A method and system for multi-modal fault root cause localization in a cable television network

Abstract

The application discloses a cable television network multi-modal fault root cause positioning method and system, and mainly relates to the technical field of cable television networks. It comprises the following steps: collecting multi-source operation and maintenance data and triggering a diagnosis event; screening topological nodes based on a service flow path, constructing a multi-modal data unit; extracting cross-modal features and introducing a topological weight matrix to construct a joint causal feature vector; combining topological level constraints to perform initial root cause inference; modifying the root cause probability based on fault propagation consistency; and outputting a diagnosis result and updating the model based on feedback. The application has the beneficial effects that it realizes the structured fusion of multi-modal data, introduces topological constraints and propagation consistency verification, significantly improves the positioning accuracy and interpretability of complex chain faults, and has self-evolution ability.

Description

Technical Field

[0001] This invention relates to the field of cable television network technology, specifically a method and system for locating the root cause of multimodal faults in cable television networks based on topology constraints. Background Technology

[0002] Currently, cable television networks are evolving towards IP-based and cloud-network convergence, with the network scale continuously expanding and the hierarchical structure becoming increasingly complex. In multi-level topologies such as provincial core networks, backbone transmission networks, and access networks, the relationships between network nodes and links exhibit highly complex characteristics. Existing fault diagnosis technologies mainly rely on network management systems based on protocols such as SNMP, whose monitoring data is limited to equipment operating parameters and alarm information, lacking the ability to directly perceive end-to-end video service quality. Traditional diagnostic methods are usually based on single data modal analysis and fail to combine network topology for constrained reasoning, resulting in a lack of modeling ability for signal propagation paths and hierarchical relationships in the diagnostic logic.

[0003] When video service anomalies such as mosaic or still frames occur, maintenance personnel need to manually correlate and analyze multi-source heterogeneous data, including video screenshots, network operating parameters, and user fault reports. This data is scattered across different systems, lacking a unified topology node identification system and time alignment mechanism. This makes it impossible to perform structured filtering and correlation analysis based on the service signal propagation path, resulting in significant data silos. Furthermore, existing diagnostic processes lack modeling and verification mechanisms for network physical propagation patterns. Even when preliminary judgments are made using multimodal data, propagation consistency verification of candidate fault root causes is not performed, easily leading to misjudgments where local feature matches do not conform to the propagation path logic. Especially in complex fault scenarios involving cross-level and chain-like propagation, existing technologies struggle to automatically identify the matching relationship between the distribution of abnormal nodes and the theoretical propagation range. Therefore, in complex cable television network topologies, traditional diagnostic methods relying on manual experience and lacking topology constraints and propagation consistency mechanisms are no longer sufficient to meet the demands of the ultra-high-definition era for high-precision, interpretable, and automated root cause localization.

[0004] Therefore, there is an urgent need for a method and system for locating the root cause of multimodal faults in cable television networks based on topology constraints to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for locating the root cause of multimodal faults in cable television networks based on topology constraints. It realizes the structured fusion of multimodal data, introduces topology constraints and propagation consistency verification, significantly improves the accuracy and interpretability of locating complex chain faults, and has self-evolution capabilities.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] On one hand, the present invention provides a method for locating the root cause of multimodal faults in a cable television network, comprising the following steps:

[0008] Step S1: Collect multi-source operation and maintenance data of the cable television network, generate diagnostic tasks according to preset diagnostic trigger conditions, and lock the time interval and service flow identifier corresponding to the diagnostic event;

[0009] Step S2: Based on the service flow identifier, locate the actual signal transmission path of the service flow in the network topology database, filter out the set of topology nodes located on the path, and perform time alignment and topology location encoding on the multi-source data according to the diagnostic time window to construct a multimodal data unit with structural semantics.

[0010] Step S3: Perform cross-modal feature extraction on video image data, network device operating parameter data and text data in the multimodal data unit, and construct a joint causal feature vector by combining the topology weight matrix and causal correlation analysis;

[0011] Step S4: Input the joint causal feature vector into the preset root cause discrimination model, and calculate the initial probability distribution of each candidate root cause by combining the topological hierarchy constraint coefficients.

[0012] Step S5: Based on the network topology and the actual distribution of detected abnormal nodes, calculate the propagation consistency matching degree between the theoretical influence range and the actual abnormal range of each candidate root cause, and correct the initial probability by physical law constraints to obtain the final root cause probability distribution.

[0013] Step S6: Determine the diagnostic results based on the final root cause probability distribution and output structured fault information; at the same time, perform incremental training and update of model parameters based on operation and maintenance feedback results.

[0014] Preferably, in step S2, constructing a multimodal data unit with structural semantics specifically involves:

[0015] Based on the service flow identifier, locate the actual transmission path of the service in the topology database and generate a set of topology nodes. ; based on the time of the diagnostic event Set a time window for the center For the collected multi-source data set Execute structured filtering rules:

[0016] ;

[0017] in, Representing data The topology node to which it belongs. The timestamp represents the data;

[0018] After filtering, data located on the business transmission path and within the time window are retained, and topological location codes are added to the data of each node, including node level identifier, number of hops from the source and upstream and downstream direction attributes, to form a spatial structure feature representation, which is finally encapsulated into a multimodal data unit.

[0019] Preferably, step S3 specifically includes:

[0020] For video image sequences Extracting deep feature vectors ,in For visual encoding models; through temporal aggregation functions Calculate inter-frame structural differences and block true intensity to generate service anomaly feature vectors. ;

[0021] For network nodes Parameters of time series Calculate gradient features and offset features And concatenate them to form a parameter state feature vector. For all path nodes The parameter features are aggregated to obtain the global network parameter state features. ;

[0022] Text data is input into a text semantic encoding model The fault semantic feature vector is obtained. ;

[0023] Set the dynamic topology weight matrix ;

[0024] Calculate the feature vector of business anomalies Gradient features of each node correlation coefficient When the correlation coefficient exceeds a preset threshold, the topological location code of the node is added to the joint feature representation;

[0025] Constructing a joint causal feature vector:

[0026] ;

[0027] The joint causal feature vector integrates multimodal features and topology information to describe the causal relationship between network state changes and service anomalies.

[0028] Preferably, the topological weight matrix The dynamic calculation process uses a graph attention network. The inputs are the real-time load, historical failure frequency, link bandwidth utilization and service flow path information of each node. The output is the attention coefficient matrix between nodes, which is used to characterize the relative importance of nodes in fault propagation.

[0029] Preferably, in step S4, calculating the initial probability distribution of each candidate root cause specifically involves:

[0030] The candidate root causes are divided into core layer, transport layer and access layer subsets according to the topology level, so that they match the topology sub-path structure generated in step S2.

[0031] Joint causal eigenvectors Input the root cause discriminant function and calculate the root cause probability:

[0032] ;

[0033] in, Indicates the first Candidate root causes; The discriminant function is expressed as:

[0034] ;

[0035] in, This is the result of a nonlinear mapping of joint causal features. The coefficient representing the degree of matching between the root cause category and the current anomalous topological distribution. and These are trainable parameters.

[0036] Preferably, step S5 specifically includes:

[0037] For each candidate root cause Based on the topology database and fault knowledge rules, determine the theoretical set of affected nodes. ;

[0038] Based on the abnormal nodes identified in steps S2 and S3, determine the set of nodes that actually detected abnormal features. ;

[0039] Calculate the propagation matching coefficient:

[0040] ;

[0041] The propagation matching coefficient is used to quantify the degree of overlap between the theoretical propagation range and the actual abnormal distribution;

[0042] Introducing propagation attenuation factor The weights of theoretical nodes that are far from the abnormal concentration area are reduced to enhance the physical rationality of the matching calculation.

[0043] For initial probability Make corrections:

[0044] ;

[0045] The final root cause probability distribution conforming to the laws of network physical propagation is obtained.

[0046] Preferably, the propagation attenuation factor The value range is [0,1]. Its value decreases exponentially with respect to the hop distance between the node and the anomaly concentration area. The farther the distance, the lower the weight. It is used to simulate the attenuation characteristics of the signal during transmission.

[0047] Preferably, step S6 specifically includes:

[0048] According to the probability distribution of the final root cause Select the fault type with the highest probability as the final diagnostic root cause, and output its confidence level, scope of influence and location information of abnormal nodes;

[0049] The diagnostic results, multimodal data units, and joint causal feature vectors are archived and stored to form traceable fault cases;

[0050] If the actual root cause confirmation results received from the operation and maintenance personnel are inconsistent with the diagnosis results, the sample will be marked as a misjudged sample and added to the training set.

[0051] A periodic incremental training mechanism is used to update the parameters of the feature extraction model in step S3 and the root cause discrimination model in step S4, so that the model can adapt to new failure modes and network topology changes.

[0052] On the other hand, the present invention also provides a multimodal fault root cause localization system for cable television networks, used to implement the multimodal fault root cause localization method for cable television networks as described above, including:

[0053] The data acquisition module is used to collect data from multiple sources and generate diagnostic tasks;

[0054] The topology constraint processing module is used to filter topology nodes based on business flow paths and construct multimodal data units.

[0055] The feature modeling module is used to extract multimodal features and construct joint causal feature vectors;

[0056] The root cause inference module is used to calculate the initial root cause probability and perform propagation consistency correction.

[0057] The diagnostic output and self-evolution module is used to output diagnostic results and update model parameters based on feedback.

[0058] Preferably, the topology constraint processing module further includes a topology weight dynamic calculation unit, which calculates the topology weight matrix between nodes based on the real-time network status and business flow path information through a graph attention network, and is used to guide the causal correlation analysis in the feature modeling process.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0060] 1. This invention breaks down data silos between video images, network parameters, and text work orders by constructing a unified multi-source data acquisition and topology node identification system, thereby achieving structured fusion analysis of cross-modal data;

[0061] 2. This invention introduces a network topology constraint mechanism, filters data based on actual business flow paths and constructs a topology weight matrix, which deeply couples the diagnostic process with the signal propagation structure, significantly improving the positioning accuracy in complex chain fault scenarios.

[0062] 3. The cross-modal causal association modeling method of the present invention maps visual anomalies, parameter changes and text semantics to a unified feature space, and combines dynamic topological weights and correlation analysis to enhance the interpretability of diagnostic results;

[0063] 4. The fault propagation consistency verification mechanism of the present invention corrects the initial probability by quantifying the overlap between the theoretical influence range and the actual anomaly distribution, ensuring that the diagnostic results conform to the actual network propagation mechanism and effectively reducing the risk of misjudgment. Attached Figure Description

[0064] Figure 1 This is a flowchart of the method of the present invention;

[0065] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0066] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0067] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0068] Example:

[0069] like Figure 1 As shown in the figure, this embodiment provides a method for locating the root cause of multimodal faults in a cable television network, including the following steps:

[0070] Step S1: Collect multi-source operation and maintenance data of the cable television network, generate diagnostic tasks according to preset diagnostic trigger conditions, and lock the time interval and service flow identifier corresponding to the diagnostic event;

[0071] Step S2: Based on the service flow identifier, locate the actual signal transmission path of the service flow in the network topology database, filter out the set of topology nodes located on the path, and perform time alignment and topology location encoding on the multi-source data according to the diagnostic time window to construct a multimodal data unit with structural semantics.

[0072] Step S3: Perform cross-modal feature extraction on video image data, network device operating parameter data and text data in the multimodal data unit, and construct a joint causal feature vector by combining the topology weight matrix and causal correlation analysis;

[0073] Step S4: Input the joint causal feature vector into the preset root cause discrimination model, and calculate the initial probability distribution of each candidate root cause by combining the topological hierarchy constraint coefficients.

[0074] Step S5: Based on the network topology and the actual distribution of detected abnormal nodes, calculate the propagation consistency matching degree between the theoretical influence range and the actual abnormal range of each candidate root cause, and correct the initial probability by physical law constraints to obtain the final root cause probability distribution.

[0075] Step S6: Determine the diagnostic results based on the final root cause probability distribution and output structured fault information; at the same time, perform incremental training and update of model parameters based on operation and maintenance feedback results.

[0076] like Figure 2 As shown, this embodiment also provides a system for implementing the above steps, including:

[0077] The data acquisition module is used to collect multi-source data such as video images, network operation parameters, and text work orders, and to generate diagnostic tasks;

[0078] The topology constraint processing module is used to filter topology nodes based on business flow paths and construct multimodal data units.

[0079] The feature modeling module is used to extract multimodal features and construct joint causal feature vectors;

[0080] The root cause inference module is used to calculate the initial root cause probability and perform propagation consistency correction.

[0081] The diagnostic output and self-evolution module is used to output diagnostic results and update model parameters based on feedback.

[0082] The topology constraint processing module also includes a topology weight dynamic calculation unit. This unit calculates the topology weight matrix between nodes based on the real-time network status and business flow path information through a graph attention network, which is used to guide the causal relationship analysis in the feature modeling process.

[0083] The following is a detailed description of the above methods, steps, and system:

[0084] In this embodiment, a structured multi-source data acquisition mechanism for the operation status of cable television networks is first constructed to provide a unified data foundation for subsequent topology constraint analysis. The system is deployed on the provincial broadcast television network intelligent operation and maintenance platform. Through a distributed acquisition architecture, it continuously acquires the operation data of the service layer, network layer and operation and maintenance management layer. All acquired data undergoes unified identification mapping processing when entering the system, converting the device number, link identifier and service ID of different systems into a unified topology node identifier system and establishing a one-to-one mapping relationship with the network topology database.

[0085] At the service perception layer, video quality monitoring probes deployed at core nodes or key network locations periodically sample frames or screenshots of the live service stream to generate video image data, and attach millisecond-level unified timestamps, channel identifiers, and acquisition node identifiers. The timestamps are calibrated using a central clock source to keep cross-system time errors within a preset threshold range. When continuous abnormal images, mosaics, or still frames are detected, a service abnormality marker is generated.

[0086] At the network operation layer, network device operating parameters and alarm information, including optical power, signal-to-noise ratio, bit error rate and device status, are collected in real time through SNMP or NETCONF interface, and time stamps and topology node codes are uniformly added. At the operation and maintenance management layer, the system accesses customer service work order data and performs log calibration processing on the time of user fault reporting.

[0087] When conditions such as abnormal video quality, alarm of critical equipment, or concentrated fault reporting within a preset time window are met, the system generates a diagnostic task identifier and locks the relevant data stream within the corresponding time interval, providing structured input for subsequent multimodal association processing based on topological constraints.

[0088] After a diagnostic event is triggered, the system locates the signal transmission path of the service in the network topology database based on the abnormal service identifier. Unlike traditional topology filtering based on physical connections, this invention generates a corresponding set of topology nodes based on the actual forwarding path of the service flow (e.g., for IP video multicast services, based on the multicast distribution tree; for point-to-point video streams, based on the MPLS LSP or SRv6 path). This ensures that the selected nodes are strictly located on the end-to-end transmission link of the service flow. Then, based on the event occurrence time... Set a time window for the center For the collected data set Execute structured filtering rules:

[0089] ;

[0090] in, Indicates the topology node to which the data belongs. This indicates the data timestamp. This filtering mechanism retains only data located within the business propagation path and within the time window, thus eliminating interference from irrelevant business data in the diagnostic process. For data within the same time window but not belonging to the aforementioned topology path, the system performs isolation and marking to prevent cross-business abnormal data from being mixed into the current diagnostic process.

[0091] After completing path filtering, the system performs time alignment and format standardization on the retained data, and adds topological location codes to the data of each node, including node level identifier, number of hops from the source, and upstream and downstream directional attributes, forming a spatial structure feature representation. Finally, the system uniformly encapsulates the image data, network operation parameters, text information and their corresponding topological location codes after topological constraint filtering, and constructs a multimodal data unit for a single diagnostic event, providing a structural semantic input basis for subsequent cross-modal causal modeling and root cause inference.

[0092] After constructing the multimodal data unit, the system needs to further establish the correlation between business anomalies and network operation status. This step uses a cross-modal correlation modeling method to convert data from different modalities into joint causal features with a unified semantic space. This invention creatively selects and combines advanced visual coding models (such as Vision Transformer), text semantic coding models (such as BERT), and network parameter temporal models, and designs a dynamic topology weight adaptive calculation mechanism to achieve structurally adaptive feature fusion.

[0093] First, visual anomaly features are extracted from the video image data in the multimodal data unit. Let the image sequence obtained from video monitoring be:

[0094] ;

[0095] The system extracts the depth feature vector of each frame of image through a visual coding model:

[0096] ;

[0097] in, This represents a visual feature extraction function (e.g., a pre-trained Vision Transformer model). To quantify the degree of video service degradation, the system calculates inter-frame structural differences and block distortion intensity based on the depth features of consecutive frames using a temporal aggregation function, where the dimension of the image feature vector for each frame is specified. (For example, temporal convolution or Transformer encoder) to obtain business anomaly feature vectors:

[0098] ;

[0099] in, Indicates the density of the mosaic tiles. Indicates the duration of the still frame. Indicates the frame rate of change. The dimension of the aggregated business anomaly feature vector. Used to quantify the degree of degradation in video services;

[0100] Subsequently, state modeling is performed on the operating parameters of the network devices. Let the time series of parameters of the k-th network node on the topology path be:

[0101] ;

[0102] in The sampling time within the time window;

[0103] The system calculates the gradient (first-order difference) of parameter changes and the anomaly offset, and constructs the network parameter state feature vector:

[0104] (Gradient features);

[0105] (Offset features);

[0106] in, The average value under historical normal conditions is used to reflect the trend of link quality degradation. Gradient features and offset features are concatenated to form a node. Parameter state feature vector For all path nodes The parameter features are aggregated to obtain the global network parameter state features. Aggregation methods can employ averaging, max pooling, or attention mechanisms;

[0107] For text-based modal data, the system inputs the user's fault report description and device alarm logs into the text semantic encoding function:

[0108] ;

[0109] in, This refers to text semantic encoding models (such as BERT). Given the dimension of the text semantic feature vector, output the fault semantic feature vector to characterize the abnormal phenomena observed by the human side;

[0110] After obtaining the three types of features, the system introduces a topology weight matrix based on the determined network topology positional relationships. This matrix is ​​not statically assigned, but dynamically calculated using a graph attention network (GAT) based on real-time network status (such as node load, historical failure frequency, and link bandwidth utilization) and service flow path information. This matrix is ​​used to describe the degree of influence of different network nodes in the service propagation path, with upstream core nodes having a higher weight than end access nodes.

[0111] To demonstrate the causal relationship of fault propagation, the system calculates the correlation strength between visual anomalies and the parameters of each network node, using the Pearson correlation coefficient. gradient features of each node The correlation (if necessary) (After dimensionality reduction or interpolation alignment), the correlation coefficient vector is obtained:

[0112] ;

[0113] When the correlation strength exceeds a preset threshold, it is determined that the node may be involved in the fault formation process, and its topological location is encoded and added to the joint feature representation.

[0114] Finally, a joint causal feature vector is generated:

[0115] ;

[0116] This feature vector not only reflects the statistical characteristics of multimodal data, but also describes the correlation path that leads to service anomalies caused by changes in network state, thus providing an input basis that conforms to the network operation mechanism for subsequent root cause inference of faults.

[0117] In obtaining the joint causal eigenvector Subsequently, the system performs initial root cause inference on the current diagnostic event based on a preset set of root causes of faults. The set of root causes of faults is constructed based on actual cable television network operation experience and covers a variety of typical fault types, such as core source anomalies, sudden increase in backbone link attenuation, optical amplifier gain anomalies, splitter port degradation, and user-side line problems. It is also structured and classified according to network hierarchy.

[0118] Specifically, the system first divides the candidate root causes into different hierarchical subsets according to the topology level, such as the core layer fault set, the transport layer fault set, and the access layer fault set. This hierarchical division is consistent with the generated topology sub-paths, so that the root cause inference process matches the service propagation structure.

[0119] Subsequently, the joint causal feature vectors will be... Input the root cause inference function:

[0120] ;

[0121] in, Indicates the i-th type of candidate root cause. This represents a discriminant function constructed for different root cause types, which comprehensively considers visual anomaly features, network parameter variation features, text semantic features, and their relationship with the topological weight matrix;

[0122] Unlike traditional planar classification models, this invention calculates... When introducing topological hierarchy constraint coefficients This is used to limit root cause categories that do not conform to the current anomaly distribution range. For example, when anomalous nodes are concentrated in the middle of the transmission link, the weight of the access layer fault category will be automatically decayed. The specific calculation form is as follows:

[0123] ;

[0124] in, This represents the nonlinear mapping result of joint causal features (e.g., via a multilayer perceptron). This represents the matching coefficient between the root cause category and the current anomalous topological distribution. and These are trainable parameters;

[0125] Through the above calculations, the system obtains the initial root cause probability distribution vector:

[0126] ;

[0127] This probability distribution reflects the degree of matching between each candidate fault type and the current multimodal anomaly features before propagation consistency verification. Since the inference process has introduced topological hierarchical constraints and causal feature structures, it can more accurately narrow the fault location range and reduce candidate root causes that do not conform to the network propagation law compared with traditional classification methods based on single-modality or simple fusion features.

[0128] After obtaining the initial root cause probability, the system proceeds to the next step, where physical constraints are applied to the probability results through fault propagation consistency analysis to further improve positioning accuracy.

[0129] After obtaining the initial root cause probability distribution, the system further combines the network topology and the actual distribution of abnormal nodes to verify the propagation consistency of candidate root causes, thereby correcting their probability values ​​so that the final diagnosis results conform to the physical propagation laws of communication networks.

[0130] First, for each candidate root cause Based on a pre-stored network topology database and fault knowledge rules, the system determines the set of nodes that may theoretically be affected under the condition that the root cause is true, denoted as: ;

[0131] This set is derived based on the transmission direction of service signals and node dependencies. For example, when a candidate root is affected by an abnormal attenuation of a certain backbone link, its theoretical impact range should cover all access nodes downstream of that link.

[0132] Subsequently, based on the obtained abnormal node identification results, the system determines the set of nodes in the current diagnostic event that actually detected abnormal features, denoted as: ;

[0133] To quantify the consistency between the theoretical propagation range and the actual anomaly distribution, the system calculates the propagation matching coefficient. Considering that the theoretical set may be very large, while the actual anomalies may only be a subset, the Jaccard similarity coefficient is used to measure the degree of overlap.

[0134] ;

[0135] in, The coefficient represents the number of nodes in the set and is between [0,1]. It takes into account the effects of missed and false alarms. When the propagation matching degree of a candidate root cause is high, it means that the current anomaly distribution matches the propagation characteristics of the root cause in the network; conversely, when the matching degree is low, it means that the root cause is difficult to explain the current anomaly pattern.

[0136] To avoid the impact of local anomalies or sampling errors on the results, the system can introduce a propagation attenuation factor. Weight attenuation is applied to theoretical nodes that are far from the abnormal concentration area, so that the propagation matching calculation is more in line with the actual link attenuation characteristics.

[0137] Based on this, the system obtains the initial probability Perform physical consistency correction and calculate the corrected root cause probability:

[0138] ;

[0139] Through the above normalization process, the final root cause probability distribution that satisfies the network topology propagation constraints is obtained;

[0140] Through a probability correction mechanism, the integration of "data-driven inference" and "physical law-constrained inference" is achieved, so that the final fault location result not only meets the multimodal feature matching logic, but also conforms to the network signal propagation mechanism, effectively reducing the risk of misjudgment and improving the location accuracy of complex chain faults.

[0141] After completing the root cause probability correction and determining the final diagnostic result, the system uses the corrected root cause probability distribution. The fault type with the highest probability is selected as the final root cause of this diagnostic event, and its confidence level and impact range information are output simultaneously. In order to enhance the operability of operation and maintenance decisions, the system combines the topology sub-path information generated in step S2 to automatically locate the specific network node or link location that may be abnormal, and generate structured diagnostic result data containing description of abnormal phenomena, changes in key parameters, a list of affected nodes and suggested handling measures.

[0142] When generating diagnostic results, the system organizes the output information according to the network hierarchy, enabling operation and maintenance personnel to clearly identify the physical location of the anomaly and the scope of business impact. At the same time, the multimodal data units, joint causal feature vectors, and final root cause results involved in this diagnosis are archived and stored to form a traceable fault case data record.

[0143] After the fault handling is completed, the maintenance personnel submit the confirmed root cause type and handling result through the work order system or a dedicated feedback interface. The system compares the confirmed result with the original diagnostic result. If there is a difference, a misjudged sample is recorded; if they match, it is recorded as a correct sample. The above samples are added to the training dataset to update the causal association modeling parameters and the root cause discrimination function parameters.

[0144] The system adopts a periodic incremental training mechanism to further optimize the model parameters. The updated training set is the union of the historical sample set and the newly confirmed sample set. By minimizing the joint loss function again, the model can adaptively adjust to new failure modes or network structure changes.

[0145] In addition, when the network topology is adjusted (such as adding nodes or changing links), the system updates the topology database synchronously and automatically adapts to the new network structure in subsequent propagation consistency calculations, thereby ensuring that the diagnostic logic is consistent with the actual network state.

[0146] Through the above feedback and update mechanism, this invention realizes a complete technical closed loop from anomaly perception and intelligent inference to experience accumulation and model evolution, enabling the system to continuously improve the accuracy of fault location and adapt to changes in network scale and business complexity.

[0147] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for locating the root cause of multimodal faults in a cable television network, characterized in that, Includes the following steps: Step S1: Collect multi-source operation and maintenance data of the cable television network, generate diagnostic tasks according to preset diagnostic trigger conditions, and lock the time interval and service flow identifier corresponding to the diagnostic event; Step S2: Based on the service flow identifier, locate the actual signal transmission path of the service flow in the network topology database, filter out the set of topology nodes located on the path, and perform time alignment and topology location encoding on the multi-source operation and maintenance data according to the diagnostic time window to construct a multimodal data unit with structural semantics. Step S3: Perform cross-modal feature extraction on video image data, network device operating parameter data and text data in the multimodal data unit, and construct a joint causal feature vector by combining the topology weight matrix and causal correlation analysis; Step S4: Input the joint causal feature vector into the preset root cause discrimination model, and calculate the initial probability distribution of each candidate root cause by combining the topological hierarchical constraint coefficients. Step S5: Based on the network topology and the actual distribution of detected abnormal nodes, calculate the propagation consistency matching degree between the theoretical influence range and the actual abnormal range of each candidate root cause, and correct the initial probability by physical law constraints to obtain the final root cause probability distribution. Step S6: Determine the diagnostic results based on the final root cause probability distribution and output structured fault information; at the same time, perform incremental training and update of model parameters based on operation and maintenance feedback results.

2. The method for locating the root cause of multimodal faults in a cable television network according to claim 1, characterized in that, In step S2, constructing a multimodal data unit with structural semantics specifically involves: Based on the service flow identifier, locate the actual transmission path of the service in the topology database and generate a set of topology nodes. ; based on the time of the diagnostic event Set a time window for the center The collected multi-source operation and maintenance data set Execute structured filtering rules: ； in, Representing data The topology node to which it belongs. The timestamp represents the data; After filtering, data located on the business transmission path and within the time window are retained, and topological location codes are added to the data of each node, including node level identifier, number of hops from the source and upstream and downstream direction attributes, to form a spatial structure feature representation, which is finally encapsulated into a multimodal data unit.

3. The method for locating the root cause of multimodal faults in a cable television network according to claim 1, characterized in that, Step S3 is as follows: For video image sequences Extracting deep feature vectors ,in For visual encoding models; Through time series aggregation functions Calculate inter-frame structural differences and block distortion intensity to generate service anomaly feature vectors. ; For network nodes Parameters of time series Calculate gradient features and offset features And concatenate them to form a parameter state feature vector. For all path nodes The parameter features are aggregated to obtain the global network parameter state features. ; Text data is input into a text semantic encoding model The fault semantic feature vector is obtained. ; Set the dynamic topology weight matrix ; Calculate the feature vector of business anomalies Gradient features of each node correlation coefficient When the correlation coefficient exceeds a preset threshold, the topological location code of the node is added to the joint feature representation; Constructing a joint causal feature vector: ； The joint causal feature vector integrates multimodal features and topology information to describe the causal relationship between network state changes and service anomalies.

4. The method for locating the root cause of multimodal faults in a cable television network according to claim 3, characterized in that, The topology weight matrix The dynamic calculation process uses a graph attention network. The inputs are the real-time load, historical failure frequency, link bandwidth utilization and service flow path information of each node. The output is the attention coefficient matrix between nodes, which is used to characterize the relative importance of nodes in fault propagation.

5. The method for locating the root cause of multimodal faults in a cable television network according to claim 1, characterized in that, In step S4, the initial probability distribution of each candidate root cause is calculated as follows: The candidate root causes are divided into core layer, transport layer and access layer subsets according to the topology level, so that they match the topology sub-path structure generated in step S2. Joint causal eigenvectors Input the root cause discriminant function and calculate the root cause probability: ； in, Indicates the first Candidate root causes; The discriminant function is expressed as: ； in, This is the result of a nonlinear mapping of joint causal features. The coefficient representing the degree of matching between the root cause category and the current anomalous topological distribution. and These are trainable parameters.

6. The method for locating the root cause of multimodal faults in a cable television network according to claim 1, characterized in that, Step S5 is as follows: For each candidate root cause Based on the topology database and fault knowledge rules, determine the theoretical set of affected nodes. ; Based on the abnormal nodes identified in steps S2 and S3, determine the set of nodes that actually detected abnormal features. ; Calculate the propagation matching coefficient: ； The propagation matching coefficient is used to quantify the degree of overlap between the theoretical propagation range and the actual abnormal distribution; Introducing propagation attenuation factor Weight decay is applied to nodes in the theoretically influential node set that are far from the abnormal concentration area to enhance the physical rationality of the matching calculation. For initial probability Make corrections: ； The final root cause probability distribution conforming to the laws of network physical propagation is obtained.

7. The method for locating the root cause of multimodal faults in a cable television network according to claim 6, characterized in that, The propagation attenuation factor The value range is [0,1]. Its value decreases exponentially with respect to the hop distance between the node and the anomaly concentration area. The farther the distance, the lower the weight. It is used to simulate the attenuation characteristics of the signal during transmission.

8. The method for locating the root cause of multimodal faults in a cable television network according to claim 1, characterized in that, Step S6 is as follows: According to the probability distribution of the final root cause Select the fault type with the highest probability as the final diagnostic root cause, and output its confidence level, scope of influence and location information of abnormal nodes; The diagnostic results, multimodal data units, and joint causal feature vectors are archived and stored to form traceable fault cases; If the actual root cause confirmation results received from the operation and maintenance personnel are inconsistent with the diagnosis results, the sample will be marked as a misjudged sample and added to the training set. A periodic incremental training mechanism is used to update the parameters of the feature extraction model in step S3 and the root cause discrimination model in step S4, so that the model can adapt to new failure modes and network topology changes.

9. A multimodal fault root cause localization system for a cable television network, used to implement the multimodal fault root cause localization method for a cable television network as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to collect multi-source operation and maintenance data and generate diagnostic tasks; The topology constraint processing module is used to filter topology nodes based on business flow paths and construct multimodal data units. The feature modeling module is used to extract multimodal features and construct joint causal feature vectors; The root cause inference module is used to calculate the initial root cause probability and perform propagation consistency correction. The diagnostic output and self-evolution module is used to output diagnostic results and update model parameters based on feedback.

10. A multimodal fault root cause localization system for a cable television network according to claim 9, characterized in that, The topology constraint processing module also includes a topology weight dynamic calculation unit. This unit calculates the topology weight matrix between nodes based on the real-time network status and business flow path information through a graph attention network, which is used to guide the causal correlation analysis in the feature modeling process.

Images