A multi-modal data source probe state monitoring and failover method and system

Abstract

The application discloses a kind of multi-modal data signal source probe state monitoring and fault transfer method and system, it is related to the field of Internet, this method includes: respectively deploying data acquisition probe at wireless access network node and core network UPF node, and with obtaining after pre-processing wireless access network node network data and pre-processing core network UPF network data;By AI monitoring model, obtain abnormal risk prediction information;According to abnormal risk prediction information, adjust the parameter configuration of wireless side and core network UPF side;Based on the parameter configuration of wireless side and core network UPF side after adjustment is fed back to service perception module, to realize fault transfer.The application has realized cross-domain, cross-layer unified probe monitoring system, also through graph neural network and reinforcement learning and other AI technology constructs dynamic, interpretable, closed-loop network fault analysis and control mechanism, can significantly improve the precision of network state perception, the timeliness of response and the intelligent level of tuning.

Description

TECHNICAL FIELD

[0001] The present specification relates to the field of Internet, more specifically, the present invention relates to a multi-modal data source probe state monitoring and fault transfer method and system. BACKGROUND

[0002] With the development of 5G and future network architecture, the network structure is increasingly complex, and after the decoupling of the data plane and the control plane, real-time perception and accurate scheduling of network state become particularly important. Especially between the radio access network (RAN) and the core network user plane function (UPF), the service path is long, the user scale is large, and the data flow is complex. Once a network failure occurs, it is easy to cause large-scale service interruption or performance degradation.

[0003] Existing network operation mainly relies on manual rule setting or traditional alarm systems to complete state monitoring and fault handling. These methods have the following shortcomings:

[0004] 1. Probe deployment is fragmented and information is isolated: data collection on the wireless side and the UPF side of the core network is often managed separately, and there is a lack of end-to-end cross-domain state integration capability;

[0005] 2. Network state monitoring is lagging: traditional systems rely on threshold alarms or static indicators for judgment, making it difficult to detect dynamic burst risks in a timely manner;

[0006] 3. Lack of intelligent optimization mechanism: even if an anomaly is detected, existing systems often rely on fixed strategies for execution and cannot adaptively optimize based on actual network evolution trends;

[0007] 4. Control response lacks explainability: network parameter adjustment lacks identification and analysis of the source of the anomaly, which can easily lead to excessive or ineffective configuration, thereby affecting network stability.

[0008] In the modern network environment of high-density access and complex service types, there is an urgent need for a fault detection and transfer method based on multi-source probe fusion, with intelligent analysis capabilities, and capable of closed-loop optimization, to support a more flexible, explainable, and efficient network operation guarantee system. SUMMARY

[0009] A series of simplified concepts are introduced in the summary section, which will be further described in detail in the specific embodiments section. The summary section of the present invention does not mean to attempt to limit the key features and necessary technical features of the claimed technical solutions, nor does it mean to attempt to determine the protection scope of the claimed technical solutions.

[0010] In a first aspect, the present invention proposes a multi-modal data source probe state monitoring and fault transfer method, the method comprising:

[0011] deploying data collection probes at the radio access network node and the core network UPF node respectively to collect radio access network node network data and core network UPF network data;

[0012] performing preprocessing operations on the radio access network node network data and the core network UPF network data to obtain preprocessed radio access network node network data and preprocessed core network UPF network data;

[0013] inputting the preprocessed radio access network node network data and the preprocessed core network UPF network data into an AI monitoring model to obtain abnormal risk prediction information, wherein the AI monitoring model includes a multi-source feature encoding module, a spatio-temporal collaborative modeling module, and a risk prediction module, the multi-source feature encoding module is used to map the preprocessed radio access network node network data and the preprocessed core network UPF network data into a unified feature space respectively, the spatio-temporal collaborative modeling module is used to construct a dynamic graph and extract spatio-temporal dependencies of the dynamic graph based on a graph neural network; the risk prediction module is used to model a historical state sequence of a node in the dynamic graph to output the abnormal risk prediction information;

[0014] adjusting parameter configurations of the radio side and the core network UPF side according to the abnormal risk prediction information;

[0015] feedback the adjusted parameter configurations of the radio side and the core network UPF side to a service awareness module to realize fault transfer.

[0016] In a feasible implementation, the network data includes network traffic, latency, packet loss, jitter, application connectivity, and application network quality.

[0017] In a feasible implementation, the risk prediction module adopts a multi-layer Transformer structure.

[0018] In a feasible implementation, the adjusting parameter configurations of the radio side and the core network UPF side according to the abnormal risk prediction information includes:

[0019] constructing a policy knowledge graph, wherein the policy knowledge graph is used to store the association between the abnormal type and the network parameter item and the mapping relationship between the network parameter item and the control action;

[0020] based on the abnormal risk prediction information, retrieving a candidate parameter adjustment strategy set matching the current abnormal type from the policy knowledge graph;

[0021] obtaining current network state information, historical parameter adjustment records, and network performance feedback information to construct a state representation for reinforcement learning;

[0022] based on the state representation, performing an optimal policy selection operation based on a deep reinforcement learning algorithm in the candidate parameter adjustment policy set to determine a target control action for adjusting the wireless side and / or the core network UPF side;

[0023] performing a corresponding parameter adjustment operation according to the target control action to achieve optimization of the network performance indicators, wherein the target control action aims to maximize network performance improvement, which is evaluated by a reward function calculated based on risk score change, network latency change, packet loss rate change and QoS feedback information.

[0024] In a possible implementation, the performing of the optimal policy selection operation based on the deep reinforcement learning algorithm in the candidate parameter adjustment policy set to determine the target control action for adjusting the wireless side and / or the core network UPF side based on the state representation comprises:

[0025] inputting the state representation into a policy selection model constructed based on deep reinforcement learning, wherein the policy selection model is used for action evaluation and policy selection in the candidate parameter adjustment policy set, and the policy selection model is trained by a deep reinforcement learning algorithm to minimize a policy loss function and maximize a reward function value constructed based on network performance feedback, the reward function being evaluated based on risk score change, network latency change, packet loss rate change and QoS indicator change;

[0026] calculating a policy score of each candidate action based on the current state representation in the candidate parameter adjustment policy set, and selecting a target control action with the highest expected long-term return, wherein the target control action is used as a network control instruction to adjust the parameter configuration of the corresponding wireless side and / or core network UPF side.

[0027] In a possible implementation, the target control action comprises one or more of a dynamic adjustment operation of cell scheduling priority performed at the wireless side, a transmit power control operation performed at the wireless side, a handover threshold optimization operation performed at the wireless side, a user plane traffic path re-routing operation performed at the core network UPF side, a QoS identifier switching operation performed at the core network UPF side and a forwarding table management operation performed at the core network UPF side.

[0028] In a second aspect, the present application further provides a multi-modal data source probe state monitoring and fault transfer system for the method of any one of the first aspect, comprising:

[0029] The collection unit is configured to deploy data collection probes at a radio access network node and a core network UPF node respectively to collect radio access network node network data and core network UPF network data, wherein the network data includes network traffic, latency, packet loss, jitter, application connectivity and application network quality.

[0030] The preprocessing unit is configured to perform a preprocessing operation on the radio access network node network data and the core network UPF network data to obtain preprocessed radio access network node network data and preprocessed core network UPF network data.

[0031] The monitoring unit is configured to input the preprocessed radio access network node network data and the preprocessed core network UPF network data into an AI monitoring model to obtain abnormal risk prediction information, wherein the AI monitoring model includes a multi-source feature encoding module, a space-time collaborative modeling module and a risk prediction module, the multi-source feature encoding module is configured to map the preprocessed radio access network node network data and the preprocessed core network UPF network data into a unified feature space respectively, the space-time collaborative modeling module is configured to construct a dynamic graph and extract a space-time dependence relationship of the dynamic graph based on a graph neural network, and the risk prediction module is configured to model a historical state sequence of a node in the dynamic graph using a multi-layer Transformer structure to output the abnormal risk prediction information.

[0032] The adjustment unit is configured to adjust parameter configurations of a radio side and a core network UPF side according to the abnormal risk prediction information.

[0033] The optimization unit is configured to feed back the adjusted parameter configurations of the radio side and the core network UPF side to a service awareness module to realize fault transfer.

[0034] In a feasible implementation, the multi-source feature encoding module includes a sliding window encoding unit, a heterogeneous domain fusion unit, a fusion mapping unit and a connection unit.

[0035] The sliding window encoding unit is configured to perform sliding window segmentation on the preprocessed radio access network node network data and the preprocessed core network UPF network data according to different time granularities to extract short-period mutation features and long-period trend features.

[0036] The heterogeneous domain fusion unit is configured to calculate feature attention weights of radio side data and core network UPF side data respectively, and perform intra-domain feature weighting based on the attention weights.

[0037] The fusion mapping unit is configured to map the weighted features to a unified semantic vector space through a deep neural network structure with shared weights to realize semantic alignment of different network domain data.

[0038] a connection unit configured to introduce a residual connection structure in the mapping process and combine a channel attention mechanism to enhance the stability of feature propagation.

[0039] In an implementation, the spatio-temporal collaborative modeling module includes a dynamic graph subunit, a graph neural network subunit, a graph evolution representation subunit, and a structure attribution unit.

[0040] The dynamic graph subunit is configured to map the feature vectors output by the multi-source feature encoding module into a plurality of graph structures at consecutive time steps, wherein each graph structure at a time step includes nodes, edges, and dynamic weights, and the weights of the edges are dynamically generated based on traffic synergy degree and time delay correlation.

[0041] The graph neural network subunit is configured to extract graph structure representations of nodes in each time step graph based on a graph attention mechanism, and the graph neural network subunit includes a time-aware embedding mechanism, an edge attribute embedding mechanism, and a skip connection mechanism.

[0042] The graph evolution representation subunit is configured to perform time series modeling on the node representations of the above-mentioned consecutive graph structures to obtain dynamic embedding features of the nodes in the historical evolution process.

[0043] The structure attribution unit is configured to interpret node states based on an attention mechanism in the graph neural network and output influence paths and key adjacent nodes of high-risk nodes.

[0044] In an implementation, the risk prediction module includes a historical state sequence subunit, a multi-layer Transformer subunit, a multi-scale risk perception subunit, and a risk causal explanation subunit.

[0045] The historical state sequence subunit is configured to extract node historical state vectors from the dynamic graph output by the spatio-temporal collaborative modeling module and construct an input sequence in combination with time position encoding.

[0046] The multi-layer Transformer subunit includes a plurality of encoding layers, each of which includes a multi-head attention layer, a feedforward network, a residual connection, and a normalization structure, and is configured to capture time series dependencies and abnormal evolution patterns in the node state sequence.

[0047] The multi-scale risk perception subunit is configured to generate abnormal trend prediction results, risk type classification labels, and confidence estimation information based on the Transformer output, respectively.

[0048] The risk causal explanation subunit is configured to identify key time slices and / or indicators that cause high-risk scores based on the attention weights and output risk explanation information.

[0049] In summary, the multi-modal data source probe state monitoring and fault transfer method provided by the application synchronously deploys data collection probes at the wireless access network node and the core network UPF node, and performs multi-source data fusion and space-time modeling based on an AI monitoring model. The method realizes unified probe deployment and data collection at the wireless side and the core network UPF side, and performs deep fusion of heterogeneous indicators through a multi-source feature coding module, solves the problem of fragmented collection and information isolation in the existing system, and improves the integrity of state perception. The space-time dependence between nodes is modeled by using a graph neural network, and deep learning is performed on the historical state sequence, which can realize early warning and evolution trend analysis of network anomalies, and is significantly better than the traditional lag detection method based on threshold or moving average. Through the attention mechanism, the system can provide structural explanation of network risk, support more accurate local parameter adjustment, and avoid blind parameter adjustment or misadjustment. The application introduces a strategy knowledge graph and a deep reinforcement learning double-driven control mechanism, which can automatically match candidate control strategies after risk identification, and adaptively select the optimal control action based on historical feedback, realize joint intelligent regulation and control of the wireless and UPF sides, and has self-evolution ability. The proposed method can adapt to different network topologies and probe protocols of different operators, and the AI model and the strategy module support modular replacement and online optimization, and have good engineering landing nature and long-term evolution potential. In summary, compared with the prior art, the application not only realizes a unified probe monitoring system across domains and layers, but also constructs a dynamic, interpretable and closed-loop network fault analysis and regulation mechanism through AI technologies such as graph neural network and reinforcement learning, which can significantly improve the accuracy of network state perception, the timeliness of response and the intelligent level of optimization, and has significant practical value and technical advancement.

[0050] The multi-modal data source probe state monitoring and fault transfer method provided by the application, other advantages, objects and characteristics of the application will be embodied in part through the following description, and will be understood by those skilled in the art through research and practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0051] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of the preferred embodiments and are not meant to limit the present description. Moreover, the same reference numerals in the attached drawings indicate the same or similar components. In the drawings:

[0052] Figure 1 A flowchart of a multi-modal data source probe state monitoring and fault transfer method provided by an embodiment of the application is shown in the figure;

[0053] Figure 2A flowchart for adjusting the parameter configuration of the wireless side and the core network UPF side according to the abnormal risk prediction information is provided for the embodiment of the present application.

[0054] Figure 3 A flowchart for determining the target control action for adjusting the wireless side and / or the core network UPF side is provided for the embodiment of the present application.

[0055] Figure 4 A multi-modal data source probe state monitoring and fault transfer system structure schematic diagram is provided for the embodiment of the present application. DETAILED DESCRIPTION

[0056] The terms "first", "second", "third", "fourth" and the like in the description and claims of the present application and above-mentioned appended drawings (if any) are used for distinguishing between similar objects, not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of the terms so construed can interchange, under appropriate circumstances, to refer to a similarly differentiated embodiment. Furthermore, the terms "comprise", "comprising", "include", "including", and the like, are intended to cover non-exclusive inclusions, such that a process, method, system, product, or apparatus that comprises a list of steps or units are not necessarily limited to those steps or units that are clearly listed, but can include other steps or units not expressly listed or inherent to such process, method, product, or apparatus. The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all.

[0057] Please refer to Figure 1 A multi-modal data source probe state monitoring and fault transfer method flowchart is provided for the embodiment of the present application, which can specifically include:

[0058] S110, data acquisition probes are respectively deployed at the wireless access network node and the core network UPF node to collect wireless access network node network data and core network UPF network data, and the above-mentioned network data includes network traffic, delay, packet loss, jitter, application connectivity and application network quality;

[0059] S120, pre-processing operation is performed on the above-mentioned wireless access network node network data and the above-mentioned core network UPF network data to obtain pre-processed wireless access network node network data and pre-processed core network UPF network data;

[0060] S130, input the preprocessed wireless access network node network data and the preprocessed core network UPF network data into an AI monitoring model to obtain abnormal risk prediction information, wherein the AI monitoring model comprises a multi-source feature encoding module, a space-time collaborative modeling module, and a risk prediction module, the multi-source feature encoding module is used to map the preprocessed wireless access network node network data and the preprocessed core network UPF network data into a unified feature space respectively, the space-time collaborative modeling module is used to construct a dynamic graph and extract its space-time dependence relationship based on a graph neural network; the risk prediction module is used to model the historical state sequence of the nodes in the dynamic graph to output abnormal risk prediction information;

[0061] S140, adjust the parameter configuration of the wireless side and the core network UPF side according to the abnormal risk prediction information;

[0062] S150, feedback to the service awareness module based on the adjusted parameter configuration of the wireless side and the core network UPF side to realize fault transfer.

[0063] In a feasible implementation, the network data includes network traffic, latency, packet loss, jitter, application connectivity, and application network quality.

[0064] In a feasible implementation, the risk prediction module adopts a multi-layer Transformer structure.

[0065] For example, data collection probes are deployed at wireless access network nodes (such as gNodeB, cell base station) and core network UPF (User Plane Function) nodes respectively. The probes are used to collect multi-dimensional performance data related to network running state in real time. The collected network data includes but is not limited to: network traffic indicators (such as uplink and downlink rates, connection numbers); latency indicators (such as one-hop latency, end-to-end latency); packet loss rate and jitter information; application layer connectivity (such as HTTP success rate, DNS response); application network quality (such as video freezing rate, MOS score, etc.).

[0066] Subsequently, a unified preprocessing operation is performed on the collected data, including timestamp alignment, noise removal, missing value filling, normalization, and feature structuring. Through the preprocessing process, standardized time series feature data for model use is obtained, ensuring that data from different network sides have consistency and modeling effectiveness.

[0067] Then, the preprocessed data is input into the AI monitoring model for intelligent analysis. The AI monitoring model consists of three parts: first, the multi-source feature encoding module is used to perform sliding window segmentation, attention weighted fusion and unified semantic space mapping on the wireless access network data and core network UPF data respectively, realizing the deep fusion of different domain data; second, the spatio-temporal collaborative modeling module is used to construct a dynamic graph, taking each network node as a vertex and the weight of the edge as the correlation between traffic coordination degree and delay, and based on the graph neural network to extract the spatio-temporal dependence features of each node at different time steps; finally, the risk prediction module uses a multi-layer Transformer structure to model the historical state sequence of the nodes in the dynamic graph, outputs the future risk score of each node, and gives the cause explanation of the high-risk nodes by combining the attention mechanism, realizing the explainable analysis of network anomalies.

[0068] After obtaining the abnormal risk prediction information, the system further dynamically adjusts the key parameters of the wireless side and the core network UPF side according to the abnormal location and type. For example, on the wireless side, the load or interference in the risk area can be relieved by adjusting the scheduling priority, controlling the transmission power, or optimizing the handover threshold; on the core network UPF side, network traffic guidance and control can be performed by path re-routing, QFI level switching or forwarding table management strategy. The above parameter configuration actions provide a candidate strategy set by the policy knowledge graph, and the optimal control strategy is selected by the deep reinforcement learning model according to the current state, so as to realize the multi-objective optimization of the network.

[0069] Finally, the adjusted parameter configuration result is fed back to the service perception module, which monitors the user service experience quality in real time and evaluates whether the adjustment effect meets the optimization target, forming a closed-loop control mechanism of perception-prediction-adjustment-feedback.

[0070] The multi-modal data source probe state monitoring and fault transfer method provided by the application synchronously deploys data collection probes at a radio access network node and a core network UPF node, and performs multi-source data fusion and space-time modeling based on an AI monitoring model. The method realizes unified probe deployment and data collection at the radio side and the core network UPF side, and performs deep fusion of heterogeneous indexes through a multi-source feature coding module, solves the problem of fragmented collection and information isolation in existing systems, and improves the integrity of state perception. The space-time dependency between nodes is modeled by using a graph neural network, and a historical state sequence is deeply learned by using a multi-layer Transformer, which can realize early warning and evolution trend analysis of network anomalies, and is significantly better than a traditional lag detection method based on a threshold or a sliding average. Through the attention mechanism, the system can provide a structural explanation of network risks, support more accurate local parameter adjustment, and avoid blind parameter adjustment or misadjustment. The application introduces a strategy knowledge graph and a deep reinforcement learning double-driven control mechanism, which can automatically match candidate control strategies after risk identification, and adaptively select the optimal control action based on historical feedback, realize joint intelligent regulation and control of the radio and UPF sides, and has self-evolution ability. The proposed method can adapt to different network topologies and probe protocols of different operators, and the AI model and the strategy module support modular replacement and online optimization, and have good engineering landing and long-term evolution potential. In summary, compared with the prior art, the application not only realizes a unified probe monitoring system across domains and layers, but also constructs a dynamic, interpretable and closed-loop network fault analysis and regulation mechanism by using AI technologies such as a graph neural network and reinforcement learning, which can significantly improve the accuracy of network state perception, the timeliness of response and the intelligent level of optimization, and has significant practical value and technical advancement.

[0071] In a feasible implementation manner, as shown in Figure 2 , Figure 2 A flowchart for adjusting parameter configurations of a radio side and a core network UPF side according to abnormal risk prediction information is provided for an embodiment of the application. The S140 adjusts the parameter configurations of the radio side and the core network UPF side according to the abnormal risk prediction information, and includes the following steps.

[0072] S1401, constructing a strategy knowledge graph, wherein the strategy knowledge graph is used to store the association relationship between abnormal types and network parameter items and the mapping relationship between network parameter items and control actions;

[0073] S1402, based on the abnormal risk prediction information, retrieving a candidate parameter adjustment strategy set matched with a current abnormal type from the strategy knowledge graph;

[0074] S1403, acquire current network state information, historical parameter adjustment records and network performance feedback information to construct a state representation for reinforcement learning;

[0075] S1404, based on the above state representation, perform optimal policy selection operation based on deep reinforcement learning algorithm in the above candidate parameter adjustment strategy set to determine the target control action for adjusting the wireless side and / or core network UPF side;

[0076] S1405, perform corresponding parameter adjustment operation according to the above target control action to realize optimization of network performance indicators, wherein the target control action aims to maximize network performance improvement, the network performance improvement is evaluated by a reward function, and the reward function is calculated based on risk score change, network delay change, packet loss rate change and QoS feedback information.

[0077] In a feasible implementation manner, as shown in Figure 3 , Figure 3 A flowchart for determining the target control action for adjusting the wireless side and / or core network UPF side is provided for the embodiment of the application, S1404, based on the above state representation, perform optimal policy selection operation based on deep reinforcement learning algorithm in the above candidate parameter adjustment strategy set to determine the target control action for adjusting the wireless side and / or core network UPF side, including:

[0078] S14041, input the above state representation into the policy selection model constructed based on deep reinforcement learning, wherein the policy selection model is used for action evaluation and policy selection in the candidate parameter adjustment strategy set, the policy selection model is trained by deep reinforcement learning algorithm to minimize the policy loss function and maximize the reward function value constructed based on network performance feedback, and the reward function is evaluated based on risk score change, network delay change, packet loss rate change and QoS indicator change;

[0079] S14042, calculate the policy score of each candidate action based on the current state representation in the above candidate parameter adjustment strategy set, and select the target control action with the highest expected long-term return, wherein the target control action is used as a network control instruction to adjust the parameter configuration of the corresponding wireless side and / or core network UPF side.

[0080] For example, first, the system constructs a strategy knowledge graph for storing the triple relationship between network anomaly types and parameter items, control actions. The graph is organized in the form of "anomaly type - parameter item - control action", such as "packet loss rate increase - UPF path - rerouting" or "delay jitter - scheduling weight - priority increase" and the like. The strategy knowledge graph not only records the adjustment behavior corresponding to the historical anomaly, but also integrates expert experience and data-driven results to provide structured support for subsequent candidate action generation.

[0081] After receiving the abnormal risk prediction information output by the AI monitoring module, the system retrieves the candidate parameter adjustment strategy set that best matches the current situation from the strategy knowledge graph according to the predicted abnormal type, location and intensity, to form a limited action space for the strategy decision stage.

[0082] Subsequently, the system obtains the running state information of the current network, including but not limited to node load, delay distribution, link utilization, historical risk level, etc., while extracting the past executed parameter adjustment action sequence and corresponding network performance feedback. The above information constitutes the state representation, denoted as:

[0083]

[0084] Among them: represents the multi-dimensional state characteristics of the current network node; A t-n:t-1 represents the executed action sequence in the past n time steps; R t-n:t-1 represents the network performance feedback (such as delay, packet loss, QoE index change, etc.) in the corresponding time.

[0085] The state vector is input into an Actor-Critic structure based on the PPO algorithm, which performs policy selection operations. The policy network (Actor) scores and samples actions from the candidate strategy set, and outputs the optimal control action under the current state The reward function R t is designed to measure the degree of network performance improvement brought by a certain strategy, and its calculation formula can be set as:

[0086] R t = λ1·ΔP t + λ2·ΔS t - λ3·C(a t ) + λ4·C t

[0087] Among them: ΔP t is the prediction consistency term, which measures the verification degree of the model on abnormal prediction, for example, whether the strategy action successfully makes the node be judged as high risk at the next moment when it is predicted that the future risk of a certain path will rise.

[0088] ΔS t is the comprehensive performance improvement item (which can be refined into a function),

[0089] ΔS t = f (ΔQoS t , ΔDelay t , ΔLoss t , ΔUtil t )

[0090] f (·) adopts a normalized dynamic weighted average function, and the weights are determined by the current load state of the network, for example: when the business is busy, ΔQoS t has a larger weight; in the network stable period, ΔDelay t and ΔUtil t are more important.

[0091] C (a t ) is the control action overhead item, C (a t ) = θ1×instruction complexity + θ2×impact range + θ3×QoS disturbance risk, considering the "disturbance cost" required to execute the action, such as the execution load of cutting UPF, upgrading scheduling priority, etc. A threshold value can be set, and a negative penalty is given when the control cost is higher than the threshold value.

[0092] C t is the predicted confidence adjustment item, C t = φ (Confidence) · (ΔRiskScore t ), where φ (·) is a confidence enhancement function (such as sigmoid), and if a high-confidence prediction result successfully avoids risk, it will increase the positive reward and strengthen the model's ability to "trust strong signals".

[0093] Through the above design, the reward function not only reflects the traditional performance improvement effect, but also introduces prediction-driven and control cost awareness, which helps the reinforcement learning model to achieve a more balanced, stable and interpretable strategy learning process in a complex network environment. The mechanism can continuously adjust the strategy network during training, so that it tends to select the optimal action that can effectively reduce network risk without causing excessive control cost, thereby improving the intelligent optimization ability of the entire system.

[0094] The embodiment constructs a strategy knowledge graph as expert experience support, combines a deep reinforcement learning strategy network for dynamic strategy selection, not only improves the intelligent level of abnormal processing, but also realizes joint optimization of multiple network indicators, has the significant advantages of strong interpretability, high execution efficiency and good adaptability, and is suitable for fault prediction and self-recovery control tasks in multiple types of network scenarios.

[0095] In an implementable embodiment, the target control action includes one or more of performing a cell scheduling threshold optimization operation on the wireless side, performing a user plane traffic path re-routing operation on the core network UPF side, performing a QoS identifier switching operation on the core network UPF side, and performing a forwarding table management operation on the core network UPF side.

[0096] For example, when the system detects an abnormality such as an increase in packet loss rate, scheduling congestion, or service queuing delay in a cell, the system can perform a dynamic adjustment operation on the scheduling priority of the cell. This operation can achieve preferential guarantee of scheduling resources and improve user experience by increasing the scheduling weight of critical services (such as VoIP and video) and reducing the scheduling priority of background services.

[0097] When the abnormality is represented by an increase in inter-cell interference, serious signal overlap, or uneven coverage, the system will trigger a transmit power control operation. Specifically, the operation includes reducing the transmit power of the interference source cell to reduce the signal leakage range, or increasing the transmit power of the cell with insufficient edge coverage to enhance the coverage capability, thereby improving the overall cell quality and handover success rate.

[0098] For frequent Ping-Pong handover, handover failure, or handover delay abnormality, the system can perform a handover threshold optimization operation. This operation can suppress unnecessary handover and improve handover success rate and network stability by dynamically adjusting the handover trigger threshold or delay time (T) such as RSRP and RSRQ.

[0099] On the core network UPF side, when the load of a specific UPF node is detected to be abnormal, the forwarding table entry is abnormal, or the path performance is deteriorated, the system can preferentially perform a user plane traffic path re-routing operation. This operation can achieve traffic bypassing and load balancing by updating the PFCP rule to guide user traffic to a low-risk, low-load alternative UPF path.

[0100] In addition, to cope with the failure of QoS guarantee for specific services, the system can perform a QoS identifier switching operation. This operation can improve the overall service level matching degree by updating the 5QI (QoS Flow Identifier) or QFI to assign the abnormal service flow to a higher service class, guaranteeing the transmission delay, bandwidth, and packet loss rate requirements.

[0101] When the UPF path has frequent abnormalities, table entries are continuously accumulated, or rule aging problems occur, the system can also perform a forwarding table management operation. This operation includes shortening the survival period of abnormal path entries, accelerating the recovery of invalid path entries, or increasing the update frequency of high-priority service entries to ensure the timeliness and efficiency of UPF forwarding tables, avoid abnormal spread, and waste of resources.

[0102] The above control actions can be automatically selected by the system according to the risk prediction result, the knowledge graph recommended strategy and the reinforcement learning evaluation result, and combined as needed to form a multi-action sequence to form joint regulation of complex scenarios. For example, when there is scheduling congestion and path abnormality in a certain cell at the same time, the system can jointly perform scheduling priority adjustment and UPF path re-routing operation to realize cross-domain and cooperative adaptive network optimization.

[0103] In summary, the target control action provided by the embodiment has the characteristics of fine operation granularity, high response accuracy and strong cross-domain linkage, and can realize closed-loop network self-optimization from prediction analysis to fine control, and is particularly suitable for complex communication network environments with concurrent multiple services and dynamic evolution of structure.

[0104] In a second aspect, as Figure 4 indicated, the present application also proposes a multi-modal data source probe state monitoring and fault transfer system for the method of any one of the first aspect, comprising:

[0105] The acquisition unit 21 is configured to deploy data acquisition probes at the radio access network node and the core network UPF node respectively to acquire radio access network node network data and core network UPF network data, wherein the network data includes network traffic, latency, packet loss, jitter, application connectivity and application network quality.

[0106] The preprocessing unit 22 is configured to perform preprocessing operations on the radio access network node network data and the core network UPF network data to obtain preprocessed radio access network node network data and preprocessed core network UPF network data.

[0107] The monitoring unit 23 is configured to input the preprocessed radio access network node network data and the preprocessed core network UPF network data into an AI monitoring model to obtain abnormal risk prediction information, wherein the AI monitoring model includes a multi-source feature encoding module, a space-time collaborative modeling module and a risk prediction module. The multi-source feature encoding module is configured to map the preprocessed radio access network node network data and the preprocessed core network UPF network data into a unified feature space, respectively. The space-time collaborative modeling module is configured to construct a dynamic graph and extract space-time dependence relationships of the dynamic graph based on a graph neural network. The risk prediction module uses a multi-layer Transformer structure to model the historical state sequence of the nodes in the dynamic graph to output the abnormal risk prediction information.

[0108] The adjustment unit 24 is configured to adjust the parameter configuration of the radio side and the core network UPF side according to the abnormal risk prediction information.

[0109] An optimization unit 25 is configured to feed back the adjusted wireless side and core network UPF side parameter configurations to the service awareness module to realize failover.

[0110] For example, the collection unit is deployed at the wireless access network node and the core network UPF node, and real-time network operation data are collected through setting data collection probes. The data include but are not limited to network traffic, time delay, packet loss, jitter, application connectivity, and application layer network quality, and comprehensively cover multiple dimensions of network performance.

[0111] The preprocessing unit performs preprocessing operations such as cleaning, denoising, and format unification on the collected raw network data to obtain structured and standardized preprocessed data corresponding to the wireless access network node and the core network UPF node.

[0112] The monitoring unit is responsible for inputting the preprocessed data into an AI monitoring model to realize intelligent prediction of abnormal risks. The AI model consists of three parts: a multi-source feature encoding module that uniformly maps heterogeneous data from the wireless side and the core network side to the same feature space to eliminate representation bias caused by source differences; a spatio-temporal collaborative modeling module that constructs a dynamic graph based on a graph neural network to extract the correlation and evolution pattern of each node in the network topology in space and time; and a risk prediction module that uses a multi-layer Transformer structure to model the historical state sequence of each node in the dynamic graph to accurately predict future abnormal risks.

[0113] The adjustment unit dynamically adjusts the relevant network parameters of the wireless side and the core network UPF side according to the abnormal risk prediction information output by the AI model to avoid potential failure risks or mitigate their impact in advance.

[0114] The optimization unit feeds back the adjusted parameter configuration results to the service awareness module to realize rapid reconfiguration of the network layer and intelligent switching of the service path, thereby completing the failover operation and ensuring the continuity and stability of network services.

[0115] In summary, the system proposed in the present application can realize abnormal monitoring and rapid adaptive optimization of key nodes in a communication network through end-to-end data collection, AI-driven risk prediction, and closed-loop failure response mechanisms, significantly improving the intelligent operation and maintenance capability and business stability of the system.

[0116] In a feasible implementation, the multi-source feature encoding module includes a sliding window encoding unit, a heterogeneous domain fusion unit, a fusion mapping unit, and a connection unit.

[0117] The sliding window encoding unit is configured to perform sliding window segmentation on the preprocessed wireless access network node data and the preprocessed core network UPF data according to different time granularities, so as to extract short-period mutation features and long-period trend features.

[0118] The heterogeneous domain fusion unit is configured to calculate feature attention weights of the wireless side data and the core network UPF side data respectively, and perform intra-domain feature weighting based on the attention weights.

[0119] The fusion mapping unit is configured to map the weighted features to a unified semantic vector space through a deep neural network structure sharing weights, so as to realize semantic alignment of data in different network domains.

[0120] The connection unit is configured to introduce a residual connection structure in the mapping process and combine a channel attention mechanism to enhance the stability of feature propagation.

[0121] For example, the sliding window encoding unit performs sliding window segmentation on the data of the wireless side and the core network side according to different time granularities. By setting multiple time window lengths, the unit can extract short-time mutation features (such as burst traffic, jitter surge, etc.) and capture trend changes in a long period (such as gradual increase in packet loss rate, continuous decline in application connectivity, etc.). The data in the window forms feature segments of local time periods after segmentation, laying a foundation for modeling the time evolution process.

[0122] Then, the heterogeneous domain fusion unit performs fusion processing on the data from different network domains. The unit calculates feature attention weights of the wireless access network node data and the core network UPF node data respectively, highlights key feature dimensions through a self-attention mechanism, and strengthens the attention to important indicators (such as RTT fluctuation, change in forwarding table hit rate, etc.). The weighted features can better reflect the importance distribution of intra-domain features, providing a premise for unified mapping.

[0123] Then, the fusion mapping unit inputs the weighted features into a deep neural network structure sharing weights. The neural network is responsible for mapping the feature vectors of the wireless side and the core network side to a unified semantic vector space, eliminating the representation differences between different network domains, so that the subsequent graph structure construction can perform topology modeling and information propagation in the same embedding dimension. This operation guarantees the cross-domain consistency of data representation, improves the generalization ability and learning effect of the entire model.

[0124] Finally, the connection unit introduces a residual connection structure in the feature propagation process, and combines with the channel attention mechanism to enhance the overall model's ability to retain deep features and identify important channels. The residual structure can alleviate the gradient vanishing problem in deep networks, making the model training more stable; the channel attention mechanism further enhances the model's expression ability for important feature channels through a dynamic weighting mechanism, thereby enhancing the model's sensitivity to abnormal behavior while maintaining feature integrity.

[0125] The embodiment realizes the time series modeling and deep integration of wireless side and core network UPF side network data through the organic combination of sliding window, attention weighting, shared mapping and channel enhancement, and builds a unified multi-source feature representation, laying a solid data foundation for subsequent dynamic graph modeling and risk prediction. The method has good modeling ability for short-term anomalies and long-term trends, and has good expressiveness, expansibility and adaptability.

[0126] In a feasible implementation, the above-mentioned spatio-temporal collaborative modeling module includes a dynamic graph subunit, a graph neural network subunit, a graph evolution representation subunit, and a structure attribution unit;

[0127] The above-mentioned dynamic graph subunit is configured to map the feature vectors output by the multi-source feature encoding module into a plurality of continuous time step graph structures, wherein each time step graph structure includes nodes, edges, and dynamic weights, and the weights of the edges are dynamically generated based on traffic synergy degree and time delay correlation;

[0128] The above-mentioned graph neural network subunit is configured to extract the graph structure representation of the nodes in each time step graph based on a graph attention mechanism, and the graph neural network subunit includes a time-aware embedding mechanism, an edge attribute embedding mechanism, and a skip connection mechanism;

[0129] The above-mentioned graph evolution representation subunit is configured to perform time series modeling on the node representation of the continuous graph structure to obtain dynamic embedding features of the nodes in the historical evolution process;

[0130] The above-mentioned structure attribution unit is configured to interpret the node state based on the attention mechanism in the graph neural network, and output the influence path and key adjacent nodes of the high-risk nodes.

[0131] Exemplarily, the dynamic graph subunit is responsible for organizing the node feature vectors output by the multi-source feature encoding module into a graph structure at multiple consecutive time steps. At each time step, the graph structure is composed of nodes, edges, and edge weights, where the nodes represent network entities (such as gNodeB nodes, UPF nodes, etc.), and the presence of edges indicates that there is a certain business connection or influence between the network entities. Specifically, the weights of the edges are dynamically generated based on the traffic correlation (e.g., concurrent connection trend) and latency correlation (e.g., path RTT fluctuation consistency) between two nodes to characterize the interaction strength between entities in the network at that time step. The subunit finally outputs a series of weighted graph structures as the basis for subsequent modeling.

[0132] Subsequently, the graph neural network subunit extracts structural information from the graph structure at each time step to generate a graph embedding representation of the nodes. To improve the model's ability to model temporal evolution and network topology changes, this subunit integrates multiple key mechanisms. First, a time-aware embedding mechanism is introduced to fuse the timestamp information of the current time step with the node features, so that the node embedding not only contains structural information but also reflects temporal change characteristics. Second, an edge attribute embedding mechanism is used to encode the traffic correlation, latency correlation, and other information contained in the edge weights into the graph convolution operation process, enabling more accurate relationship modeling. Finally, a skip connection mechanism is used to connect different graph convolution layers to avoid the over-smoothing problem that occurs in deep graph neural networks, enhancing the stability and non-linear representation ability of the model.

[0133] After the graph neural network extracts the node representation at each time step, the graph evolution representation subunit models the node embedding sequence at all time steps to obtain the dynamic feature trajectory of each node in the historical evolution process. This subunit can use a recurrent neural network (such as GRU) or a Transformer-based time series modeling structure to model the representation of the same node in consecutive graph structures to capture its behavior evolution pattern and abnormal development trend, and finally output a dynamic embedding representation with temporal dependence for subsequent risk prediction or visualization analysis.

[0134] Finally, the structural attribution subunit explains the node state based on the attention weights in the graph neural network. By analyzing the attention distribution of high-risk nodes, the system can identify the influence path and key adjacent nodes behind their high-risk scores, i.e., identify the "upstream nodes" or key propagation links that cause abnormalities. The output of this module not only contains the risk score of the high-risk node but also contains its structural attribution path graph, which can be used as a reference for subsequent parameter adjustment or network policy formulation modules, thereby improving the accuracy and explainability of fault handling.

[0135] The embodiment realizes high-fidelity modeling of dynamic correlation between network entities and fine description of node behavior evolution process through the synergy of graph modeling and time series modeling, and outputs abnormal influence path of high-risk nodes through structural attribution mechanism, significantly improving the insight and intelligent intervention ability of complex network running state. The module is suitable for various types of network anomaly analysis scenarios, and has strong scalability and engineering adaptability.

[0136] In an available embodiment, the risk prediction module comprises a historical state sequence subunit, a multi-layer Transformer subunit, a multi-scale risk perception subunit, and a risk causal explanation subunit.

[0137] The historical state sequence subunit is configured to extract node historical state vectors from the dynamic graph output by the spatiotemporal collaborative modeling module, and construct input sequences in combination with time position encoding.

[0138] The multi-layer Transformer subunit comprises a plurality of encoding layers, each of which comprises a multi-head attention layer, a feedforward network, a residual connection, and a normalization structure, and is configured to capture temporal dependencies and abnormal evolution patterns in the node state sequence.

[0139] The multi-scale risk perception subunit is configured to generate abnormal trend prediction results, risk type classification labels, and confidence estimation information based on the Transformer output.

[0140] The risk causal explanation subunit is configured to identify key time segments and / or indicators that lead to high-risk scores based on the attention weights, and output risk explanation information.

[0141] For example, the historical state sequence subunit extracts embedding feature vectors of a target node at multiple time steps from the dynamic graph output by the spatiotemporal collaborative modeling module. In order to preserve the time sequence information, the subunit introduces a time position encoding mechanism, encodes the timestamp corresponding to each time step embedding, injects time information into the state representation, and constructs an input sequence with time perception ability. The sequence is input into the risk prediction model, reflecting the dynamic change trajectory of the node in the historical evolution process, which is the key basis for capturing temporal patterns.

[0142] Subsequently, the input sequence enters a multi-layer Transformer subunit for deep modeling. The subunit is composed of multiple stacked Transformer encoding layers, each of which contains a multi-head attention mechanism, a feedforward network structure, a residual connection, and a layer normalization mechanism. The multi-head attention mechanism can model the dependencies between time segments in the sequence in a global range, capturing the evolution trend and potential abnormal precursors of node states; the residual connection and normalization operation improve the training stability and nonlinear expression ability of the network. Through multi-layer stacking, the model can learn complex temporal evolution laws and provide high-quality representations for downstream risk prediction tasks.

[0143] After obtaining the Transformer output, a multi-scale risk perception subunit further interprets it and outputs multiple tasks. This subunit realizes the multi-dimensional expression ability of risk in a multi-channel structure, evaluating risk from different scales and perspectives. On the one hand, it generates abnormal trend prediction results to quantify the degree of abnormal risk change in the future; on the other hand, it classifies the risk type, such as determining whether it is caused by time delay fluctuation, packet loss anomaly, or path congestion, and outputs the corresponding label; in addition, the unit also generates confidence estimation information for each prediction result to help judge the credibility of the prediction conclusion and provide a risk reliability reference for the subsequent regulation module.

[0144] Finally, the risk causal explanation subunit performs interpretable analysis on the abnormal prediction results based on the attention weights inside the Transformer model. This subunit analyzes the time segments and index dimensions that play a dominant role in the generation of abnormal scores and identifies which historical time or performance indicators have the greatest contribution to high-risk scores. For example, if it is found that the region with a sudden increase in packet loss rate in the past period is strongly focused on by the attention weight, it is labeled as a "key abnormal driving factor" and output in the form of structured explanation information for the strategy selection module and operation and maintenance personnel to reference.

[0145] The risk prediction module provided by the embodiment not only has fine temporal modeling capability and comprehensive risk perception capability, but also has strong explainability and auxiliary decision-making capability. By introducing a multi-layer Transformer structure and a multi-task risk output mechanism, the system can realize early warning and hierarchical evaluation of network state; and through causal analysis driven by the attention mechanism, the transparency and controllability of the prediction results are improved, significantly enhancing the application value of the intelligent operation and maintenance system in complex network environments.

[0146] The application further provides a computer program product, which comprises a computer program or computer executable instructions stored in a computer readable storage medium. A processor of an electronic device reads the computer program or computer executable instructions from the computer readable storage medium, and the processor executes the computer program or computer executable instructions, so that the electronic device performs any step of the multi-modal data source probe state monitoring and fault transfer method described above.

[0147] The above examples are only used to illustrate the technical solutions of the present application, but not to limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-modal data source probe state monitoring and failover method, characterized by, The method comprises the following steps: Deploying data collection probes at the wireless access network node and the core network UPF node respectively to collect wireless access network node network data and core network UPF network data; Performing preprocessing operations on the wireless access network node network data and the core network UPF network data to obtain preprocessed wireless access network node network data and preprocessed core network UPF network data; Inputting the preprocessed wireless access network node network data and the preprocessed core network UPF network data into an AI monitoring model to obtain abnormal risk prediction information, wherein the AI monitoring model comprises a multi-source feature encoding module, a spatio-temporal collaborative modeling module, and a risk prediction module, the multi-source feature encoding module is used to map the preprocessed wireless access network node network data and the preprocessed core network UPF network data into a unified feature space respectively, the spatio-temporal collaborative modeling module is used to construct a dynamic graph and extract the spatio-temporal dependence relationship of the dynamic graph based on a graph neural network, the nodes in the dynamic graph correspond to wireless access network nodes and core network UPF nodes, and the weights of the edges are dynamically generated based on traffic synergy degree and time delay correlation; the risk prediction module is used to model the historical state sequence of the nodes in the dynamic graph to output the abnormal risk prediction information; Adjusting the parameter configurations of the wireless side and the core network UPF side according to the abnormal risk prediction information; Feeding back the adjusted parameter configurations of the wireless side and the core network UPF side to a service perception module to realize fault transfer; The method of adjusting the parameter configurations of the wireless side and the core network UPF side according to the abnormal risk prediction information comprises: Constructing a policy knowledge graph, wherein the policy knowledge graph is used to store the association relationship between abnormal types and network parameter items and the mapping relationship between network parameter items and control actions; Based on the abnormal risk prediction information, retrieving a candidate parameter adjustment strategy set matching the current abnormal type from the policy knowledge graph; Obtaining current network state information, historical parameter adjustment records, and network performance feedback information to construct a state representation for reinforcement learning; Based on the state representation, performing optimal strategy selection operations based on a deep reinforcement learning algorithm in the candidate parameter adjustment strategy set to determine a target control action for adjusting the wireless side and / or the core network UPF side; Performing corresponding parameter adjustment operations according to the target control action to optimize network performance indicators, wherein the target control action aims to maximize network performance improvement, the network performance improvement is evaluated by a reward function, and the reward function is calculated based on risk score changes, network delay changes, packet loss rate changes, and QoS feedback information.

2. The multi-modal data source probe state monitoring and failover method of claim 1, wherein, The network data includes network traffic, time delay, packet loss, jitter, application connectivity, and application network quality.

3. The multi-modal data source probe state monitoring and failover method of claim 1, wherein, The risk prediction module adopts a multi-layer Transformer structure.

4. The multi-modal data source probe state monitoring and failover method of claim 1, wherein, The optimal policy selection operation based on the deep reinforcement learning algorithm is performed in the candidate parameter adjustment strategy set based on the state representation to determine a target control action for adjusting the wireless side and / or the core network UPF side, including: The state representation is input into a policy selection model constructed based on deep reinforcement learning, wherein the policy selection model is used for action evaluation and policy selection in the candidate parameter adjustment strategy set, and the policy selection model is trained by a deep reinforcement learning algorithm to minimize a policy loss function and maximize a value calculated by a reward function constructed based on network performance feedback, wherein the reward function is calculated based on risk score changes, network delay changes, packet loss rate changes, and QoS indicator changes; The policy score of each candidate action is calculated based on the current state representation in the candidate parameter adjustment strategy set, and a target control action with the highest expected long-term return is selected, wherein the target control action is used as a network control instruction to adjust the parameter configuration of the corresponding wireless side and / or core network UPF side; The target control action includes one or more of dynamic adjustment of cell scheduling priority on the wireless side, transmit power control on the wireless side, handover threshold optimization on the wireless side, user plane traffic path re-routing on the core network UPF side, QoS identifier switching on the core network UPF side, and forwarding table management on the core network UPF side.

5. A multi-modal data source probe status monitoring and failover system for implementing the method of any one of claims 1 to 4, characterized in that, It includes: The acquisition unit is used for deploying data acquisition probes at wireless access network nodes and core network UPF nodes respectively to collect wireless access network node network data and core network UPF network data, wherein the network data includes network traffic, delay, packet loss, jitter, application connectivity, and application network quality. The preprocessing unit is used for preprocessing the wireless access network node network data and the core network UPF network data to obtain preprocessed wireless access network node network data and preprocessed core network UPF network data. The monitoring unit inputs the preprocessed wireless access network node network data and the preprocessed core network UPF network data into an AI monitoring model to obtain abnormal risk prediction information, wherein the AI monitoring model includes a multi-source feature encoding module, a spatio-temporal collaborative modeling module, and a risk prediction module, the multi-source feature encoding module is used to map the preprocessed wireless access network node network data and the preprocessed core network UPF network data into a unified feature space respectively, the spatio-temporal collaborative modeling module is used to construct a dynamic graph and extract spatio-temporal dependence relationships of the dynamic graph based on a graph neural network; the risk prediction module uses a multi-layer Transformer structure to model the historical state sequence of the nodes in the dynamic graph to output the abnormal risk prediction information; The adjustment unit is used for adjusting the parameter configuration of the wireless side and the core network UPF side according to the abnormal risk prediction information. An optimization unit is configured to feed back the adjusted wireless side and core network UPF side parameter configurations to the service awareness module to achieve failover.

6. The multi-modal data source probe status monitoring and failover system of claim 5, wherein, The multi-source feature encoding module comprises a sliding window encoding unit, a heterogeneous domain fusion unit, a fusion mapping unit and a connection unit. The sliding window encoding unit is configured to perform sliding window segmentation on the preprocessed wireless access network node network data and the preprocessed core network UPF network data according to different time granularities, to extract short-period mutation features and long-period trend features. The heterogeneous domain fusion unit is configured to calculate feature attention weights of wireless side data and core network UPF side data respectively, and perform intra-domain feature weighting based on the attention weights, to obtain weighted feature outputs. The fusion mapping unit is configured to map the weighted features output by the heterogeneous domain fusion unit to a unified semantic vector space through a deep neural network structure with shared weights, to realize semantic alignment of different network domain data. The connection unit is configured to introduce a residual connection structure in the mapping process and combine a channel attention mechanism to enhance the stability of feature propagation.

7. The multi-modal data source probe status monitoring and failover system of claim 5, wherein, The spatio-temporal collaborative modeling module comprises a dynamic graph atlas subunit, a graph neural network subunit, a graph atlas evolution representation subunit and a structure attribution factor unit. The dynamic graph atlas subunit is configured to map the feature vectors output by the multi-source feature encoding module into a plurality of continuous time step graph structures, wherein each time step graph structure comprises nodes, edges and dynamic weights, the nodes correspond to wireless access network nodes or core network UPF nodes, and the weights of the edges are dynamically generated based on traffic collaboration degree and time delay correlation. The graph neural network subunit is configured to extract graph structure representations of nodes in each time step graph based on a graph attention mechanism, and comprises a time-aware embedding mechanism, an edge attribute embedding mechanism and a skip connection mechanism. The graph atlas evolution representation subunit is configured to perform time series modeling on the node representations of the continuous time step graph structures, to obtain dynamic embedding features of the nodes in the historical evolution process. The structure attribution factor unit is configured to interpret node states based on an attention mechanism in the graph neural network, and output an influence path and key adjacent nodes of a high-risk node.

8. The multi-modal data source probe status monitoring and failover system of claim 6, wherein, The risk prediction module comprises a historical state sequence subunit, a multi-layer Transformer subunit, a multi-scale risk awareness subunit and a risk causal explanation subunit. The historical state sequence subunit is configured to extract node historical state vectors from the dynamic graph atlas output by the spatio-temporal collaborative modeling module, and construct a node state sequence as an input sequence of the multi-layer Transformer subunit. The multi-layer Transformer subunit comprises a plurality of encoding layers, each encoding layer comprising a multi-head attention layer, a feedforward network, a residual connection and a normalization structure, for capturing time series dependencies and abnormal evolution patterns in the node state sequence. The multi-scale risk awareness subunit is configured to generate an abnormal trend prediction result, a risk type classification label and confidence estimation information based on the Transformer output, respectively. The risk causal explanation subunit is configured to identify key time segments and / or indicators that cause a high risk score based on the attention weights and output risk explanation information.

Images