value of the source node, and assign the interference transmission potential value to the corresponding edge path.

[0039] Based on the topological structure of the interference causal graph, a directed traversal is performed starting from a node with no incoming edges. The directed traversal step is to propagate potential energy along the topological order of the interference causal graph until all paths are scanned, forming the cumulative interference transmission potential energy value of each node.

[0040] Write the initial potential energy value of each node, the disturbance propagation potential energy value of the incoming edge, and the cumulative disturbance propagation potential energy value into the corresponding dimension of the tensor to obtain the causal potential energy tensor.

[0041] Optionally, step four specifically includes:

[0042] The causal potential tensor is extracted by the encoder of the Gumbel-Softmax variational autoencoder model to extract the feature structure in the causal potential tensor of each node and generate the corresponding latent vector representation set. The latent vector representation set is used to abstractly represent the high-dimensional pattern features of interference in the propagation process.

[0043] Discrete sampling is performed on the latent vector representation set based on the Gumbel-Softmax reparameterization mechanism to generate discrete latent variables with class discrimination ability;

[0044] Discrete latent variables are input into the decoder of the Gumbel-Softmax variational autoencoder model to reconstruct and map the discrete latent variables, and output the interference type label and risk level label corresponding to the node. The interference type label is used to represent the interference category corresponding to each node, including sudden interference, periodic interference, structural interference and unknown anomaly interference.

[0045] The risk level label is used to indicate the potential severity of the interference during its propagation, and is divided into high risk, medium risk and low risk levels.

[0046] Optionally, step five specifically includes:

[0047] Based on the output interference type label and risk level label, the interference scenario is divided into multiple control modes, including sudden interference emergency mode, periodic interference suppression mode, structural interference stabilization mode and unknown interference adaptive mode.

[0048] For each control mode, a corresponding control action library is established. The control action library consists of a chip-level adjustable parameter mapping table, including a modulation and coding scheme parameter table, a transmit power control table, a spectrum switching configuration table, a time slot scheduling priority table, and a link resource reallocation table.

[0049] Read the interference type label and risk level label of the interfered nodes in the current communication behavior graph, assign control priority to the nodes according to the severity of the risk level label, and execute resource adjustment strategy according to the control priority;

[0050] Retrieve a set of candidate control actions that match the current disturbance type from the control action library, and generate a candidate sequence of control actions;

[0051] The candidate sequence of control actions is subjected to multi-objective screening, which selects the control action with the lowest energy consumption change rate and channel occupancy balance to form an optimal set of control actions. The optimal set of control actions includes modulation and coding switching, dynamic adjustment of transmit power, spectrum jump switching, time slot scheduling adjustment, and link reconfiguration.

[0052] Optionally, step six specifically includes:

[0053] The optimal set of control actions is input to the control interface of the communication chip and matched with the latest state data of the corresponding node in the current communication behavior graph to extract the target address and parameter control items for regulation.

[0054] Based on the target address, control commands are sent through the communication chip bus to perform parameter adjustment operations on the communication control unit, modulation unit, power amplification unit, spectrum allocation unit, time slot scheduling unit and link management unit inside the communication chip;

[0055] During parameter adjustment based on parameter control items, real-time monitoring of communication status feedback data is conducted. The communication status feedback data includes updated physical layer signal strength, link layer performance indicators, and control layer response latency.

[0056] By comparing the communication status data and communication status feedback data before and after parameter adjustment, the interference suppression effect index is calculated. The interference suppression effect index includes the signal quality improvement, the reduction ratio of bit error rate, the stability of data path and the trend of energy consumption change.

[0057] All parameter adjustment records and communication status feedback data during the control execution process are summarized into the control execution result.

[0058] The beneficial effects of this invention are:

[0059] This invention constructs a communication behavior map integrating communication state data and functional structure, and introduces an improved NOTEARS algorithm for causal structure learning. This successfully uncovers the interference propagation paths and causal influence strengths between different functional modules within a communication chip, establishing a causal potential energy tensor for dynamic evolution modeling, significantly improving the perception and interpretation capabilities of complex interference patterns. By employing a Gumbel-Softmax variational autoencoder model for high-dimensional latent variable modeling and unsupervised discrete classification of the causal potential energy tensor, it not only achieves accurate identification of interference types and risk levels but also possesses adaptability to unknown interference scenarios. Based on this, the system automatically matches control modes according to interference labels, constructs a targeted control action library, and selects the optimal control action set through multi-objective optimization methods such as energy consumption and channel occupancy balancing. Ultimately, it achieves real-time intelligent control of chip-level modulation, power, spectrum, time slots, and link parameters. This invention excels in interference suppression, communication quality improvement, resource scheduling efficiency, and system robustness, effectively addressing the challenges posed by multi-source heterogeneous interference to communication chip performance, and possesses significant engineering practical value and promising prospects for widespread application. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a schematic diagram of the structure of a communication chip signal interference prediction and optimization system based on artificial intelligence proposed in this invention;

[0062] Figure 2 This is an overall flowchart of an artificial intelligence-based method for predicting and optimizing signal interference in communication chips, as proposed in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figure 1 A communication chip signal interference prediction and optimization system based on artificial intelligence includes the following modules:

[0065] The data acquisition and preprocessing module is used to acquire multi-dimensional communication status data from the communication chip and preprocess it to generate multi-dimensional communication status data with a unified structure.

[0066] The communication behavior graph construction module is used to construct a communication behavior graph based on multi-dimensional communication state data with a unified structure and the control dependencies, physical connection paths and data flow paths between various functional modules in the communication chip.

[0067] The causal structure learning module is used to extract the state data sequence of nodes in the communication behavior graph based on the improved NOTEARS algorithm, generate a preliminary interference causal graph, and construct a causal potential tensor.

[0068] The interference classification and risk identification module is used to input the causal potential energy tensor into the Gumbel-Softmax variational autoencoder model, generate a set of latent vector representations through the encoder, and output interference type labels and risk level labels by the decoder after discrete sampling.

[0069] The control strategy generation module is used to classify the control modes of the interference scenario based on the output interference type label and risk level label, build a control action library, and output the optimal set of control actions.

[0070] The real-time control execution module is used to input the optimal set of control actions to the communication chip control interface, execute real-time control actions according to the target address and parameter control items, and output the control execution results.

[0071] This step integrates multi-dimensional communication data modeling, causal structure learning, and intelligent control strategy generation to achieve accurate identification and dynamic optimization of signal interference within the communication chip. The system can not only extract potential interference sources from complex communication behaviors but also adaptively generate optimal control actions based on risk levels and execute them in real time, significantly improving the chip's anti-interference capability and communication performance stability.

[0072] refer to Figure 2 A method for predicting and optimizing signal interference in communication chips based on artificial intelligence includes the following steps:

[0073] Step 1: Collect multi-dimensional communication status data of the communication chip under different operating conditions, and preprocess the data to generate multi-dimensional communication status data with a unified structure;

[0074] Step 2: Construct a communication behavior graph based on multidimensional communication state data with a unified structure;

[0075] Step 3: Use the improved NOTEARS algorithm to learn the causal structure of the communication behavior graph, obtain the initial interference causal graph, and construct the causal potential tensor;

[0076] Step 4: Input the causal potential tensor into the Gumbel-Softmax variational autoencoder model, and output the disturbance type label and risk level label corresponding to the node;

[0077] Step 5: Generate the corresponding set of optimal control actions based on the interference type label and risk level label;

[0078] Step 6: Execute real-time control actions based on the optimal set of control actions and the current communication status, and output the control execution results.

[0079] This step achieves intelligent processing of signal interference within the communication chip through six consecutive steps, from data acquisition, spectrum construction, causal learning to interference identification and control response, forming a closed-loop feedback interference prediction and optimization mechanism, which effectively enhances the communication system's ability to perceive, classify and adaptively control complex interference.

[0080] In this embodiment, the communication status data includes physical layer signal characteristics, link layer performance indicators, and control layer command information; the physical layer signal characteristics include signal-to-noise ratio, spectral power density distribution, instantaneous frequency drift amplitude, and channel occupancy rate; the link layer performance indicators include bit error rate, packet loss rate, average retransmission count, average link delay, and data throughput; the control layer command information includes power control commands, frequency selection configuration, frequency hopping parameters, and time slot scheduling parameters; the preprocessing step includes outlier removal, missing value imputation, and normalization processing for different types of data to generate multidimensional communication status data with a unified structure.

[0081] This step, by clarifying the hierarchical structure and specific indicators of communication status data, covering multi-dimensional features of the physical layer, link layer, and control layer, achieves a comprehensive and detailed characterization of the communication chip's operating status. At the same time, the introduction of preprocessing procedures such as outlier removal, missing value imputation, and normalization ensures the stability and consistency of the input data, providing a high-quality data foundation for subsequent communication behavior graph construction and causal analysis, and effectively improving the accuracy and robustness of interference identification and optimization strategy generation.

[0082] In this embodiment, the step of constructing the communication behavior graph is specifically as follows:

[0083] Multidimensional communication state data in different communication layers are defined as communication behavior graph nodes, which include physical layer nodes, link layer nodes and control layer nodes.

[0084] The physical layer node represents the underlying signal state, the link layer node represents the transmission performance state, and the control layer node represents the command input state.

[0085] Based on the control dependencies, physical connection paths, and data flow paths among the functional modules in the communication chip, directed edges are established in the communication behavior graph.

[0086] The starting point of the directed edge is an upstream module or variable, and the ending point is a controlled module or variable. The types of directed edges include control relationship edges, physical connection edges, and data flow edges.

[0087] This step constructs a communication behavior graph that integrates multi-dimensional state information from the physical layer, link layer, and control layer. This results in a heterogeneous graph structure where nodes represent different communication layer states and directed edges depict the control, connection, and data transmission relationships between modules. This effectively reveals the dynamic dependencies between multiple modules within the communication chip. This graph structure not only enhances the structured representation of the chip's internal operating state but also provides a precise graph input foundation for subsequent causal structure learning, improving the accuracy and interpretability of interference tracing and risk prediction.

[0088] In this embodiment, the step of performing causal structure learning on the communication behavior graph using the improved NOTEARS algorithm specifically includes:

[0089] Extract the state data sequence of each node in the communication behavior map within a set time window to obtain a set of state data sequences. The state data sequence is the communication state data corresponding to each node, which is used to reflect the dynamic changes of the communication state over time.

[0090] Perform multi-scale processing on the set of state data sequences, the multi-scale processing steps being:

[0091] For each state data sequence, a sliding time window of different preset lengths is constructed sequentially;

[0092] Within each sliding time window, the statistical characteristics of each state data sequence are calculated, including the mean, range, magnitude of change, slope of change, standard deviation, and average increment between adjacent time steps.

[0093] The statistical features of all scales are combined and concatenated into a unified multi-scale feature representation, and the multi-scale feature representations of all nodes are summarized to obtain a multi-scale feature representation set;

[0094] Calculate the Pearson correlation coefficient between any two node multiscale feature representations in the multiscale feature representation set, and summarize the Pearson correlation coefficients of all node pairs to construct the initial feature correlation matrix;

[0095] Based on the feature correlation matrix and the temporal order of the nodes, the positive causal relationship between the nodes is determined;

[0096] The positive causal relationships between all nodes are combined with the Pearson correlation coefficient to form an initial causal adjacency matrix. The value of each element in the initial causal adjacency matrix represents the strength of the causal influence between nodes.

[0097] The initial causal adjacency matrix is ​​topologically sorted. If a loop structure is found in the causal path, the loop is broken by removing the edge with the smallest Pearson correlation coefficient in the loop, and the causal adjacency matrix is ​​obtained.

[0098] From the de-looped causal adjacency matrix, edges with Pearson correlation coefficients higher than a set threshold are selected as causal path retention edges, and a preliminary interference causal graph is output. The nodes of the preliminary interference causal graph are the corresponding multi-scale feature representations, the edges are the corresponding causal path retention edges, and the edge weights are the corresponding Pearson correlation coefficients.

[0099] The fluctuation amplitude, variance and number of anomalous jumps of the multi-scale feature representation corresponding to each node in the preliminary disturbance causal graph are calculated within a preset time window, and the initial potential value of the corresponding node is obtained by weighting according to the preset weights.

[0100] For each directed causal edge in the initial interference causal graph, from the source node to the target node, read the Pearson correlation coefficient of the corresponding edge and sum it with the initial potential value of the source node, and assign the interference transmission potential value to the corresponding edge path.

[0101] Based on the topological structure of the interference causal graph, a directed traversal is performed starting from a node with no incoming edges. The directed traversal step is to propagate potential energy along the topological order of the interference causal graph until all paths are scanned, forming the cumulative interference transmission potential energy value of each node.

[0102] Write the initial potential energy value of each node, the disturbance propagation potential energy value of the incoming edge, and the cumulative disturbance propagation potential energy value into the corresponding dimension of the tensor to obtain the causal potential energy tensor.

[0103] This step utilizes an improved NOTEARS algorithm, combined with multi-scale statistical feature extraction and correlation analysis, to efficiently construct a preliminary interference causal graph and causal potential tensor from the communication behavior map. Compared to traditional structurally differentiable optimization methods, the proposed sliding window-based multi-scale processing method and Pearson correlation coefficient construction strategy improve the controllability and computational efficiency of causal relationship learning. Simultaneously, the interference potential modeling method, which integrates node feature volatility and causal propagation paths, enhances the semantic expression of the causal structure and the ability to measure the intensity of interference impact, providing a solid foundation for subsequent interference type identification and risk level assessment.

[0104] In this embodiment, step four specifically includes:

[0105] The causal potential tensor is extracted by the encoder of the Gumbel-Softmax variational autoencoder model to extract the feature structure in the causal potential tensor of each node and generate the corresponding latent vector representation set. The latent vector representation set is used to abstractly represent the high-dimensional pattern features of interference in the propagation process.

[0106] Discrete sampling is performed on the latent vector representation set based on the Gumbel-Softmax reparameterization mechanism to generate discrete latent variables with class discrimination ability;

[0107] Discrete latent variables are input into the decoder of the Gumbel-Softmax variational autoencoder model to reconstruct and map the discrete latent variables, and output the interference type label and risk level label corresponding to the node. The interference type label is used to represent the interference category corresponding to each node, including sudden interference, periodic interference, structural interference and unknown anomaly interference.

[0108] The risk level label is used to indicate the potential severity of the interference during its propagation, and is divided into high risk, medium risk and low risk levels.

[0109] This step introduces a Gumbel-Softmax variational autoencoder to perform high-dimensional latent pattern recognition and discretization classification of the causal potential tensor, achieving efficient identification of communication interference types and risk levels under unsupervised learning conditions. Compared to traditional clustering or classification methods, this step utilizes a reparameterization mechanism to map the continuous latent space into discriminative discrete variables, thereby enhancing the model's ability to identify sudden, periodic, structural, and unknown anomaly-type interference. Simultaneously, by combining the propagation potential energy of nodes in the causal graph to dynamically estimate the interference impact intensity, more accurate risk level quantification is achieved. This significantly improves the generalization ability and response accuracy in complex communication interference scenarios, providing a solid decision-making basis for subsequent control strategy formulation.

[0110] In this embodiment, step five specifically includes:

[0111] Based on the output interference type label and risk level label, the interference scenario is divided into multiple control modes, including sudden interference emergency mode, periodic interference suppression mode, structural interference stabilization mode and unknown interference adaptive mode.

[0112] For each control mode, a corresponding control action library is established. The control action library consists of a chip-level adjustable parameter mapping table, including a modulation and coding scheme parameter table, a transmit power control table, a spectrum switching configuration table, a time slot scheduling priority table, and a link resource reallocation table.

[0113] Read the interference type label and risk level label of the interfered nodes in the current communication behavior graph, assign control priority to the nodes according to the severity of the risk level label, and execute resource adjustment strategy according to the control priority;

[0114] Retrieve a set of candidate control actions that match the current disturbance type from the control action library, and generate a candidate sequence of control actions;

[0115] The candidate sequence of control actions is subjected to multi-objective screening, which selects the control action with the lowest energy consumption change rate and channel occupancy balance to form an optimal set of control actions. The optimal set of control actions includes modulation and coding switching, dynamic adjustment of transmit power, spectrum jump switching, time slot scheduling adjustment, and link reconfiguration.

[0116] This step achieves differentiated dynamic control strategies for different interference types and risk levels by constructing a multi-mode control mechanism that matches interference tags. The system divides control modes based on the tag attributes of interfered nodes in the communication behavior graph and allocates resources according to control priorities. Simultaneously, it utilizes chip-level parameter mapping to establish a multi-dimensional control action library and performs multi-objective optimization screening to generate the set of control actions with optimal energy efficiency and stability. Compared to traditional static control strategies, this solution has stronger adaptability and real-time performance, enabling intelligent steady-state optimization of communication performance and balanced energy consumption control in complex, dynamic interference environments.

[0117] In this embodiment, step six specifically includes:

[0118] The optimal set of control actions is input to the control interface of the communication chip and matched with the latest state data of the corresponding node in the current communication behavior graph to extract the target address and parameter control items for regulation.

[0119] Based on the target address, control commands are sent through the communication chip bus to perform parameter adjustment operations on the communication control unit, modulation unit, power amplification unit, spectrum allocation unit, time slot scheduling unit and link management unit inside the communication chip;

[0120] During parameter adjustment based on parameter control items, real-time monitoring of communication status feedback data is conducted. The communication status feedback data includes updated physical layer signal strength, link layer performance indicators, and control layer response latency.

[0121] By comparing the communication status data and communication status feedback data before and after parameter adjustment, the interference suppression effect index is calculated. The interference suppression effect index includes the signal quality improvement, the reduction ratio of bit error rate, the stability of data path and the trend of energy consumption change.

[0122] All parameter adjustment records and communication status feedback data during the control execution process are summarized into the control execution result.

[0123] This step constructs a closed-loop communication control mechanism through the execution and feedback analysis of the optimal set of control actions. The system can accurately locate the target address and control parameters of the interfered node, and control multiple key functional units within the communication chip at the instruction level, achieving synchronous optimization of multi-dimensional communication parameters. During the control process, the system acquires and analyzes feedback data from the physical layer, link layer, and control layer in real time, dynamically evaluating the interference suppression effect. Quantitative feedback is generated through indicators such as bit error rate, signal strength, and energy consumption trends, thereby driving continuous iterative updates to the control strategy. This mechanism, which integrates state awareness and execution feedback, significantly improves the communication system's response speed and control effectiveness in complex interference environments.

[0124] Example 1:

[0125] To verify the feasibility of this invention in practice, it was applied to a high-performance communication chip platform for system deployment and interference suppression experiments. This communication chip is used in a high-speed railway vehicle-to-ground communication system, operating in the 5.8GHz high-frequency band. The communication environment is characterized by high-speed movement, frequent frequency hopping, and complex electromagnetic interference sources, facing various structural, periodic, and sudden interference events, which greatly affect communication reliability and system stability.

[0126] In our experimental scenarios, we selected several typical areas with high concentrations of interference events, including tunnel entrances and exits, sections with numerous turnouts, and high-voltage power grid crossings. These areas are often accompanied by sudden lightning pulses, interference reflected from trackside monitoring equipment, cross signals caused by multipath effects, and periodic electromagnetic interference caused by track resonance. Traditional communication chips in such environments primarily employ power control or frequency avoidance based on empirical rules, lacking the ability to accurately identify and adaptively optimize the causes of multi-source interference.

[0127] The system constructed in this invention first uses a communication behavior graph to uniformly model the multi-dimensional state data of the physical layer, link layer, and control layer within the chip, abstracting the control dependencies, physical connections, and data flow paths between functional units into a directed graph structure. Based on this, an improved NOTEARS algorithm is used to learn causal relationships from the node state data, further generating a causal potential tensor to model interference propagation paths and potential energy accumulation regions. The system uses a Gumbel-Softmax variational autoencoder model to embed, discriminate, and reconstruct interference patterns in a high dimension, outputting interference type labels (such as burst interference, periodic interference, structural interference, and unknown interference) and risk level labels (high, medium, and low), thereby achieving prior identification of potential interference.

[0128] To address different types and levels of interference, the system retrieves a multi-dimensional adjustable parameter configuration table (including modulation scheme, transmit power, spectrum switching, time slot scheduling, and link allocation) from the underlying control action library. This table is then matched with the current chip state and behavior graph nodes to form the optimal control action sequence. Through energy consumption and performance multi-objective constraints, the system improves communication stability while suppressing unnecessary power consumption, ensuring overall link efficiency.

[0129] In the comparative experiment, we compared our method with three mainstream interference handling methods. These methods represent the main technical paths for anti-interference in current communication chips, but they generally suffer from problems such as low recognition rate, high response delay, and lagging control strategies in scenarios with multiple overlapping interferences and real-time changes.

[0130] To comprehensively evaluate the advantages of the method of this invention, we collected data from 1623 real interference events over a continuous 48-hour period in an experimental environment, and quantified various indicators, including interference identification accuracy, response delay, signal quality improvement (represented by SNR), bit error rate reduction ratio, link stability (retransmission rate), and energy consumption trend. The following is a detailed summary of the statistical results.

[0131] Table 1 Comparative test results of different interference optimization techniques

[0132]

[0133] As shown in Table 1, the method of this invention significantly outperforms the other three traditional methods in terms of interference identification accuracy, achieving a recognition level of 92.4%, especially maintaining good recognition performance in complex scenarios with multiple types of interference superimposed. In terms of response speed, the system utilizes a causal propagation chain constructed using graph structures and tensor representations to achieve rapid identification and label output. Compared to the burst frequency jump avoidance strategy based on spectrum sensing and the multi-level judgment mechanism of the graph traversal heuristic offline prediction model, it has a stronger real-time advantage.

[0134] During control execution, this invention achieves a signal quality improvement of over 11 dB and an average reduction in bit error rate of 86.3% by employing a multi-objective screening strategy while maintaining low energy consumption growth. Especially in scenarios with frequent structural interference, spectrum-aware burst frequency jump avoidance strategies and graph traversal heuristic offline prediction models often fail to identify abnormal nodes in complex causal chains. This method, however, can mine deep interference propagation paths through potential energy transfer mechanisms, enabling link reconstruction and resource reallocation strategies, thereby effectively improving link stability. This system not only possesses high-precision classification and level judgment capabilities for interference types but also outperforms traditional methods in multiple optimization dimensions, demonstrating excellent engineering deployability and real-time interference management capabilities.

[0135] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An artificial intelligence based communication chip signal interference prediction and optimization system, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to acquire multi-dimensional communication status data from the communication chip and preprocess it to generate multi-dimensional communication status data with a unified structure. The communication behavior graph construction module is used to construct a communication behavior graph based on multi-dimensional communication state data with a unified structure and the control dependencies, physical connection paths and data flow paths between various functional modules in the communication chip. The causal structure learning module is used to extract the state data sequences of nodes in the communication behavior graph based on the improved NOTEARS algorithm, generate a preliminary interference causal graph, and construct a causal potential tensor. Specifically: Extract the state data sequence of each node in the communication behavior map within a set time window to obtain a set of state data sequences. The state data sequence is the communication state data corresponding to each node, which is used to reflect the dynamic changes of the communication state over time. Perform multi-scale processing on the set of state data sequences, the multi-scale processing steps being: For each state data sequence, a sliding time window of different preset lengths is constructed sequentially; Within each sliding time window, the statistical characteristics of each state data sequence are calculated, including the mean, range, magnitude of change, slope of change, standard deviation, and average increment between adjacent time steps. The statistical features of all scales are combined and concatenated into a unified multi-scale feature representation, and the multi-scale feature representations of all nodes are summarized to obtain a multi-scale feature representation set; Calculate the Pearson correlation coefficient between any two node multiscale feature representations in the multiscale feature representation set, and summarize the Pearson correlation coefficients of all node pairs to construct the initial feature correlation matrix; Based on the feature correlation matrix and the temporal order of the nodes, the positive causal relationship between the nodes is determined; The positive causal relationships between all nodes are combined with the Pearson correlation coefficient to form an initial causal adjacency matrix. The value of each element in the initial causal adjacency matrix represents the strength of the causal influence between nodes. The initial causal adjacency matrix is ​​topologically sorted. If a loop structure is found in the causal path, the loop is broken by removing the edge with the smallest Pearson correlation coefficient in the loop, and the loop-free causal adjacency matrix is ​​obtained. From the de-looped causal adjacency matrix, edges with Pearson correlation coefficients higher than a set threshold are selected as causal path retention edges, and a preliminary interference causal graph is output. The nodes of the preliminary interference causal graph are the corresponding multi-scale feature representations, the edges are the corresponding causal path retention edges, and the edge weights are the corresponding Pearson correlation coefficients. The fluctuation amplitude, variance and number of anomalous jumps of the multi-scale feature representation corresponding to each node in the preliminary disturbance causal graph are calculated within a preset time window, and the initial potential value of the corresponding node is obtained by weighting according to the preset weights. For each directed causal edge in the initial interference causal graph, from the source node to the target node, read the Pearson correlation coefficient of the corresponding edge and sum it with the initial potential value of the source node, and assign it as the interference transfer potential value of the corresponding edge. Based on the topological structure of the interference causal graph, a directed traversal is performed starting from a node with no incoming edges. The directed traversal step is to propagate potential energy along the topological order of the interference causal graph until all paths are scanned, forming the cumulative interference transmission potential energy value of each node. Write the initial potential energy value of each node, the disturbance propagation potential energy value of the incoming edge, and the cumulative disturbance propagation potential energy value into the corresponding dimension of the tensor to obtain the causal potential energy tensor. The interference classification and risk identification module is used to input the causal potential energy tensor into the Gumbel-Softmax variational autoencoder model, generate a set of latent vector representations through the encoder, and output interference type labels and risk level labels by the decoder after discrete sampling. The control strategy generation module is used to classify the control modes of the interference scenario based on the output interference type label and risk level label, build a control action library, and output the optimal set of control actions. The real-time control execution module is used to input the optimal set of control actions to the communication chip control interface, execute real-time control actions according to the target address and parameter control items, and output the control execution results.

2. The system according to claim 1, wherein, The modules are connected in the following way: Step 1: Collect multi-dimensional communication status data of the communication chip under different operating conditions, and preprocess the data to generate multi-dimensional communication status data with a unified structure; Step 2: Construct a communication behavior graph based on multidimensional communication state data with a unified structure; Step 3: Use the improved NOTEARS algorithm to learn the causal structure of the communication behavior graph, obtain the initial interference causal graph, and construct the causal potential tensor; Step 4: Input the causal potential tensor into the Gumbel-Softmax variational autoencoder model, and output the disturbance type label and risk level label corresponding to the node; Step 5: Generate the corresponding set of optimal control actions based on the interference type label and risk level label; Step 6: Execute real-time control actions based on the optimal set of control actions and the current communication status, and output the control execution results.

3. The system of claim 2, wherein the system is configured to: The communication status data includes physical layer signal characteristics, link layer performance indicators, and control layer command information. The physical layer signal characteristics include signal-to-noise ratio, spectral power density distribution, instantaneous frequency drift amplitude, and channel occupancy rate. The link layer performance indicators include bit error rate, packet loss rate, average retransmission count, average link delay, and data throughput. The control layer command information includes power control commands, frequency selection configuration, frequency hopping parameters, and time slot scheduling parameters. The preprocessing steps include outlier removal, missing value imputation, and normalization of different types of data to generate multidimensional communication status data with a unified structure.

4. The system according to claim 2, wherein, The construction of the communication behavior graph specifically involves: Multidimensional communication state data in different communication layers are defined as communication behavior graph nodes, which include physical layer nodes, link layer nodes and control layer nodes. The physical layer node represents the underlying signal state, the link layer node represents the transmission performance state, and the control layer node represents the command input state. Based on the control dependencies, physical connection paths, and data flow paths among the functional modules in the communication chip, directed edges are established in the communication behavior graph. The starting point of the directed edge is an upstream module or variable, and the ending point is a controlled module or variable. The types of directed edges include control relationship edges, physical connection edges, and data flow edges.

5. The system of claim 2, wherein the system is configured to: Step four specifically involves: The causal potential tensor is extracted by the encoder of the Gumbel-Softmax variational autoencoder model to extract the feature structure in the causal potential tensor of each node and generate the corresponding latent vector representation set. The latent vector representation set is used to abstractly represent the high-dimensional pattern features of interference in the propagation process. Discrete sampling is performed on the latent vector representation set based on the Gumbel-Softmax reparameterization mechanism to generate discrete latent variables with class discrimination ability; Discrete latent variables are input into the decoder of the Gumbel-Softmax variational autoencoder model to reconstruct and map the discrete latent variables, and output the interference type label and risk level label corresponding to the node. The interference type label is used to represent the interference category corresponding to each node, including sudden interference, periodic interference, structural interference and unknown anomaly interference. The risk level label is used to indicate the potential severity of the interference during its propagation, and is divided into high risk, medium risk and low risk levels.

6. The system of claim 2, wherein the system is configured to: Step five specifically involves: Based on the output interference type label and risk level label, the interference scenario is divided into multiple control modes, including sudden interference emergency mode, periodic interference suppression mode, structural interference stabilization mode and unknown interference adaptive mode. For each control mode, a corresponding control action library is established. The control action library consists of a chip-level adjustable parameter mapping table, including a modulation and coding scheme parameter table, a transmit power control table, a spectrum switching configuration table, a time slot scheduling priority table, and a link resource reallocation table. Read the interference type label and risk level label of the interfered nodes in the current communication behavior graph, assign control priority to the nodes according to the severity of the risk level label, and execute resource adjustment strategy according to the control priority; Retrieve a set of candidate control actions that match the current disturbance type from the control action library, and generate a candidate sequence of control actions; The candidate sequence of control actions is subjected to multi-objective screening, which selects the control action with the lowest energy consumption change rate and channel occupancy balance to form an optimal set of control actions. The optimal set of control actions includes modulation and coding switching, dynamic adjustment of transmit power, spectrum jump switching, time slot scheduling adjustment, and link reconfiguration.

7. The system according to claim 2, wherein the system is further configured to: Step six specifically involves: The optimal set of control actions is input to the control interface of the communication chip and matched with the latest state data of the corresponding node in the current communication behavior graph to extract the target address and parameter control items for regulation. Based on the target address, control commands are sent through the communication chip bus to perform parameter adjustment operations on the communication control unit, modulation unit, power amplification unit, spectrum allocation unit, time slot scheduling unit and link management unit inside the communication chip; During parameter adjustment based on parameter control items, real-time monitoring of communication status feedback data is conducted. The communication status feedback data includes updated physical layer signal strength, link layer performance indicators, and control layer response latency. By comparing the communication status data and communication status feedback data before and after parameter adjustment, the interference suppression effect index is calculated. The interference suppression effect index includes the signal quality improvement, the reduction ratio of bit error rate, the stability of data path and the trend of energy consumption change. All parameter adjustment records and communication status feedback data during the control execution process are summarized into the control execution result.

Images