Method and system for identifying intention in operator OLT fault processing scene
By employing methods such as data cleaning, deduplication verification, full intent database construction, and domain knowledge fusion, combined with the dual-tower model and BERT semantic representation, the problems of low efficiency and high misjudgment rate in OLT fault handling were solved, realizing the intelligent transformation of OLT fault handling and improving operation and maintenance efficiency and network reliability.
Patent Information
- Application Number
- CN202510886605.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2025-10-28
AI Technical Summary
Existing technologies are inefficient in OLT fault handling, rely on expert experience, have a high misjudgment rate, and existing NLP intent recognition technologies suffer from insufficient domain adaptation, data noise interference, and rigidity in dynamic scenarios in OLT fault scenarios.
采用数据清洗、去重校验、全意图库建设、双塔模型与BERT语义表征相结合的方法，结合领域知识融合和动态意图优化，通过MinHash LSH和人工抽样检验进行数据预处理，使用双塔模型和FastText模型进行意图匹配，结合OLT拓扑数据和健康状态数据进行故障判断，进行增量学习和规则处理以适配新故障场景。
实现了OLT故障处理的智能化转型，提高了故障定位效率，降低了误判率，提升了运维效率和网络可靠性，降低了运维成本。
Smart Images
Figure CN120849608A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of natural language processing technology, specifically to an intent recognition method and system for handling OLT faults in telecom operators. Background Technology
[0002] OLT (Optical Line Terminal) is the core equipment of fiber optic access networks, and its failure directly affects the network services of large-scale users. Currently, operators rely on manual processing of fault tickets, but OLT fault descriptions usually contain highly specialized terminology (such as "PON port optical attenuation" or "OLT disconnection"), unstructured text noise (such as redundant descriptions and spelling errors), and complex logical relationships (such as the topological relationship between OLT and ONU).
[0003] Traditional manual analysis is inefficient and relies heavily on expert experience, resulting in time-consuming fault location and a high misjudgment rate. Existing NLP intent recognition technologies (such as keyword matching and single BERT models) perform well in general scenarios, but they suffer from problems such as "insufficient domain adaptation," "data noise interference," and "rigidity of dynamic scenarios" in OLT fault scenarios. To improve operational efficiency, reduce operational costs, enhance network reliability, and promote comprehensive intelligent network operation and maintenance, an intent recognition method for operator OLT fault handling scenarios is designed. Summary of the Invention
[0004] The purpose of this invention is to provide an intent recognition method and system for operator OLT fault handling scenarios, so as to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: an intent recognition method in a carrier OLT fault handling scenario, comprising the following data cleaning and deduplication verification steps:
[0006] S101: Based on OLT fault scenarios, the original data is cleaned by removing irrelevant characters, filtering stop words, and filtering by rules.
[0007] S102: Use a combination of MinHash LSH (Local Sensitive Hash) and manual sampling verification to perform deduplication verification on the cleaned data.
[0008] Preferably, the following steps are included in the construction of the full intent database and intent matching and classification:
[0009] S201: Construct a comprehensive intent library, define OLT fault-related intents, and build an intent tagging system by combining historical work orders and expert knowledge. This system covers "optical module fault", "link interruption", "OLT out of management", "create scheduling work group", "disband scheduling work group", "generate fault recovery information", "information release", "fault delimitation and location", "fault recovery verification", "one-click handling", "generate fault summary report", "query OLT uplink port alarm", "query OLT uplink port traffic", "query OLT service user count", "query OLT data center", "query OLT data center power outage status", "query OLT data center access control", "query OLT data center related projects", "query OLT logical topology", and "query handling plan", covering physical layer, protocol layer and service layer fault types.
[0010] S202: Perform large-granularity intent matching. Use a dual-tower model to encode the user input text and intent library text into vectors respectively. Perform preliminary intent matching by calculating cosine similarity. At the same time, use the FastText model to quickly match high-frequency intents and identify keyword combinations.
[0011] S203: Perform fine-grained intent classification, use BERT semantic representation to perform deep semantic encoding on user input text, capture contextual associations, use convolutional layers and pooling to extract local semantic features, generate text vector representations, and use a Softmax classifier to output the final intent probability distribution, and select the intent label with the highest confidence.
[0012] Preferably, the following steps are included: domain knowledge fusion and dynamic intent optimization.
[0013] S301: Perform domain knowledge fusion, combine OLT topology data to constrain the scope of intent recognition, and introduce OLT health status data to assist in judging the possibility of faults;
[0014] S302: Use incremental learning and rule processing to perform dynamic intent optimization, update the intent library based on user feedback, adapt to new fault scenarios, and correct model output based on business logic.
[0015] Preferably, the following multi-model comparison and verification steps are included:
[0016] S401: Perform multi-model comparison and validation, compare the accuracy and recall of the single-tower model and the dual-tower model on the test set, and verify the impact of each module on the effect through ablation experiments.
[0017] Preferably, the following indicators are used to evaluate the effectiveness of intent recognition: the "proportion of reduction in fault work order processing time" and the "first-time recognition accuracy" indicators are used to evaluate the effectiveness of intent recognition.
[0018] A system for intent recognition in an operator OLT fault handling scenario includes a data preprocessing module, which is used for:
[0019] Based on OLT fault scenarios, the original data is cleaned by removing irrelevant characters, filtering stop words, and filtering by rules.
[0020] The cleaned data is deduplicated using a combination of MinHash LSH (Local Sensitive Hash) and manual sampling verification.
[0021] Preferably, it includes an intent library construction and matching classification module, which is used for:
[0022] Build a full intent library, define intents related to OLT faults, and combine historical work orders and expert knowledge to construct an intent tag system covering fault types at the physical layer, protocol layer, and business layer.
[0023] Large-granularity intent matching is performed by using a dual-tower model to encode user input text and intent database text into vectors respectively, and preliminary intent matching is performed by calculating cosine similarity. At the same time, the FastText model is used to quickly match high-frequency intents and identify keyword combinations.
[0024] We perform fine-grained intent classification, use BERT semantic representation to perform deep semantic encoding on user input text to capture contextual relationships, use convolutional layers and pooling to extract local semantic features, generate text vector representations, and use a Softmax classifier to output the final intent probability distribution, and select the intent label with the highest confidence.
[0025] Preferably, it includes a domain knowledge fusion and dynamic optimization module, which is used for:
[0026] Combine OLT topology data to constrain the scope of intent recognition;
[0027] Introduce OLT health status data to help determine the likelihood of a fault;
[0028] Dynamic intent optimization is achieved through incremental learning and rule processing. The intent library is updated based on user feedback to adapt to new fault scenarios, and the model output is corrected based on business logic.
[0029] Preferably, it includes a multi-model comparison and verification module, which is used to: compare the accuracy and recall of the single-tower model and the dual-tower model on the test set; and verify the impact of each module on the intent recognition effect through ablation experiments.
[0030] Preferably, it includes an effectiveness evaluation module, which is used to evaluate the effectiveness of the intent recognition system using the metrics of "proportion of reduction in fault work order processing time" and "first-time recognition accuracy".
[0031] Compared with the prior art, the beneficial effects of the present invention are:
[0032] This invention proposes an intent recognition method and system for operator OLT fault handling scenarios. Through domain-enhanced intent recognition technology, it transforms natural language dialogue into executable maintenance actions, solving the core pain points of slow response, high misjudgment rates, and high costs in traditional OLT fault handling. This achieves a transformation from "human experience-driven" to "intelligent decision-driven." Its application value lies not only in the quantitative improvement of efficiency and accuracy but also in providing long-term support for operator network stability, user satisfaction, and business competitiveness through standardized processes and preventative maintenance capabilities. Attached Figure Description
[0033] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0034] To make the objectives, technical solutions, and advantages of the present invention clear and complete, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only some, not all, embodiments of the present invention, and are merely illustrative of the embodiments of the present invention. They are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0035] Example 1, please refer to Figure 1 The present invention provides a technical solution: 1. Data preprocessing and optimization
[0036] S101: Based on OLT fault scenarios, raw data is cleaned by removing irrelevant characters, filtering stop words, and filtering by rules.
[0037] Character cleaning: Remove irrelevant characters (such as special symbols, garbled characters), redundant spaces, and non-text content.
[0038] Stop word filtering: Remove general words (such as "of" and "and") that are not related to OLT faults, and retain domain keywords (such as "OLT", "optical attenuation", and "PON port").
[0039] Rule filtering: Filter low-quality text (such as duplicate submissions and invalid descriptions) based on regular expressions or domain rules.
[0040] S102: Use MinHash LSH (Locality Sensitive Hash) and manual sampling checks for deduplication verification:
[0041] II. Construction and Matching of OLT Fault Scenario Intent Library
[0042] S201: Construction of a comprehensive intent library. Define intents related to OLT faults and construct an intent tagging system by combining historical work orders and expert knowledge. This includes "Optical module fault," "Link interruption," "OLT out of management," "Create scheduling workgroup," "Disband scheduling workgroup," "Generate fault recovery information," "Information release," "Fault delimitation and location," "Fault recovery verification," "One-click handling," "Generate fault summary report," "Query OLT uplink port alarms," "Query OLT uplink port traffic," "Query the number of OLT service users," "Query the OLT data center," "Query the OLT data center power outage status," "Query the OLT data center access control," "Query the OLT data center related projects," "Query the OLT logical topology," and "Query handling plan," covering physical layer, protocol layer, and service layer fault types.
[0043] S202: Large-grained intent matching. A dual-tower model is used to encode user input text and intent library text into vectors respectively, and preliminary intent matching is performed by calculating cosine similarity. The FastText model is used to quickly match high-frequency intents and identify keyword combinations, such as "OLT + disconnect + restart".
[0044] S203: Fine-grained intent classification. BERT semantic representation is used to perform deep semantic encoding on user input text, capturing contextual relationships, such as "frequent OLT disconnections" implying hardware aging issues. Convolutional layers and pooling are used to extract local semantic features (such as fault location and device model) to generate text vector representations. A Softmax classifier is used to output the final intent probability distribution, and the intent label with the highest confidence is selected.
[0045] III. Fault Scenario Modeling and Optimization
[0046] S301: Domain Knowledge Fusion. By combining OLT topology data (such as OLT hierarchy and associated ONU devices), the scope of intent identification is constrained (e.g., an OLT fault only relates to its downstream devices). OLT health status data (such as CPU utilization and optical power) is introduced to assist in determining the likelihood of a fault (e.g., high optical attenuation indicates a fiber optic problem).
[0047] S302: Dynamic intent optimization is performed using incremental learning and rule processing. The intent library is updated based on user feedback (such as the "Results Analysis" module) to adapt to new fault scenarios (such as fault codes specific to new OLT devices). The model output is corrected based on business logic (e.g., in the case of "OLT disconnection," the power supply should be checked first, rather than the fiber optic cable).
[0048] IV. Verification and Deployment
[0049] S401: Multi-model comparison and validation, comparing the accuracy and recall of the single-tower model (BERT) and the dual-tower model (FastText+BERT) on the test set. Ablation experiments are used to verify the impact of each module (such as data cleaning and topological association) on the performance.
[0050] S402: Use the metrics of "proportion of reduction in fault work order processing time" and "accuracy of first identification" for evaluation.
[0051] As a preferred technical solution of the present invention, step S101 is based on the OLT fault scenario, and the original data is cleaned by removing irrelevant characters, filtering stop words, and filtering rules.
[0052] As a preferred embodiment of the present invention, step S102 uses MinHash LSH (Local Sensitive Hash) and manual sampling inspection for deduplication verification.
[0053] As a preferred technical solution of the present invention, step S201 involves the construction of a full intent library. This includes defining OLT fault-related intents and constructing an intent tagging system by combining historical work orders and expert knowledge. This includes steps such as "Optical module fault," "Link interruption," "OLT out of management," "Create scheduling workgroup," "Disband scheduling workgroup," "Generate fault recovery information," "Information release," "Fault delimitation and location," "Fault recovery verification," "One-click handling," "Generate fault summary report," "Query OLT uplink port alarms," "Query OLT uplink port traffic," "Query the number of OLT service users," "Query the OLT data center," "Query the OLT data center power outage status," "Query the OLT data center access control," "Query the OLT data center related projects," and "Query the OLT logical topology." As a preferred technical solution of the present invention, step "Query handling plan" covers physical layer, protocol layer, and service layer fault types.
[0054] As a preferred embodiment of the present invention, step S202 involves large-granularity intent matching. A dual-tower model is used to encode the user input text and the intent library text into vectors respectively, and preliminary intent matching is performed through cosine similarity calculation. The FastText model is then used to quickly match high-frequency intents and identify keyword combinations. For example, as a preferred embodiment of the present invention, the steps include "OLT + disconnect + restart".
[0055] As a preferred embodiment of the present invention, step S203 involves fine-grained intent classification. BERT semantic representation is used to perform deep semantic encoding on the user input text, capturing contextual relationships, such as "frequent OLT disconnections" implying hardware aging issues. Convolutional layers and pooling are used to extract local semantic features (such as fault location and device model) to generate text vector representations. A Softmax classifier is used to output the final intent probability distribution, and the intent label with the highest confidence is selected.
[0056] As a preferred embodiment of the present invention, step S301 involves domain knowledge integration. Combining OLT topology data (such as OLT hierarchy and associated ONU devices), the scope of intent identification is constrained (e.g., an OLT fault only relates to its downstream devices). OLT health status data (such as CPU utilization and optical power) is introduced to assist in determining the likelihood of a fault (e.g., high optical attenuation indicating fiber optic problems).
[0057] As a preferred embodiment of the present invention, step S302 uses incremental learning and rule processing for dynamic intent optimization. The intent library is updated based on user feedback (e.g., the "Result Analysis" module) to adapt to new fault scenarios (e.g., fault codes specific to new OLT devices). The model output is corrected based on business logic (e.g., "OLT disconnection" should prioritize checking the power supply over the optical fiber).
[0058] As a preferred technical solution of the present invention, step S401 involves multi-model comparison and verification, comparing the accuracy and recall of the single-tower model (BERT) and the dual-tower model (FastText+BERT) on the test set. Ablation experiments are used to verify the impact of each module (such as data cleaning and topological association) on the results.
[0059] As a preferred technical solution of the present invention, step S402 uses the indicators of "proportion of reduction in fault work order processing time" and "accuracy of first identification" for evaluation.
[0060] Example 2, based on Example 1, proposes a system for intent recognition in the operator OLT fault handling scenario according to claim 5, including a data preprocessing module. This module is used to: clean the original data based on the OLT fault scenario by removing irrelevant characters, filtering stop words, and filtering by rules; and perform deduplication verification on the cleaned data using a combination of MinHash LSH local sensitive hashing and manual sampling inspection.
[0061] The module includes an intent library construction and matching / classification module. This module is used for: building a full intent library, defining OLT fault-related intents, and constructing an intent label system covering physical layer, protocol layer, and business layer fault types by combining historical work orders and expert knowledge; performing large-granular intent matching, using a dual-tower model to encode user input text and intent library text into vectors respectively, performing preliminary intent matching through cosine similarity calculation, and using the FastText model to quickly match high-frequency intents and identify keyword combinations; and performing small-granular intent classification, using BERT semantic representation to perform deep semantic encoding on user input text, capturing contextual relationships, using convolutional layers and pooling to extract local semantic features, generating text vector representations, and using a Softmax classifier to output the final intent probability distribution, selecting the intent label with the highest confidence.
[0062] It includes a domain knowledge fusion and dynamic optimization module, which is used to: combine OLT topology data to constrain the scope of intent recognition; introduce OLT health status data to help judge the possibility of failure; use incremental learning and rule processing to perform dynamic intent optimization, update the intent library through user feedback, adapt to new failure scenarios, and correct the model output based on business logic.
[0063] It includes a multi-model comparison and validation module, which is used to: compare the accuracy and recall of single-tower and dual-tower models on the test set; and verify the impact of each module on the intent recognition effect through ablation experiments.
[0064] It includes an effectiveness evaluation module, which is used to evaluate the effectiveness of the intent recognition system using the metrics of "proportion of reduction in fault work order processing time" and "first-time recognition accuracy".
[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. An intent recognition method in a carrier OLT fault handling scenario, characterized in that: The data cleaning and deduplication verification steps include the following: S101: Based on OLT fault scenarios, the original data is cleaned by removing irrelevant characters, filtering stop words, and filtering by rules. S102: Use a combination of MinHash LSH (Local Sensitive Hash) and manual sampling verification to perform deduplication verification on the cleaned data.
2. The intent recognition method in a carrier OLT fault handling scenario according to claim 1, characterized in that: The following steps are included in the construction of the full intent database and the intent matching and classification process: S201: Construct a comprehensive intent library, define intents related to OLT faults, and build an intent tagging system by combining historical work orders and expert knowledge. This system covers "optical module fault," "link interruption," "OLT out of control," "create scheduling workgroup," "disband scheduling workgroup," "generate fault recovery information," "information release," "fault delimitation and location," "fault recovery verification," "one-click handling," "generate fault summary report," "query OLT uplink port alarms," "query OLT uplink port traffic," "query OLT service user count," "query OLT data center," "query OLT data center power outage status," "query OLT data center access control," "query OLT data center related engineering," "query OLT logical topology," and "query handling plan," covering physical layer, protocol layer, and service layer fault types. S202: Perform large-granularity intent matching. Use a dual-tower model to encode the user input text and intent library text into vectors respectively. Perform preliminary intent matching by calculating cosine similarity. At the same time, use the FastText model to quickly match high-frequency intents and identify keyword combinations. S203: Perform fine-grained intent classification, use BERT semantic representation to perform deep semantic encoding on user input text, capture contextual associations, use convolutional layers and pooling to extract local semantic features, generate text vector representations, and use a Softmax classifier to output the final intent probability distribution, and select the intent label with the highest confidence.
3. The intent recognition method in a carrier OLT fault handling scenario according to claim 2, characterized in that: This includes the following steps for domain knowledge integration and dynamic intent optimization: S301: Perform domain knowledge fusion, combine OLT topology data to constrain the scope of intent recognition, and introduce OLT health status data to assist in judging the possibility of faults; S302: Use incremental learning and rule processing to perform dynamic intent optimization, update the intent library based on user feedback, adapt to new fault scenarios, and correct model output based on business logic.
4. The intent recognition method in a carrier OLT fault handling scenario according to claim 3, characterized in that: The following multi-model comparison and validation steps are included: S401: Perform multi-model comparison and validation, compare the accuracy and recall of the single-tower model and the dual-tower model on the test set, and verify the impact of each module on the effect through ablation experiments.
5. The intent recognition method in a carrier OLT fault handling scenario according to claim 4, characterized in that: The following metrics were used to evaluate the effectiveness of intent recognition: the percentage reduction in fault work order processing time and the first-time recognition accuracy.
6. A system for the intent recognition method in the operator OLT fault handling scenario according to claim 5, characterized in that: Includes a data preprocessing module, which is used for: Based on OLT fault scenarios, the original data is cleaned by removing irrelevant characters, filtering stop words, and filtering by rules. The cleaned data is deduplicated using a combination of MinHash LSH (Local Sensitive Hash) and manual sampling verification.
7. The system according to claim 6, characterized in that: Includes an intent library building and matching classification module, which is used for: Build a full intent library, define intents related to OLT faults, and combine historical work orders and expert knowledge to construct an intent tag system covering fault types at the physical layer, protocol layer, and business layer. Large-granularity intent matching is performed by using a dual-tower model to encode user input text and intent database text into vectors respectively, and preliminary intent matching is performed by calculating cosine similarity. At the same time, the FastText model is used to quickly match high-frequency intents and identify keyword combinations. We perform fine-grained intent classification, use BERT semantic representation to perform deep semantic encoding on user input text to capture contextual relationships, use convolutional layers and pooling to extract local semantic features, generate text vector representations, and use a Softmax classifier to output the final intent probability distribution, and select the intent label with the highest confidence.
8. The system according to claim 7, characterized in that: This includes a domain knowledge fusion and dynamic optimization module, which is used for: Combine OLT topology data to constrain the scope of intent recognition; Introduce OLT health status data to help determine the likelihood of a fault; Dynamic intent optimization is achieved through incremental learning and rule processing. The intent library is updated based on user feedback to adapt to new fault scenarios, and the model output is corrected based on business logic.
9. A system according to claim 8, characterized in that: It includes a multi-model comparison and validation module, which is used to: compare the accuracy and recall of single-tower and dual-tower models on the test set; and verify the impact of each module on the intent recognition effect through ablation experiments.
10. A system according to claim 9, characterized in that: It includes an effectiveness evaluation module, which is used to evaluate the effectiveness of the intent recognition system using the metrics of "proportion of reduction in fault work order processing time" and "first-time recognition accuracy".