A broadband network one-key intelligent diagnosis method and system based on user perception data

Abstract

The application provides a broadband network one-key intelligent diagnosis method and system based on user perception data, and the overall technical path is developed around six links of "data fusion, feature modeling, intelligent reasoning, root cause mining, suggestion generation and closed-loop tracking", a fusion diagnosis system for poor quality users, complaint users and low satisfaction users is constructed, the method passes through network operation data, user business perception data, Wi-Fi device data, terminal identification data, fault reporting data and other multi-source heterogeneous information by constructing a broadband fusion diagnosis intelligent agent, and constructs a user network perception state atlas; combined with machine learning technologies such as time series modeling, event clustering, causal reasoning and feature enhancement, real-time portrait construction, root cause tracing and closed-loop optimization suggestion output of key users are realized, so as to effectively support customer service personnel and smart home engineers to quickly locate and accurately dispose in the user complaint, poor quality and low satisfaction scenarios, and improve the overall network service quality and customer satisfaction.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, communication networks, and operation and maintenance, specifically a one-click intelligent diagnostic method and system that integrates multi-source sensing data, knowledge graphs, and machine learning algorithms to perform network quality perception, fault root cause analysis, and optimization suggestion generation for home broadband environments. This method is applicable to various application scenarios, including telecom operators, intelligent customer service platforms, home network management systems, and remote / on-site support from installation and maintenance engineers. Background Technology

[0002] With the continuous improvement of broadband access speeds and the widespread adoption of smart home devices, users' subjective requirements for home network quality are increasing, and user experience in network services has become a crucial component of operators' core competitiveness. Currently, most broadband operation and maintenance systems still rely on the "account resource tree" as their core, employing a link-driven analysis model. This involves sequentially checking physical links such as "account → ONU → splitter → OLT → BAS," depending on static indicator alarms and manual rule-based judgment for fault diagnosis. However, this approach suffers from limitations such as coarse diagnostic granularity, a single reasoning path, and untimely response when dealing with multi-user group faults, dynamic changes in network topology, and unstructured fault reporting scenarios. It struggles to meet the efficient and intelligent diagnostic needs of modern complex home network scenarios.

[0003] Existing home broadband diagnostic solutions mainly focus on the following three areas: Link status driven methods, such as CN106936632A, determine the scope of the fault by monitoring the gateway link status, but they only rely on local device anomaly reporting and static link status judgment, lacking an understanding of the user's actual perception experience and multi-terminal collaborative environment. Adjacency comparison type reverse diagnosis method: such as CN108289031A, which proposes to identify the fault range by comparing the operation indicators of the target user and the neighboring users. Although it improves the positioning accuracy, it relies on a fixed topology and device-level performance indicators and cannot identify problems such as Wi-Fi interference and terminal capabilities. Business data analysis solutions, such as CN110943874A, introduce home gateway logs and business performance data, and identify anomalies through rule matching and event templates. Although they have a certain degree of intelligence, they still rely on static rule bases and preset logic, and lack dynamic learning and closed-loop iteration capabilities. Resource access fault identification mechanisms, such as CN116962242A, focus on detecting abnormal web page access and automatically determine the type of problem through DNS, connectivity analysis, etc. However, the diagnostic granularity is mainly limited to the application layer and lacks systematic coverage of access layer, Wi-Fi layer and terminal side problems.

[0004] In summary, existing home broadband intelligent diagnostic technologies have the following main shortcomings: 1. Limited data dimensions and lack of integrated perception capabilities: Most solutions only focus on link status or device performance indicators, lacking the ability to integrate and process information such as user perception, terminal type, Wi-Fi channel quality, and interference, resulting in incomplete identification of the root cause of the problem.

[0005] 2. Static diagnostic logic and rigid patterns: The common approach of rule matching and indicator threshold comparison is difficult to adapt to dynamic network changes in multiple scenarios and among multiple users, and cannot proactively discover potential problems.

[0006] 3. Lack of intelligent reasoning and learning mechanisms: Although some methods introduce models or knowledge tables, their learning ability is weak, the models are updated slowly, and they cannot be continuously optimized or migrated to new environments.

[0007] 4. Lack of a closed-loop optimization path: Most current systems are limited to problem discovery and preliminary suggestions, lacking an automatic closed loop from diagnosis to optimization, which cannot support the future development trend of intelligent operation and maintenance of home networks.

[0008] 5. Segmented service targets: Most patents are designed for professional maintenance personnel, with complex interfaces and interaction mechanisms, and lack visual prompts and self-operation guidance for non-professional home users.

[0009] Therefore, there is an urgent need for a one-click intelligent diagnostic method that is designed for the complex environment of home networks, and has the capabilities of multi-dimensional data fusion, intelligent reasoning and decision-making, and closed-loop optimization. This method would support efficient processing by maintenance engineers and provide convenient operation for end users, thereby comprehensively improving the level of intelligence in broadband network maintenance. Summary of the Invention

[0010] This invention aims to overcome the core technical bottlenecks in existing home broadband network diagnostic systems, such as low accuracy in identifying perception problems, lack of root cause analysis paths, high generalization of recommendations, and fragmented diagnostic processes. It proposes a one-click intelligent diagnostic method and system that integrates multi-source user perception data, machine learning modeling, structured attribution reasoning, and knowledge graph-driven optimization mechanisms.

[0011] This invention provides a one-click intelligent diagnosis method for broadband networks based on user-perceived data. The overall technical path revolves around six major stages: data fusion, feature modeling, intelligent reasoning, root cause mining, suggestion generation, and closed-loop tracking. It constructs a fusion diagnosis system for users with poor quality, complaining users, and users with low satisfaction.

[0012] This method constructs a broadband converged diagnostic intelligent agent, integrating multi-source heterogeneous information such as network operation data, user service perception data, Wi-Fi device data, terminal identification data, and fault reporting data to build a user network perception status map. Combining machine learning techniques such as time series modeling, event clustering, causal reasoning, and feature enhancement, it enables real-time profiling of key users, root cause tracing, and closed-loop optimization suggestions. This effectively supports customer service personnel and smart home engineers in quickly locating and accurately handling user complaints, poor quality, and low satisfaction scenarios, thereby improving overall network service quality and customer satisfaction.

[0013] To achieve the above objectives, the present invention provides the following technical solution: A one-click intelligent diagnostic method for broadband networks based on user-perceived data includes the following steps: Data fusion and feature construction steps: Using user accounts as the main index, collect network link performance indicators, user behavior profile features, Wi-Fi operation status data, terminal device identification features, uplink device and access link parameters, service history and perception tag data from multiple source systems, and form a structured input vector through feature engineering processing; The steps for building a model training and inference strategy are as follows: a supervised learning method is used to train the perception scoring model, outputs the user's network perception quality score, and sets diagnostic labels based on the scoring results; Root cause analysis steps: For users whose perception scores are below a set threshold, the problem is attributed through Shapley value decomposition, causal graph reasoning, anomaly clustering and time series matching; Optimization suggestion generation and feedback steps: Based on the attribution results, match the knowledge graph to generate optimization suggestions for multiple roles, and collect feedback data to achieve closed-loop iteration between the model and the knowledge base.

[0014] Preferably, in the data fusion and feature construction step: The multi-source system includes a network management system, an optical access management platform, a home Wi-Fi status monitoring platform, a terminal identification module, and a customer service system; The feature engineering process includes data cleaning, normalization, and encoding. A sliding window method is used to count the frequency of anomalies and the rate of variation. The feature is generated by combining the user's complaint text with vector encoding of a large language model.

[0015] Preferably, in the step of constructing the model and training the inference strategy: The perception scoring model uses the XGBoost algorithm and outputs a scoring range of [0, 10]. The model training labels are derived from user satisfaction scores, complaint handling conclusions, order dispatch results, and weakly supervised information on the main complaint classification. The model is deployed on the SparkML inference service framework, supporting batch inference and real-time invocation, with a single diagnostic response latency controlled within 2 seconds.

[0016] Preferably, the root cause analysis step includes: The contribution of each feature to the scoring results is calculated based on the Shapley value, and a Top-N outlier ranking is generated. Multi-level causal chains are constructed using structural causal graphs to eliminate redundant edges and weak dependencies; KMeans and DBSCAN are used to cluster the characteristics of users with poor historical quality and extract typical abnormal combination patterns. A dynamic time warping algorithm is introduced to perform user temporal behavior similarity matching, and an analogy attribution library is constructed.

[0017] Preferably, the optimization suggestion generation and feedback step includes: Based on the attribution results, match the suggestion rules in the knowledge graph to generate visual prompts for end users and professional handling suggestions for customer service and engineers; It supports automatic work order dispatch and work order generation. When the score is below the threshold and hits the high confidence fault flag, a work order is automatically generated and dispatched. Collect feedback from users or engineers for incremental model training, knowledge graph weight updates, and path confidence adjustments.

[0018] This invention also provides a one-click intelligent diagnostic system for broadband networks based on user-perceived data, characterized in that it includes: the aforementioned one-click intelligent diagnostic method for broadband networks based on user-perceived data, further comprising: The data fusion module is used to collect and process multi-source heterogeneous data to construct a user network perception state map; The scoring model module uses an ensemble learning algorithm to output a perceptual score and provides an explanation of the feature contribution. The root cause analysis module combines structural cause-effect diagrams and diagnostic item maps to generate traceable abnormal attribution paths; The abnormal trajectory matching module uses the DTW algorithm to calculate the similarity of users' historical time-series behavior and builds an analogy attribution library; The suggestion generation and feedback module relies on the knowledge graph to achieve linkage and matching between problem tags and suggestion rules, and updates the model and graph based on feedback.

[0019] Preferably, the system is deployed on a distributed data processing platform based on Spark, Kafka, HDFS, and Kubernetes container environments; It supports three diagnostic task triggering methods: user self-triggering, customer service assistance triggering, and system-initiated batch triggering. It has minute-level diagnostic task scheduling and real-time concurrent scoring inference capabilities, with a single round of concurrent inference capability exceeding 1000 users.

[0020] Preferably, the scoring model module: XGBoost was used as the core classifier, and Grid Search and Bayesian optimization algorithms were used for parameter tuning during training. The model evaluation metrics include AUC > 0.85, MSE < 1.2, and F1-score ≥ 0.79; It supports batch inference at midnight every day and real-time calls at the minute level, with an average of over 800,000 inference calls per day.

[0021] Preferably, the root cause analysis module: Construct a diagnostic knowledge graph, which consists of anomalies, causal paths, problem labels, and suggested nodes, and store it in a graph database; It supports semantic retrieval based on path tags or anomaly combinations, providing structured input for suggestion generation.

[0022] Preferably, the suggestion generation and feedback module: The generated suggestions include visual prompts for end users, professional handling suggestions from customer service and engineers, and automated work orders; The feedback mechanism includes user clicks, which are collected by the system and used for incremental training and graph augmentation. A closed-loop path of diagnosis-optimization-learning is achieved, supporting the continuous evolution of the system under different network environments and terminal structures.

[0023] This invention provides a one-click intelligent diagnosis method and system for broadband networks based on user-perceived data, which has the following beneficial effects: In terms of data perception and fusion capabilities, the system fully integrates multi-dimensional heterogeneous data such as user network connection quality, Wi-Fi operating status, terminal identification information, user behavior profiles and historical service feedback, and constructs a unified structured feature input system, which significantly improves the coverage of the diagnostic process and the accuracy of problem location. In terms of intelligent analysis and attribution capabilities, the system uses an integrated learning model to output user perception scores and completes the structured identification of the root causes of problems through Shapley values, causal graphs, event clustering and time trajectory comparison. It can automatically classify and reuse reasoning for typical problem patterns, enhancing the interpretability and robustness of diagnosis. In terms of optimization suggestion generation and closed-loop mechanism, the system calls knowledge graph based on attribution results to generate customized suggestions for three roles: users, customer service and engineers. It also supports automatic order dispatch and user feedback collection mechanism. Feedback data is used to feed back into model training and knowledge graph enhancement to build an intelligent closed-loop path of diagnosis-optimization-learning. In summary, this invention effectively overcomes the limitations of traditional link alarm-driven methods, realizing full-link modeling, intelligent analysis, and automatic response to user-perceived problems in home broadband networks. It is applicable to various application scenarios such as complaint handling, proactive maintenance, and self-diagnosis, and has significant practical value in improving diagnostic efficiency, reducing labor costs, and enhancing user satisfaction. It also has good engineering feasibility and broad commercial prospects. Attached Figure Description

[0024] Figure 1 The overall system logic architecture of this invention clarifies the overall path and functional distribution of the multi-source sensing data flow diagnosis and optimization module; Figure 2 This invention represents the application model in actual operation and maintenance scenarios, including a business process where customer service and installation and maintenance engineers trigger one-click diagnosis through the system. Figure 3 The technical architecture of the broadband fusion diagnostic intelligent agent in this invention consists of a data processing framework, a feature modeling platform, and a model service system. Figure 4 The data flow path for model inference services in this invention specifically includes key steps such as diagnostic triggering, feature processing, model invocation, and result feedback. Figure 5 The system architecture for the user terminal identification task in this invention includes data preprocessing, nested modeling, Transformer encoding, and multi-source information fusion processes. Detailed Implementation

[0025] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. However, it should be understood that the present invention can be implemented in various forms. The exemplary and non-limiting embodiments shown in the drawings and described below are not intended to limit the invention to the specific embodiments illustrated.

[0026] The broadband network one-click intelligent diagnostic system proposed in this invention is designed around the goals of user perception modeling and automated operation and maintenance. The overall architecture consists of five core modules: The data fusion module collects and processes multi-source data such as network link status, Wi-Fi parameters, terminal capabilities, and user behavior characteristics in a unified manner. The scoring model module uses an ensemble algorithm to output a perceptual score and provides an explanation of feature contribution based on the Shapley value; The root cause analysis module combines structural cause-effect diagrams and diagnostic item maps to generate traceable abnormal attribution paths; The abnormal trajectory matching module introduces the DTW algorithm to calculate the similarity of users' historical time-series behavior and builds an analogy attribution library to improve the completeness of attribution and adaptability to complex scenarios. The suggestion generation and feedback module leverages a knowledge graph to link and match problem tags with suggestion rules, and updates the model and graph based on user actions. The system supports distributed deployment and is suitable for applications such as customer service diagnostics, self-service troubleshooting, and network quality inspection.

[0027] The system is deployed on a distributed data processing platform (based on Spark, Kafka, HDFS, and Kubernetes container environment), supports minute-level diagnostic task scheduling and real-time concurrent scoring inference, and has high throughput and low latency industrial deployment capabilities. It is suitable for various business scenarios such as user complaint response, quality inspection, and self-diagnosis in home broadband scenarios.

[0028] This invention also provides a one-click intelligent diagnosis method for broadband networks based on user-perceived data, the specific method of which is as follows: Step 1: Data Fusion and Feature Construction In the diagnostic process, the system first uses the PPPoE account as the primary index, combined with the user's city / region and home device identifier, to extract relevant user operational data from multiple sources, including the Network Management System (RMS), Optical Access Management Platform (OLT), Home Wi-Fi Status Monitoring Platform, Terminal Identification Module, and Customer Service System, constructing a cross-level, cross-device, and cross-behavioral data fusion set. The collected data characteristics mainly cover the following six categories: (1) Network link and protocol layer performance indicators: including DNS resolution latency, TCP three-way handshake latency, HTTP first byte latency, PPPoE connection establishment time, uplink and downlink actual rates, rate deviation, downlink packet loss rate, TCP connection establishment failure rate, etc., reflecting the connection quality from the application layer to the transport layer; (2) User behavior profile features: By analyzing access records, user activity, application usage preferences (video, social, web, games, etc.), terminal distribution and switching frequency, traffic consumption patterns, etc., user behavior time series are constructed to characterize user perception status and network pressure distribution. (3) Wi-Fi side operation status data: including terminal connection frequency band, actual negotiated rate (Mbps), supported protocol type (such as 802.11n / ac / ax), channel occupancy rate, SNR, RSSI, telecom channel interference index, master-slave routing structure, etc., reflecting the stability and anti-interference capability of the user's local area network communication link; (4) Terminal device identification features: Based on the terminal identification engine, nested vectorized features are constructed by integrating DHCP allocation fields, User-Agent, Vendor information, mDNS broadcast, device historical active time periods and other dimensions to help judge device capabilities (such as maximum negotiation rate, whether it supports 5GHz, whether it is a gigabit network card). (5) Parameters of upstream equipment and access link: including optical network unit receiving / transmitting optical power (dBm), PON port BIP error rate, ONU online / offline frequency, LAN port negotiation rate, OLT board temperature, BRAS port congestion, etc.; (6) Service history and perception label data: historical complaint records, user satisfaction scores, whether there is repair request, repair type, work order dispatch result, and previous diagnosis conclusion are introduced as weak label feature input for the auxiliary learning model.

[0029] All original data are subjected to data cleaning, normalization and encoding via the feature engineering module, and then structured input vectors are constructed. For data processing, a sliding window from T-7 day to T-1 day is adopted to count the abnormal frequency and variation rate of each diagnostic index, so as to form statistical features of base value category; meanwhile, in combination with user's chief complaint text, a large language model is invoked to perform vector encoding and tendency classification on the text (such as "video playback stuttering", "Wi-Fi frequent disconnection" and etc.), to generate representation features of extend value category, which is used to enhance the model's ability to recognize subjective experience problems. This process is executed in parallel by the SparkCore engine to ensure the data processing efficiency and model input consistency in the scenario of high-concurrency diagnostic tasks.

[0030] Step 2: Construct model training and reasoning strategy In the model reasoning stage, the system respectively constructs a daily batch perception scoring model and a monthly structured user portrait model to support minute-level real-time diagnostic evaluation and periodic trend analysis. The daily batch scoring model is the core module, its input feature is the structured diagnostic feature vector constructed in Step 1, and its output is the user's network perception quality score, with a value range of [0,10], wherein a lower score indicates worse network experience. The system sets three levels of diagnostic labels according to the scoring result: Score ≤ 3 is determined as "severe quality degradation", 3 < Score ≤ 6 is determined as "moderate perception problem", and Score > 6 is determined as "normal". The model is trained by supervised learning, the labels are derived from server-side historical data such as user satisfaction scores, complaint handling conclusions and work order dispatch results, meanwhile the chief complaint classification result is introduced as weak supervision information to assist label construction, so as to improve label density and model generalization ability. Training samples are derived from diagnostic data of multiple cities across the country in the past three months, the samples are divided into a training set, a verification set and a test set according to the ratio of 70%:15%:15%, and sampling strategies such as SMOTE are adopted to ensure the balanced distribution of various labels.

[0031] After multiple rounds of comparative experiments, including XGBoost, LightGBM, Random Forest, and Multilayer Perceptron (MLP), XGBoost was ultimately selected as the deployment model to balance AUC performance and inference latency requirements. During model training, a combination of Grid Search and Bayesian optimization algorithms was used for parameter tuning, adjusting core parameters such as learning rate, tree depth, and subsampling ratio. The final result on the test set achieved an AUC > 0.85, MSE < 1.2, and an F1-score above 0.79. The model is deployed on the SparkML inference service framework, with a Kafka scheduling mechanism that automatically triggers batch inference processes daily at midnight. It also supports minute-level real-time diagnostic calls. The system's single-round concurrent inference capacity exceeds 1000 users, and the single diagnostic response latency is controlled within 2 seconds, meeting the requirements for rapid perception capabilities in high-concurrency environments.

[0032] Step 3: Root Cause Analysis For users whose perception score is below a set threshold (e.g., Score ≤ 3), the system will automatically enter the root cause analysis process. The attribution module will then accurately break down and structurally label the perception degradation problem. First, after scoring inference, the system synchronously outputs the marginal contribution of each input feature. The Shapley value decomposition method is used to quantify the influence weight of each feature in the scoring decision, constructing an individual-level diagnostic item contribution ranking list (Top-N). Based on this, statistical analysis of the conditional probability, mutual information, and frequency of collaboration between each feature combination and perception degradation events is performed on a large-scale training sample to form a diagnostic item association graph. This graph is used to label data-driven structural relationships such as "decreased negotiation rate is often accompanied by video stuttering, and high DNS latency co-occurs with connection failures."

[0033] To further eliminate spurious correlations and extract core causal paths, the system introduces structural equation modeling and causal graph reasoning mechanisms to construct multi-level causal chains such as "Wi-Fi negotiation capability → terminal connection rate → video buffering time," and filters out redundant edges and weak dependencies through path weight adjustment mechanisms. The system has established typical causal path clusters for various scenarios (such as [Wi-Fi bottleneck], [optical link degradation], and [terminal protocol compatibility issues]) to guide subsequent optimization suggestion matching.

[0034] Simultaneously, the system uses KMeans and DBSCAN algorithms to perform unsupervised clustering of historical poor-quality user characteristics, automatically extracting typical abnormal combination patterns, such as attribution profiles like "old equipment + 2.4GHz + slow DNS + page loading failure," and marking them as reusable problem templates. To enhance the ability to transfer historical experience, the system introduces the Dynamic Time Warping (DTW) algorithm to perform trajectory similarity matching on user feature time series, finding historical cases similar to the current abnormal trajectory and constructing an "analogy attribution library," supporting recommended root cause output and completion-style graph enhancement. Finally, this module generates structured root cause output, including contribution item ranking, causal chain path labels, analogy case references, and attribution explanation text, serving as direct input to the next suggestion generation engine.

[0035] Furthermore, the system constructs a diagnostic knowledge graph to support the structured interpretation of attribution results and recommendation matching. The graph consists of anomalies, causal paths, problem labels, and recommendation nodes. Nodes and their relationships are automatically extracted from historical diagnostic records, attribution outputs, and engineer action data, and stored in a graph database. It supports semantic retrieval based on path labels or anomaly combinations, providing structured input for subsequent recommendation generation. Step 4: Optimize the suggestion generation and feedback mechanism After root cause analysis is completed, the system automatically enters the optimization suggestion generation stage. By matching the attribution results (Top anomalies, causal path labels, cluster profiles) with suggestion rules in the knowledge graph, it generates structured, multi-role-adaptive suggestion solutions. The suggestion generation logic in this stage is based on the following three types of interactive objects for categorized output: (1) Visual prompts and suggestions for end users: The system maps abnormal items to easy-to-understand problem descriptions and recommends corresponding operation suggestions, such as: "The current connected router is on the 2.4GHz band. It is recommended to switch to 5GHz to obtain a higher speed"; "It was detected that the connected terminal only supports 802.11n. It is recommended to replace it with a device that supports 802.11ac or higher protocols"; such suggestions are displayed in pop-up windows on terminal pages such as user APP and home gateway management interface, and support "click to execute" or "confirm completion" feedback.

[0036] (2) Professional handling suggestions for customer service and engineers: The system outputs processing instructions for complex attribution paths (such as master-slave topology anomaly, link degradation, equipment aging, etc.), such as: "The CPU load of the main router is higher than 90% for a long time. It is recommended to check on-site whether there is abnormal access of IoT devices". The suggestions are presented through the customer service CRM system or the installation and maintenance dispatch platform, and can be integrated with work order information to assist in troubleshooting.

[0037] (3) Intelligent dispatching and automatic generation of fault work orders: When the score is lower than 2.5 and the attribution result hits the high confidence fault flag (such as network outage, frequent disconnection, packet loss exceeding the threshold of key nodes, etc.), the system will automatically generate a work order with diagnostic tags, synchronize it to the dispatching system through the interface, and perform the nearest dispatching based on the user's geographical location, equipment type and installation and maintenance strategy, so as to realize the fully automatic closed loop of diagnosis-dispatch.

[0038] (4) Feedback closed-loop mechanism: All suggestions outputs support users or engineers to perform response operations on the terminal, such as "I have completed the optimization" or "The problem has been solved". The system will collect this behavior and form a feedback record. Such feedback will be used as a diagnostic result collection signal, written into the feedback sample database, and participate in the incremental retraining of the scoring model, the update of the suggestion graph rule weights and the adjustment of the confidence of abnormal paths, so as to realize the model self-learning and optimization iteration, and support the continuous evolution of the system under different network environments and terminal structures.

[0039] Example: The one-click intelligent diagnostic system described in this invention supports multi-entry triggering of diagnostics from customer service agents, user self-service terminals, and installation and maintenance engineers. After the system is launched, it is applied to diagnosing perceived problems, root cause analysis, and optimization recommendations for home broadband users. The diagnostic task processing flow is as follows: Figures 1 to 5 As shown.

[0040] Step 1: Triggering the diagnostic task like Figure 1 As shown, the system supports the following three types of diagnostic triggering methods: User self-trigger: When a user clicks the "Network Checkup" button through the home gateway management app interface, the system records their account and connected device information and immediately creates a diagnostic request; Customer service assistance triggered: When a user calls to report a network problem (such as "video streaming is choppy at home"), the customer service representative clicks the "One-Click Diagnosis" button in the CRM interface, and the system receives the input PPPoE account, complaint text, and other request parameters. System-initiated batch triggering: The system dispatch center executes a diagnostic scheduling script every morning at midnight, automatically filtering accounts with poor internet quality in the past 24 hours (such as abnormal upload and download speeds, frequent disconnections), and creating diagnostic tasks in batches.

[0041] Upon receiving any type of diagnostic request, the system will uniformly construct a diagnostic task request body (including fields such as trigger source, account ID, trigger time, problem description, and terminal information), push it to the "Diagnostic Task Management Module" for queuing and processing, and trigger the subsequent data preparation process.

[0042] Step 2: Connecting multi-source data

[0043] All data windows are unified to a sliding window period of "T-7 to T-1". Within this time range, the system counts the frequency of anomalies, average values ​​and rates of change of various user indicators to form basic statistical features. At the same time, it generates biased embedding representations for text fields (such as chief complaints) as extended features.

[0044] Taking user "Zhang" as an example, his diagnostic request was initiated at 19:42 on July 8, 2024. The system automatically retrieved his running data for the past 7 days, and the key raw data fields generated are as follows: Link quality: The optical modem's received optical power is -20.8 dBm (normal); the average DNS resolution latency is 94ms; the PPPoE connection is uninterrupted; Wi-Fi status: The currently connected device is Xiaomi TV A55, which only supports the 2.4GHz / 802.11n protocol, with a negotiated speed of 72 Mbps, and has dropped the connection 4 times in the past 3 days; Terminal operating load: The main router's CPU load has been maintained above 91% for a long time, and there is a phenomenon of frequent WAN port disconnection (3 times / day); Behavioral characteristics: Video usage time during evening peak hours > 6 hours / day; Chief complaint identification result: After analysis by the large language model, it is classified as a problem tendency of "video stuttering + frequent Wi-Fi interruption".

[0045] After being cleaned, normalized, and windowed, the data is then processed by Spark's feature engineering module to construct a unified vectorized structure for direct use by the subsequent model inference module.

[0046] Step 3: Perceptual rating model inference like Figure 2 and Figure 4 As shown, after constructing the feature vector, the system immediately calls the scoring model to quantitatively evaluate the current user's network experience perception status. The scoring model is a supervised learning-based regression structure model, with the diagnostic feature vector as input. The feature vector is input into the scoring model to evaluate the user's network experience perception status. The model adopts an ensemble learning structure, with XGBoost as the base classifier, and outputs a user perception score. The score range is [0, 10], with lower scores indicating worse perception.

[0047] The scoring function takes the following form:

[0048]

[0049] Score ≤ 3: Severely poor quality 3<Score ≤ 6: Mild poor quality Score>6: Normal The model adopts XGBoost as the core classifier, fuses high-importance features such as behavior, terminal, link, and Wi-Fi status, and has strong nonlinear modeling capability and feature interaction expression ability. Model training uses actual diagnostic data from the operator platform in the past 3 months, covering complaining users, proactively fault-reporting users and users labeled with low satisfaction. The composition of annotation labels includes: Main supervision labels: derived from user satisfaction scores (satisfactory / medium / poor) and dispatching conclusions (whether on-site service is required); Weak label signals: tendency labels generated by large language models based on the classification results of users' chief complaint texts (such as "freezing", "Wi-Fi disconnection"); Auxiliary labels: including feedback data such as click behavior on diagnostic suggestions and speed improvement after diagnosis; To solve the problem of label imbalance, the system adopts a data augmentation strategy combining undersampling and SMOTE, and ensures the stability of feature distribution through cross-validation. The final model parameters are as follows: tree depth is 6, learning rate is 0.1, subsampling rate is 0.8, training rounds are 500, evaluation indicators are AUC>0.85, MSE<1.2, F1-score is close to 0.8, which meets the online requirements.

[0050] The parameters of the model have been frozen during the training stage. In the inference stage, the model is deployed in the SparkML service framework under the K8S container environment, supporting batch task execution every early morning (such as batch scoring of abnormal accounts generated on the previous day), and also supporting real-time online invocation at minute level (such as user actively clicking "network physical examination" to trigger scoring inference). The system has high concurrency capability: the concurrent processing capability of a single task can reach 1000 users, the average daily inference invocation times exceed 800,000, and the response delay of a single diagnosis is controlled within 1.8 seconds.

[0051] In addition, the model output also includes the Shapley value contribution weight corresponding to each input feature, for the subsequent root cause analysis module to perform importance ranking and path attribution. All scoring results and diagnostic metadata are uniformly written into the diagnostic task result table after generation, and enter the next attribution processing process.

[0052] The input vector obtains the score value through inference by the XGBoost perception model. Taking the user "Zhang" as an example, the inference result is:

[0053] It belongs to the category of severe perception problems and enters the next step of abnormal attribution analysis program Step 4: Abnormal attribution analysis If the score falls within the poor quality range, the system immediately triggers root cause analysis. The attribution module comprehensively utilizes feature importance ranking, Shapley value decomposition, and anomaly graph modeling to identify the specific contribution of diagnostic items to perceived degradation.

[0054] The attribution process is as follows: 1) Standardize the input data and calculate the marginal influence of each feature Xi using the Shapley value algorithm formula:

[0055]

[0056]

[0057]

[0058]

[0059]

[0060] Ranked by contribution (TOP-N) Taking user "Zhang" as an example, the top 3 attribution results are as follows:

[0061] 2) Causal path construction and graph matching To avoid misidentifying spurious factors, the system performs path analysis based on structural equation modeling and causal graphs, constructing a logical chain of problem evolution through a predefined path library and data-driven learning results. Taking Zhang as an example, the system identified an explainable path: "Old terminal → Limited negotiation rate → Increased video cache → Increased stuttering experience," and labeled the path as: [Video-related - Wi-Fi bottleneck] 3) Problem clustering and analogical attribution The system further clusters current samples and historical poor-quality user characteristics based on the KMeans and DBSCAN clustering algorithms to extract common failure modes (such as "old terminal + channel congestion + high download failure rate"). Simultaneously, it uses the DTW algorithm to perform time-series similarity matching on abnormal evolution trajectories, retrieves analogous cases, and achieves "attribution template completion" capabilities.

[0062] Step 5: Suggestions for Generation and Closed-Loop Feedback After completing root cause analysis, the system will generate personalized processing suggestions by calling the predefined "Optimization Suggestion Knowledge Graph" module based on the attribution output (such as Top Anomalies, Causal Path Labels, Problem Template Types, etc.). The suggestion generation logic performs semantic mapping between attribution labels and suggestion nodes in the graph to achieve consistent matching of problem and countermeasure, and automatically distributes them to the corresponding terminals according to the user role. These suggestions mainly include the following three categories: like Figure 2 As shown, the system generates personalized suggestions based on attribution tags and distributes them according to roles: (1) Recommendations for end users (for home users via APP) Based on the perception score and the terminal's capabilities, the system outputs easy-to-understand operational suggestions. Taking Mr. Zhang as an example, after completing the diagnosis, the client received the following suggestions. "The currently connected device only supports the 2.4GHz band. It is recommended to connect to 5GHz to improve the speed." "Limited terminal negotiation rate detected. It is recommended to replace the device with one that supports the 802.11ac protocol." "During the evening rush hour, when multiple devices are using high-definition video simultaneously, it is recommended to reduce concurrent playback."

[0063] It is recommended to display the diagnostic results on the home gateway app in a graphic format, supporting buttons such as "Execute Recommendations" and "Optimized" to facilitate user self-operation and result feedback.

[0064] (2) Suggestions from customer service representatives and installation / maintenance engineers The system synchronizes attribution tags and handling suggestions to the service personnel interface through the CRM and dispatch platform interface to assist in problem explanation and service suggestions: "The main router's CPU load has been too high for an extended period. We recommend on-site investigation to check for abnormal IoT device connections." "The terminal protocol version is too low, which may cause a rate bottleneck. We recommend assisting the user in replacing the terminal."

[0065] These suggestions are presented through structured templates, making it easier for customer service staff to quickly interpret and communicate them, thereby improving the first-response resolution rate.

[0066] 3) Automatic order dispatching mechanism and work order generation If the system determines that the user's score is ≤2.5 and there are high-confidence fault characteristics (such as abnormal optical power, link interruption, frequent disconnection, etc.), it will trigger the automatic work order dispatch logic, generate a work order containing fields such as diagnostic tags, fault description, and equipment status, and push it to the local installation and maintenance team through the work order dispatch platform to achieve closed-loop collaboration of "intelligent diagnosis + intelligent work order dispatch".

[0067] (4) Feedback closed-loop data acquisition and model learning The system continuously monitors the response behavior of users or engineers to suggestions, including: Did the user click on the suggested action? Should it be marked as "resolved"? Whether the rating will improve subsequently, etc.

[0068] All behavioral signals and suggestion effects will be written into the "closed-loop feedback database" and participate in the next round of incremental model training and knowledge graph suggestion weight update, thereby building a closed-loop optimization system with self-learning capabilities to continuously improve suggestion coverage and user response matching.

[0069] Step Six: System Performance Indicators and Application Results The one-click intelligent diagnostic system described in this embodiment was deployed and launched in early 2024 at a provincial operator branch, covering more than 20 cities and prefectures across the province for residential broadband users. The system is integrated into the existing customer service platform, installation and maintenance dispatch system, and user gateway management App, supporting various business scenarios such as complaint handling, self-diagnosis, and proactive detection, and has been fully put into daily operation and maintenance.

[0070] According to actual operational data provided by the operator, as of June 2025, the system had processed over 12 million user perception scoring and diagnostic tasks, with an average of over 850,000 inference calls per day. Key operational achievements are as follows: 1) The consistency rate between the model score and the conclusions of manual handling reached 84.3%; 2) The percentage of users who did not require order dispatch after self-diagnosis was 48%. 3) The diagnostic response delay should be controlled within 1.8 seconds; 4) User satisfaction rating increased by +17.6%; 5) The low-problem work order identification and filtering rate reaches 23.9%.

[0071] The system features an online model learning and evolution mechanism, supporting automatic scheduling of training tasks every 10 days. It annotates newly collected behavioral labels such as "return visit scores," "suggestion response behaviors," and "closed-loop markers" within the recent period to generate incremental training sets. After verification, the new model version is deployed to replace the old version, while simultaneously updating the suggestion matching weight parameters in the knowledge graph to ensure the model maintains long-term robustness and suggestion coverage breadth under changes in network structure, terminal type, and user behavior.

[0072] The examples described above are merely preferred embodiments of the present invention, and while their descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A one-click intelligent diagnostic method for broadband networks based on user-perceived data, characterized in that, Includes the following steps: Data fusion and feature construction steps: Using user accounts as the main index, collect network link performance indicators, user behavior profile features, Wi-Fi operation status data, terminal device identification features, uplink device and access link parameters, service history and perception tag data from multiple source systems, and form a structured input vector through feature engineering processing; The steps for building a model training and inference strategy are as follows: a supervised learning method is used to train the perception scoring model, outputs the user's network perception quality score, and sets diagnostic labels based on the scoring results; Root cause analysis steps: For users whose perception scores are below a set threshold, the problem is attributed through Shapley value decomposition, causal graph reasoning, anomaly clustering and time series matching; Optimization suggestion generation and feedback steps: Based on the attribution results, match the knowledge graph to generate optimization suggestions for multiple roles, and collect feedback data to achieve closed-loop iteration between the model and the knowledge base; In the data fusion and feature construction steps: The multi-source system includes a network management system, an optical access management platform, a home Wi-Fi status monitoring platform, a terminal identification module, and a customer service system; The feature engineering process includes data cleaning, normalization, and encoding. A sliding window method is used to count the frequency of anomalies and the rate of variation. The user's complaint text is combined with vector encoding of a large language model to generate representational features. In the steps of constructing the model and training the inference strategy: The perception scoring model uses the XGBoost algorithm and outputs a scoring range of [0, 10]. The model training labels are derived from user satisfaction scores, complaint handling conclusions, order dispatch results, and weakly supervised information on the main complaint classification. The model is deployed on the SparkML inference service framework, supporting batch inference and real-time invocation, with a single diagnostic response latency controlled within 2 seconds. The root cause analysis steps include: The contribution of each feature to the scoring results is calculated based on the Shapley value, and a Top-N outlier ranking is generated. Multi-level causal chains are constructed using structural causal graphs to eliminate redundant edges and weak dependencies; KMeans and DBSCAN are used to perform unsupervised clustering on the characteristics of users with poor historical quality, and typical abnormal combination patterns are extracted. A dynamic time warping algorithm is introduced to perform user temporal behavior similarity matching, and an analogy attribution library is constructed. The optimization suggestion generation and feedback steps include: Based on the attribution results, match the suggestion rules in the knowledge graph to generate visual prompts for end users and professional handling suggestions for customer service and engineers; It supports automatic work order dispatch and work order generation. When the score is below the threshold and hits the high confidence fault flag, a work order is automatically generated and dispatched. Collect feedback from users or engineers for incremental model training, knowledge graph weight updates, and path confidence adjustments.

2. A one-click intelligent diagnostic system for broadband networks based on user-perceived data, characterized in that: The system, applied to the one-click intelligent diagnosis method for broadband networks based on user-perceived data as described in claim 1, further includes: The data fusion module is used to collect and process multi-source heterogeneous data to construct a user network perception state map; The scoring model module uses an ensemble learning algorithm to output a perceptual score and provides an explanation of the feature contribution. The root cause analysis module combines structural cause-effect diagrams and diagnostic item maps to generate traceable abnormal attribution paths; The abnormal trajectory matching module uses the DTW algorithm to calculate the similarity of users' historical time-series behavior and builds an analogy attribution library; The suggestion generation and feedback module relies on the knowledge graph to achieve linkage and matching between problem tags and suggestion rules, and updates the model and graph based on feedback.

3. The broadband network one-click intelligent diagnostic system based on user-perceived data according to claim 2, characterized in that: The system is deployed on a distributed data processing platform, based on Spark, Kafka, HDFS, and Kubernetes container environment; It supports three diagnostic task triggering methods: user self-triggering, customer service assistance triggering, and system-initiated batch triggering. It has minute-level diagnostic task scheduling and real-time concurrent scoring inference capabilities, with a single round of concurrent inference capability exceeding 1000 users.

4. The broadband network one-click intelligent diagnostic system based on user-perceived data according to claim 2, characterized in that, The scoring model module: XGBoost was used as the core classifier, and Grid Search and Bayesian optimization algorithms were used for parameter tuning during training. The model evaluation metrics include AUC > 0.85, MSE < 1.2, and F1-score ≥ 0.79; It supports batch inference at midnight every day and real-time calls at the minute level, with an average of over 800,000 inference calls per day.

5. The broadband network one-click intelligent diagnostic system based on user-perceived data according to claim 2, characterized in that, The root cause analysis module: Construct a diagnostic knowledge graph, which consists of anomalies, causal paths, problem labels, and suggested nodes, and store it in a graph database; It supports semantic retrieval based on path tags or anomaly combinations, providing structured input for suggestion generation.

6. The broadband network one-click intelligent diagnostic system based on user-perceived data according to claim 2, characterized in that, The suggestion generation and feedback module: The generated suggestions include visual prompts for end users, professional handling suggestions from customer service and engineers, and automated work orders; The feedback mechanism includes user click operations. The system collects user click operation records and uses them for incremental training and graph augmentation. This enables a closed-loop path from diagnosis to optimization to learning, supporting the continuous evolution of the system under different network environments and terminal structures.

Images