IPTV service fault diagnosis method and device
By introducing cost-sensitive Gini splitting criteria and expert rules into IPTV service fault diagnosis, and combining bivariate splitting and GNN to generate topology vectors, the problem of the inability of existing technologies to effectively handle cost-sensitive faults is solved, achieving efficient fault identification and automated diagnosis, and improving the interpretability of the model and the accuracy of fault location.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-10
AI Technical Summary
Existing technologies cannot effectively handle cost-sensitive faults in IPTV service fault diagnosis and ignore expert knowledge in the operation and maintenance field, resulting in models that do not conform to common business sense.
We introduce cost-sensitive Gini splitting criteria and expert rules, combine them with bivariate splitting, construct decision trees, and combine them with GNN to generate topological vectors, thereby improving fault identification capabilities and model interpretability.
It improves the ability to identify high-cost faults, reduces labor costs, enables rapid fault location and automated diagnosis, and enhances the interpretability and generalization ability of the model.
Smart Images
Figure CN121842467A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of IPTV service technology, and specifically relates to an IPTV service fault diagnosis method and device. Background Technology
[0002] With the increasing number of IPTV users and the expansion of service coverage, traditional manual monitoring methods are no longer sufficient. IPTV systems typically involve multiple components, such as content providers, streaming media servers, CDNs (Content Delivery Networks), and set-top boxes. The interdependence of these components increases the complexity of operation and maintenance. To ensure service quality, improve fault handling efficiency, and reduce labor costs, automated IPTV operation and maintenance alarms have become particularly important. Furthermore, with the increasing complexity of the processes, it is necessary to accurately identify the segment where the alarm occurred and locate the fault.
[0003] Existing technologies employ decision trees for fault analysis, but traditional CART, using Gini impurity, fails to consider the cost differences in misclassification across different categories. In fault diagnosis, missed detections of certain faults can lead to significant losses, and traditional CART cannot handle such cost-sensitive scenarios. Furthermore, traditional CART, being entirely data-driven, may overlook expert knowledge in the operations and maintenance domain. For example, under certain conditions, operations personnel may know to prioritize checking certain features, but traditional CART might select other features due to randomness in the data, causing the model to contradict business common sense. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides an IPTV service fault diagnosis method and apparatus. It introduces cost sensitivity, expert rules, and bivariate splitting to make the decision tree more closely match actual business needs, improve the ability to identify high-cost faults, and enhance the interpretability of the model.
[0005] This invention provides the following technical solution: Firstly, a method for diagnosing IPTV service faults is provided, including the following steps: Collect and preprocess quality inspection data of users' IPTV services; The preprocessed quality inspection data is constructed into high-level semantic features, including network health, user experience level, and business platform load pressure. Based on the heterogeneous graph of network devices on which IPTV services depend, GNN is used to generate the topology vector of the BRAS / OLT to which the user terminal device belongs, and this vector is used as an additional feature. Based on high-level semantic features and additional features, cost-sensitive Gini is used as the splitting criterion, univariate splitting or bivariate joint splitting is used as the splitting condition, and expert rules are introduced to guide the splitting to construct a decision tree. Based on the high-level semantic features and additional features of the user's IPTV service, a decision tree is used for fault diagnosis, outputting the predicted fault type, prediction confidence, and critical split path of the user's IPTV service.
[0006] Optionally, the quality detection data includes: user information, probe data, and playback data. The user information includes: MAC address, STBID, and account. The probe data includes: source IP, destination IP, multicast address of the viewing channel, and program contentid. The playback data includes: number of buffering events, number of TS packet losses, total number of TS packets, number of buffering events, buffering duration, live streaming switching latency, on-demand loading latency, number of cache overflows, number of video interruptions, percentage of video interruptions on the service platform, and percentage of cache overflows on the terminal.
[0007] Optionally, the high-level semantic features are constructed by weighted summation, specifically: the network health is obtained by weighted summation of packet loss rate, jitter, and RTT; the user experience level is obtained by weighted summation of start-up latency, number of stutters, resolution switching frequency, and cache underflow ratio, and then classified into levels; the business platform load pressure is obtained by weighted summation of concurrent users, CDN traffic peak, and cache hit rate.
[0008] Optionally, the node types of the heterogeneous graph include: user terminal equipment, PON port, OLT, BRAS and CDN nodes; the edge types include physical connection, logical affiliation and service relationship.
[0009] Optionally, the formula for calculating the cost-sensitive Gini is: ; in, For nodes Cost-sensitive Gini The total number of categories, For nodes Medium category proportion, For nodes Medium category proportion, To classify the real categories Misclassified as a category The cost.
[0010] Optionally, the decision tree is constructed by using cost-sensitive Gini coefficients as the splitting criterion, taking univariate splitting or bivariate joint splitting as the splitting condition, and guiding the splitting by introducing expert rules. Specifically: Step S41: For the current node of the decision tree, obtain the cost-sensitive Gini reduction of all single-variable and bivariate combinations in the candidate feature set, where the bivariate combination includes predefined causal feature pairs and combinations of high-level semantic features and additional features; Step S42: Check whether the current node satisfies the predefined expert rules. If it does, add the features specified by the satisfied expert rules to the candidate feature set. Step S43: Select the optimal splitting condition based on the cost-sensitive Gini decrease, and split the current node to generate child nodes; Step S44: Repeat steps S41-S43 for the split child nodes until the termination condition is met to form a leaf node, and output the decision tree.
[0011] Optionally, when using decision trees for fault diagnosis, the cumulative reduction in cost-sensitive Gini values of variables in the decision tree is also recorded, and these reductions are sorted in reverse order as the variable importance output by the decision tree. Based on variable importance, splitting conditions, and business experience, alarm rules and thresholds are set, and alarm rules are applied to real-time quality detection data and high-level semantic features to determine whether all alarm conditions are met. If they are met, an alarm is generated. The alarm conditions include: whether the variable value exceeds the threshold, whether unnecessary alarms are filtered according to preset filtering rules, and whether multiple related alarms that occur at the same time are merged.
[0012] Secondly, an IPTV service fault diagnosis device is provided, comprising: The data acquisition module is used to collect quality inspection data of users' IPTV services and perform preprocessing. The high-level semantic feature acquisition module is used to construct high-level semantic features from the preprocessed quality detection data, including network health, user experience level, and business platform load pressure. The additional feature acquisition module is used to generate the topology vector of the BRAS / OLT to which the user terminal equipment belongs based on the heterogeneous graph of the network devices on which IPTV services depend, using GNN, and then use it as an additional feature. The decision tree construction module is used to construct decision trees based on high-level semantic features and additional features, using cost-sensitive Gini as the splitting criterion, taking univariate splitting or bivariate joint splitting as the splitting condition, and introducing expert rules to guide the splitting. The fault diagnosis module is used to perform fault diagnosis using decision trees based on the high-level semantic features and additional features of the user's IPTV service, and outputs the predicted fault type, prediction confidence, and critical split path for the user's IPTV service. Thirdly, a computer device is provided, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the steps of the IPTV service fault diagnosis method according to any one of the first aspects.
[0013] Fourthly, a computer-readable storage medium is provided for storing a computer program; when the computer program is executed by a processor, it implements the steps of the IPTV service fault diagnosis method described in any one of the first aspects.
[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention automates IPTV service data acquisition, processing, and fault prediction, proactively identifying service issues, enabling rapid problem localization, and reducing manual costs. Furthermore, it uses a decision tree method to determine service status while considering multiple features, achieving automatic threshold adjustment, making it more effective in handling complex problems. It can also capture more complex data changes and adjust accordingly. This invention maps raw quality inspection data into high-level semantic indicators through business rules, improving generalization ability and interpretability. This invention employs a fusion architecture of GNN and decision trees. GNN is used to predict the correlation between device topologies, while decision trees are used for highly interpretable causal reasoning, improving root cause localization accuracy. The decision tree in this invention adds cost-sensitive Gini computation, expert rule guidance, and bivariate joint splitting, achieving rule-driven and data-driven approaches, increasing the protection priority of high-value services, and better aligning with actual business needs. Attached Figure Description
[0015] Figure 1 This is a flowchart of the IPTV service fault diagnosis method of the present invention; Figure 2 This is a flowchart of the alarm generation process of the present invention; Figure 3 This is a flowchart of the alarm processing of the present invention; Figure 4 This is a specific example of a decision tree provided in this invention. Detailed Implementation
[0016] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be used to limit the scope of protection of the present invention. It should be noted that the term "comprising" and any variations thereof in the specification, claims and the above-mentioned drawings of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or devices.
[0017] Example 1 like Figure 1 As shown, an IPTV service fault diagnosis method is provided, including the following steps: S1: Collect quality inspection data of users' IPTV services and perform preprocessing; S2: Construct high-level semantic features from the preprocessed quality inspection data, including network health, user experience level, and business platform load pressure; S3: Based on the heterogeneous graph of network devices on which IPTV services depend, use GNN to generate the topology vector of the BRAS / OLT to which the user terminal device belongs, and use it as an additional feature; S4: Based on high-level semantic features and additional features, cost-sensitive Gini is used as the splitting criterion, univariate splitting or bivariate joint splitting is used as the splitting condition, and expert rules are introduced to guide the splitting to construct a decision tree; S5: Based on the high-level semantic features and additional features of the user's IPTV service, use a decision tree to perform fault diagnosis and output the predicted fault type, prediction confidence, and critical split path of the user's IPTV service.
[0018] In this embodiment, the quality detection data in step S1 includes: user information, probe data, and playback data. The method of obtaining the quality detection data refers to the prior art.
[0019] User information includes: MAC address, STBID, and account.
[0020] The probe data includes: source IP, destination IP, multicast address of the channel being viewed, and program contentid.
[0021] Playback data includes: number of buffer drops, number of lost TS packets, total number of TS packets, number of buffer drops, buffer duration, live stream switching latency, on-demand loading latency, number of cache overflows, number of video stream interruptions, percentage of video stream interruptions on the service platform, and percentage of cache overflows on the terminal. TS stands for Transport Stream in the IPTV / video field.
[0022] Preprocessing typically includes cleaning the data, handling outliers and missing values, scaling necessary feature fields, and generating pre-cleaned results.
[0023] In some other embodiments, if the user's viewing type is video-on-demand, the source IP in the probe data is associated with the CDN node server IP in the CDN node resource data, and the associated CDN node identifier is added to the probe information. If the user's viewing type is multicast, the multicast address in the probe data is associated with the live channel metadata data, and the associated channel name is added to the probe information.
[0024] In step S2, the high-level semantic features are constructed by weighted summation.
[0025] In this embodiment, network health is scored by weighted summation of packet loss rate, jitter, and RTT (round-trip time).
[0026] User experience levels are determined by weighting and summing factors such as playback latency, number of buffering incidents, resolution switching frequency, and cache underflow rate. Specifically, it can be divided into 5 levels (0-4).
[0027] The load pressure of the business platform is scored by weighted summation of concurrent users, CDN traffic peak and cache hit rate.
[0028] In this embodiment, it is also necessary to set the time window length for multi-feature data aggregation. Specifically, the time window length is set, the data boundaries are divided according to the specified time window length, the data of different time windows are grouped and calculated, key features related to business quality are selected from the dataset, and the data is aggregated according to different feature dimensions of the standardized wide table data within each time window to generate high-level semantic features.
[0029] A specific example is given, where each indicator is normalized according to the business tolerance threshold (such as packet loss rate 1%, jitter 30ms, RTT 100ms), and then weighted and merged according to the impact weight (typical weight: packet loss rate 0.5, jitter 0.3, RTT 0.2); the output range is [0,1], and the higher the value, the healthier the network link.
[0030] The node types in the heterogeneous graph in step S3 include: user terminal equipment, PON ports, OLTs, BRASs, and CDN nodes; the edge types include physical connections, logical affiliations, and service relationships. IPTV services rely on complex network topologies (user-PON-OLT-BRAS), and traditional flattened features are insufficient to capture cross-device dependencies.
[0031] Step S3 uses a graph neural network (GNN) to embed and encode the topology devices. Specifically, GraphSAGE is used to encode the entire network topology end-to-end; a fixed-dimensional topology context vector is generated for each device node; the vector implicitly contains information such as the status of neighboring devices, path redundancy, and load propagation effects; the topology vector of the user's BRAS / OLT is used as an additional feature and concatenated into the user's aggregate record, thereby enabling this application to perceive the impact of abnormal neighboring devices on the current user.
[0032] Traditional CART uses Gini impurity as the splitting criterion, while this solution uses cost-sensitive Gini, as the cost of misjudging different fault types varies greatly. For example, misjudging a CCTV live broadcast interruption as normal has far greater consequences than missing a report of a local channel's buffering. Therefore, cost-sensitive Gini is adopted as the splitting criterion.
[0033] The formula for calculating the cost-sensitive Gini is: ; in, For nodes Cost-sensitive Gini The total number of categories, For nodes Medium category proportion, For nodes Medium category proportion, To classify the real categories Misclassified as a category The cost.
[0034] In step S4 of this embodiment, the introduction of expert rules to guide splitting specifically involves: a predefined rule base; before each split, checking whether the current node satisfies any rule; if it does, the candidate feature set consists of the rule-specified feature and the original high-gain feature; if not, the original high-gain feature is used, and a rule matching reward is added to the splitting score to increase the priority of relevant features. Purely data-driven approaches may overlook common operational knowledge. For example, when the user experience is extremely poor (UEL≤1), CDN or terminal issues must be investigated first, rather than network links, to better align with actual business needs.
[0035] In step S4 of this embodiment, the bivariate joint splitting as a splitting condition specifically means: IPTV failures are often caused by multi-indicator causal chains. Strong causal feature pairs are pre-selected through Granger causality testing. In the splitting candidate set, in addition to single variables, a new logical combination condition is added, for example: if (packet loss rate > X) AND (cache underflow > Y), if (CDN_CPU > A) OR (cache hit rate). <B)。
[0036] In this embodiment, step S4 specifically includes the following sub-steps: Step S41: For the current node of the decision tree, obtain the cost-sensitive Gini reduction of all univariate and bivariate combinations in the candidate feature set. The bivariate combinations include predefined causal feature pairs and combinations of high-level semantic features and additional features.
[0037] Of course, before step S41, an initialization operation is required, namely setting the maximum depth (default 10), the minimum number of leaf samples (5), and the Gini descent threshold (0.001).
[0038] Step S42: Check whether the current node satisfies the predefined expert rules. If it does, add the features specified by the satisfied expert rules to the candidate feature set.
[0039] Step S43: Select the optimal splitting condition based on the cost-sensitive Gini decrease and split the current node to generate child nodes.
[0040] Step S44: Repeat steps S41-S43 for the split child nodes until the termination condition is met to form a leaf node, and output the decision tree.
[0041] Step S4 involves recursively splitting the aggregated multidimensional feature dataset (as shown in Table 1) starting from the root node. For each node: a. Calculate the Cost-Gini (cost-sensitive Gini) decrease for all univariate and predefined bivariate combinations; b. Check if expert rules are triggered and dynamically adjust the candidate set; c. Select the optimal splitting condition; d. Apply adaptive depth control (high-priority channel + 2 layers). The splitting terminates when: Cost-Gini decrease < threshold, number of samples < minimum leaf samples, or the dynamic maximum depth is reached. This outputs a business-aware decision tree.
[0042] Table 1: Specific examples of multidimensional feature datasets
[0043] Step S5 involves anomaly prediction and result output, specifically: high-level semantic features and additional features constructed based on quality detection data from the new historical time window (relative to decision tree training) are input into the trained decision tree; traversing from the root node to the leaf node, all splitting conditions on the path are recorded; the final output includes: predicted category (e.g., CDN node overload); prediction confidence: the percentage of samples of the target category in the leaf node; key splitting paths: used to generate readable root cause explanations (e.g., "Due to CDN_CPU>90% and cache hit rate<50%, it is determined to be CDN overload").
[0044] In some other embodiments, when using decision trees for fault diagnosis, the cumulative decrease in cost-sensitive Gini values of variables in the decision tree is also recorded, and these decreases are sorted in reverse order as the variable importance output by the decision tree.
[0045] like Figure 2As shown, alarm rules and thresholds are set based on variable importance, splitting conditions, and business experience. Alarm rules are applied to real-time quality detection data and high-level semantic features to determine whether all alarm conditions are met. If they are met, an alarm is generated; otherwise, no alarm is generated. Alarm conditions include: whether the variable value exceeds the threshold, whether unnecessary alarms are filtered according to preset filtering rules, and whether multiple related alarms occurring simultaneously are merged.
[0046] In other embodiments, alarm information is also categorized, such as Figure 3 As shown, specifically, alarm information is categorized into new alarms, persistent alarms, and recovering alarms based on their frequency of occurrence. Alarms generated in the current time window are combined with alarms in the persistent alarm table using a set operation. If a monitored alarm does not exist in the persistent alarm table, it is considered a new alarm; if a persistent alarm does not exist in the monitored alarm table, it is considered a recovering alarm; their intersection is a persistent alarm. Based on the feature importance generated by the decision tree and the alarm classification, alarms are prioritized, and the final alarm result is generated by combining feature summary information, associating lower-importance alarms with the primary alarms.
[0047] This embodiment provides an example of a decision tree, such as... Figure 4 As shown: Example 1: First, check if the CDN traffic is zero. If confirmed, further check if the UEL (User Edge Load) is less than or equal to 1. Assuming the condition is met, next determine if the channel priority (channel_priority) is the highest level (value 1). If so, further verify if the start-up latency is greater than 10 seconds. If this condition is also met, then evaluate if the NHS (Network Health Score) is greater than or equal to 0.95. In this case, the NHS (Network Health Score) meets the condition, and then perform a topo_sim_bras evaluation, i.e., whether the similarity between the current BRAS topology vector and the normal mode is greater than or equal to 0.85. A high topo_sim_bras score indicates that there are no anomalies on the network side, and the problem is ultimately determined to be a program source failure.
[0048] Example 2: Start by checking if CDN traffic is zero. After confirmation, if the UEL (User Experience Level) is less than or equal to 1, since the channel priority is not the highest, the process shifts to determining if the NHS is less than 0.6. If the NHS is indeed below 0.6, and the anomaly propagation score of the downstream OLT (neighbor_anomaly_olt) exceeds 0.7, then the system will diagnose that the problem lies in an anomaly at the OLT / PON layer, rather than a problem with the BRAS itself.
[0049] Example 3: Initially, check if the CDN traffic is not zero. Then evaluate if the UEL is less than or equal to 1. If the condition is met, check if the cache hit rate is lower than 0.5. If so, further check if the PLP (Platform Load Pressure) is higher than 0.8. Assuming this condition is also met, finally evaluate topo_sim_bras to confirm if it is greater than or equal to 0.85. Since the topo_sim_bras score is high, it indicates that there is no problem with the network, thus concluding that the problem is caused by CDN node overload.
[0050] Example 2 An IPTV service fault diagnosis device, comprising: The data acquisition module is used to collect quality inspection data of users' IPTV services and perform preprocessing. The high-level semantic feature acquisition module is used to construct high-level semantic features from the preprocessed quality detection data, including network health, user experience level, and business platform load pressure. The additional feature acquisition module is used to generate the topology vector of the BRAS / OLT to which the user terminal equipment belongs based on the heterogeneous graph of the network devices on which IPTV services depend, using GNN, and then use it as an additional feature. The decision tree construction module is used to construct decision trees based on high-level semantic features and additional features, using cost-sensitive Gini as the splitting criterion, taking univariate splitting or bivariate joint splitting as the splitting condition, and introducing expert rules to guide the splitting. The fault diagnosis module is used to perform fault diagnosis using decision trees based on the high-level semantic features and additional features of the user's IPTV service, and outputs the predicted fault type, prediction confidence and critical split path of the user's IPTV service.
[0051] For more detailed information on the above modules, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.
[0052] Example 3 The present invention provides a computer device, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the steps of the above-described IPTV service fault diagnosis method.
[0053] For more detailed information on the above methods, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.
[0054] Example 4 The present invention provides a computer-readable storage medium for storing a computer program; when the computer program is executed by a processor, it implements the steps of the above-described IPTV service fault diagnosis method.
[0055] For more detailed information on the above methods, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.
[0056] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatuses, devices, and storage media disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to in the method section.
[0057] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.
[0058] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.
Claims
1. An IPTV service fault diagnosis method, characterized in that, The method comprises the following steps: Collecting quality detection data of user IPTV service and pre-processing the data; Constructing high-level semantic features from the pre-processed quality detection data, including network health, user experience level and service platform load pressure; Based on the heterogeneous graph of network equipment dependent on IPTV service, using GNN to generate the topology vector of BRAS / OLT to which the user terminal equipment belongs, and taking it as an additional feature; Based on the high-level semantic features and the additional features, using cost-sensitive Gini as the splitting criterion, using single-variable splitting or double-variable joint splitting as the splitting condition, and guiding the splitting by introducing expert rules to construct a decision tree; Using the decision tree to perform fault diagnosis according to the high-level semantic features and the additional features of the user IPTV service, and outputting the predicted fault type, the prediction confidence and the key splitting path of the user IPTV service.
2. The IPTV service fault diagnosis method of claim 1, wherein, The quality detection data includes user information, probe data and playback data, the user information includes MAC address, STBID and account number, the probe data includes source IP, destination IP, viewing channel multicast address and program content ID, and the playback data includes freezing frequency, TS packet loss number, TS total packet number, freezing frequency, freezing duration, live switching delay, on-demand loading delay, buffer underflow frequency, video flow interruption frequency, service platform video flow interruption frequency ratio and terminal buffer overflow ratio.
3. The IPTV service fault diagnosis method of claim 1, wherein, The high-level semantic features are constructed by weighted summation, specifically: the network health is scored by weighted summation of packet loss rate, jitter and RTT; the user experience level is scored by weighted summation of start-up delay, freezing frequency, resolution switching frequency and buffer overflow ratio, and then graded to obtain; and the service platform load pressure is scored by weighted summation of concurrent user number, CDN traffic peak value and cache hit rate.
4. The IPTV service fault diagnosis method of claim 1, wherein, The node types of the heterogeneous graph include user terminal equipment, PON port, OLT, BRAS and CDN node; and the edge types include physical connection, logical attribution and service relationship.
5. The IPTV service fault diagnosis method of claim 1, wherein, The calculation formula of the cost-sensitive Gini is: ； where, is the cost-sensitive Gini for node , is the total number of classes, is the proportion of class in node , is the proportion of class in node , is the cost of misclassifying true class as class .
6. The IPTV service fault diagnosis method of claim 1, wherein, The decision tree is constructed by using cost-sensitive Gini as the splitting criterion, using single-variable splitting or double-variable joint splitting as the splitting condition, and guiding the splitting by introducing expert rules, specifically: Step S41: For the current node of the decision tree, obtain the cost-sensitive Gini reduction of all single variables and double variable combinations in the candidate feature set, wherein the double variable combinations include predefined causal feature pairs and combinations of high-level semantic features and additional features; Step S42: Check whether the current node meets the pre-defined expert rules, if yes, add the features specified by the met expert rules to the candidate feature set; Step S43: Select the optimal splitting condition according to the cost-sensitive Gini reduction, and split the current node to generate child nodes; Step S44: Repeat steps S41-S43 for the split child nodes until the termination condition is met to form leaf nodes, and output the decision tree.
7. The IPTV service fault diagnosis method of claim 1, wherein, In the fault diagnosis using the decision tree, the cost-sensitive Gini reduction of the variables accumulated in the decision tree is also recorded, and the variable importance is arranged in reverse order and output as the decision tree. Based on the variable importance, the split condition and the business experience, the alarm rules and thresholds are set, and the alarm rules are applied to the real-time detected quality detection data and high-level semantic features to determine whether all alarm conditions are met. If so, an alarm is generated. The alarm conditions include whether the variable value exceeds the threshold, whether unnecessary alarms are filtered according to the preset filtering rules, and whether multiple related alarms occurring at the same time are merged.
8. An IPTV service failure diagnosis apparatus, characterized by comprising: The method comprises: a data acquisition module for collecting quality detection data of user IPTV services and performing preprocessing; a high-level semantic feature acquisition module for constructing the preprocessed quality detection data into high-level semantic features, including network health, user experience level and service platform load pressure; an additional feature acquisition module for generating a topology vector of a BRAS / OLT to which a user terminal device belongs based on a heterogeneous graph of network equipment relied on by the IPTV service using GNN, and taking the topology vector as an additional feature; a decision tree construction module for constructing a decision tree based on the high-level semantic features and the additional features, using cost-sensitive Gini as a split criterion, taking single-variable split or double-variable joint split as a split condition, and guiding the split by introducing expert rules; a fault diagnosis module for using the decision tree to perform fault diagnosis according to the high-level semantic features and the additional features of the user IPTV service, and outputting the predicted fault type, the predicted confidence and the key split path of the user IPTV service.
9. A computer device, comprising: A processor and a memory; wherein the processor executes the computer program saved in the memory to realize the steps of the IPTV service fault diagnosis method in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, A computer program for storing; the computer program is executed by the processor to realize the steps of the IPTV service fault diagnosis method in any one of claims 1-7.