Intelligent fuel gas quality inspection method and equipment based on vector matching and dynamic rule engine

Abstract

The invention discloses an intelligent fuel gas quality inspection method and device based on vector matching and a dynamic rule engine. The method comprises the steps that data preprocessing is conducted through audio and text cleaning; setting a vectorization and semantic understanding layer for converting the preprocessed text into a high latitude vector; setting a configurable quality inspection rule engine which is used for realizing four layers of configurable quality inspection rule engines of quality inspection operators, quality inspection conditions, quality inspection rules and quality inspection tasks; and a decision and feedback layer is arranged and used for correcting error data in a manual rechecking mode. Vector matching is carried out based on an M3E-Base text embedding model, and quality inspection operators, conditions, tasks and rules which can be dynamically configured are combined. By fusing an advanced semantic understanding technology and a configurable quality inspection rule engine, gas customer service quality management is promoted to evolve towards automation, intellectualization and precision, and powerful technical support is provided for safe operation and excellent service of gas enterprises.

Description

Gas Intelligent Quality Inspection Method and Equipment Based on Vector Matching and Dynamic Rule Engine Technical Field

[0001] This invention relates to the interdisciplinary field of gas service and artificial intelligence technology, specifically to a gas intelligent quality inspection method and equipment based on vector matching and dynamic rule engine. Background Technology

[0002] As my country's gas industry continues to expand, natural gas consumption is growing rapidly, and the user base is expanding, higher demands are being placed on the quality and efficiency of gas companies' customer service. Gas customer service centers need to handle various business scenarios, including business inquiries, fault reporting, leak alarms, bill inquiries, and transfer applications. Their service process not only concerns user experience but also directly relates to public safety.

[0003] Traditional customer service quality management relies heavily on a "manual sampling" model, where quality inspectors randomly select a small number of call recordings for subsequent listening and evaluation. This approach is no longer sufficient to handle massive amounts of customer service interaction data and suffers from numerous bottlenecks, including low coverage, strong subjectivity, and delayed feedback.

[0004] To address these issues, the industry has begun exploring the introduction of artificial intelligence technologies, particularly Automatic Speech Recognition (ASR), Natural Language Processing (NLP), and intelligent voice quality inspection systems, into the gas customer service process, aiming to achieve more efficient, comprehensive, and objective service quality monitoring and risk management.

[0005] Although intelligent quality inspection is a clear development direction, existing technical solutions still have significant shortcomings in the specific field of gas, which has high safety requirements, and cannot meet the needs of refined and intelligent management. Specifically, this invention aims to solve the following key technical problems: 1. Low quality inspection coverage and efficiency: The sampling ratio of traditional manual quality inspection is usually extremely low (less than 1-5% of all calls), resulting in a large number of problems in service conversations going undetected, making it impossible to achieve full-volume quality inspection.

[0006] For gas customer service, this means they may not be able to promptly detect and correct procedural errors or inappropriate wording when handling emergency alarms such as "gas leaks," posing a safety hazard. While simple keyword matching technology can improve efficiency to some extent, it cannot understand contextual semantics, resulting in a high rate of false positives and false negatives.

[0007] 2. Quality inspection standards are subjective and lack simplistic dimensions. Manual quality inspection relies heavily on the personal experience and judgment of inspectors, making it difficult to guarantee the objectivity and consistency of evaluation standards. Existing automated tools are mostly limited to detecting preset sensitive words or checking basic indicators such as speech speed and silence, lacking the ability to assess deeper service quality aspects such as the semantic coherence of the dialogue, the customer's true intentions, and the accuracy of business knowledge. For example, they cannot effectively determine whether the customer service representative's explanation of the "gas meter transfer process" is accurate and comprehensive.

[0008] 3. Difficulty in Adapting to the Professionalism and Complexity of the Gas Industry: Gas customer service conversations contain numerous technical terms (such as "gas meter number," "pressure anomaly," and "riseline inspection") and specific business scenarios (such as "in-home safety inspection" and "hazard handling"). Without targeted training and adaptation, the semantic understanding accuracy of general-purpose intelligent quality inspection models in this field will be significantly reduced. Furthermore, gas safety regulations and service standards are dynamically updated, making it difficult for fixed, rule-based quality inspection systems to quickly adapt to these changes.

[0009] 4. Weak Multimodal Data Processing and Correlation Capabilities: Gas customer service quality inspection should not analyze voice or text in isolation. Ideally, it should be correlated and analyzed with diverse information such as work order systems, user files, and gas meter readings. For example, a customer service representative's promise over the phone that "a repairman will arrive within 2 hours" needs to be compared with the actual dispatch and processing time in the work order system to verify the fulfillment of the promise. Current technology lacks the ability to effectively correlate and vectorize such cross-modal information.

[0010] In summary, the core technical problem that this invention aims to solve is: how to overcome the shortcomings of existing gas customer service quality inspection technologies in terms of processing efficiency, semantic understanding depth, industry professionalism, and multimodal information association, and to provide an intelligent quality inspection solution that can accurately understand the semantics of the gas field, support dynamic configuration throughout the entire process, and perform in-depth post-event analysis, thereby comprehensively improving the quality and safety risk control level of gas customer service. Summary of the Invention

[0011] This invention proposes a gas intelligent quality inspection method based on vector matching and a dynamic rule engine. Specifically, it involves a gas intelligent customer service quality inspection system and method based on M3E-Base text embedding model for vector matching, combined with dynamically configurable quality inspection operators, conditions, tasks and rules; it can at least solve one of the technical problems in the background art.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: a gas intelligent quality inspection method based on vector matching and a dynamic rule engine, which executes the following steps through computer equipment: S1, collecting and preprocessing multi-source data, including voice call data, online chat records, work order texts, and user feedback forms; cleaning the collected audio and text data to complete data preprocessing; S2, converting the preprocessed text data into high-dimensional vectors to support quality inspection rule matching based on semantic similarity operators; S3, executing quality inspection tasks through a configurable quality inspection rule engine, which supports hierarchical configuration of quality inspection operators, quality inspection conditions, quality inspection rules, and quality inspection tasks; S4, manually reviewing the quality inspection results, correcting erroneous data, and using the corrected data to fine-tune the semantic understanding model to improve the recognition accuracy of semantic similarity operators.

[0013] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0014] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0015] As can be seen from the above technical solutions, the gas intelligent quality inspection method based on vector matching and dynamic rule engine of the present invention can solve the following technical problems: 1. Traditional quality inspection methods are inefficient and have insufficient coverage. Traditional gas customer service quality inspection mainly relies on manual sampling, which is not only inefficient, but also has a very low sampling coverage (usually less than 5%), resulting in a large number of service quality problems and safety hazards in customer service conversations not being detected in a timely manner.

[0016] This "human wave tactic" is ill-suited to the rapidly growing number of gas users and the surge in customer service data, making quality inspection work a bottleneck for the efficiency of enterprise operations and management.

[0017] 2. Weak semantic understanding and lack of contextual analysis: Existing automated quality inspection systems mostly rely on simple keyword matching and fixed rules, failing to understand the deep semantics and contextual relationships of dialogues. For example, in emergency events such as gas leak alarms, traditional systems struggle to accurately determine whether customer service procedures are compliant and whether all key safety information has been conveyed. This invention introduces the M3E-Base model for vector matching, aiming to achieve deep semantic understanding of customer service dialogue content, thereby accurately identifying complex scenarios such as whether safety procedures are being followed correctly.

[0018] 3. Rigid quality inspection rules make it difficult to quickly adapt to business changes. Safety regulations and service standards in the gas industry are dynamically updated, while traditional quality inspection systems based on fixed rules are inflexible. Once the rules are set, modification and updates are very cumbersome and cannot respond quickly to business changes.

[0019] The configurable quality inspection operators, conditions, tasks, and rules system proposed in this invention allows managers to flexibly arrange quality inspection logic in a low-code manner, enabling the system to adapt quickly to emerging business scenarios and control requirements.

[0020] 4. Multimodal data isolation prevents the formation of a comprehensive quality inspection view. The quality and efficiency evaluation of gas customer service often requires the integration of various information, including voice conversations, generated work order texts, and even on-site photos. Current technologies typically process this data in isolation, lacking effective methods for correlation analysis.

[0021] This invention aims to connect these multimodal data and establish correlations between information from different modalities using vector technology, thereby forming a comprehensive and integrated quality assessment view of a customer service transaction. For example, comparing the customer service representative's mention of "unclear meter readings" with images of the gas meter at the scene improves the accuracy of the assessment.

[0022] Specifically, compared with the prior art, the beneficial effects of this invention are as follows: Closed-loop management from "quality inspection" to "quality and efficiency": The system can not only discover problems, but also optimize the customer service knowledge base, training system and business processes by deeply mining the quality inspection results, forming a continuous improvement closed loop of "monitoring-analysis-optimization-prevention", thereby fundamentally improving the overall service quality and operational efficiency of gas companies.

[0023] Significantly reduce safety risks: By real-time identification and high-priority alarms for dialogues involving safety red lines (such as "gas leak" and "suspected gas theft"), the system can transform passive response into proactive prevention, greatly reducing the probability of safety accidents and strengthening the safety line of gas use.

[0024] Empowering employees and optimizing the experience: The system provides new employees with AI-based real-time script assistance and business guidance, shortening the training cycle and improving the professional capabilities of frontline customer service staff. At the same time, more efficient and precise service directly enhances user satisfaction and trust.

[0025] In summary, this invention, by integrating advanced semantic understanding technology with a configurable quality inspection rule engine, aims to systematically solve long-standing pain points in the field of gas customer service quality inspection, promote the evolution of gas customer service quality management towards automation, intelligence, and precision, and provide strong technical support for the safe operation and excellent service of gas companies. Attached Figure Description

[0026] Figure 1 is a flowchart of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0028] As shown in Figure 1, the gas intelligent quality inspection method based on vector matching and a dynamic rule engine described in this embodiment constructs a data-driven, semantically understanding, and dynamically configurable intelligent quality inspection pipeline. Its overall architecture is shown in the table below. The system consists of four logical layers, progressively building upon each other to ultimately achieve accurate and efficient automated quality inspection; as shown in the table below:

[0029] The core technical solutions at each level will be described in detail below.

[0030] Specifically, the embodiments of the present invention include the following steps: S1, collecting and preprocessing multi-source data, including voice call data, online chat records, work order texts, and user feedback forms; cleaning the collected audio and text data to complete data preprocessing; S2, converting the preprocessed text data into high-dimensional vectors to support quality inspection rule matching based on semantic similarity operators; S3, executing quality inspection tasks through a configurable quality inspection rule engine, which supports hierarchical configuration of quality inspection operators, quality inspection conditions, quality inspection rules, and quality inspection tasks; S4, manually reviewing the quality inspection results, correcting erroneous data, and using the corrected data to fine-tune the semantic understanding model to improve the recognition accuracy of semantic similarity operators.

[0031] The following further explains: I. Data Collection and Preprocessing This layer is responsible for transforming the raw, unstructured customer service interaction data into standardized data that can be processed by the rule engine.

[0032] The system supports access to multiple data sources, including voice calls, online chat logs, work order texts, and user feedback forms. Voice data is transcribed into text using Automatic Speech Recognition (ASR) technology, and the transcribed text is then cleaned (e.g., removing interjections and correcting errors) and standardized (e.g., unifying date and address formats).

[0033] II. Vectorization and Semantic Understanding: Semantic Vectorization Based on M3E-Base: This is the core of this invention. A text embedding model specifically optimized for Chinese is used to convert the preprocessed customer service text into a 1024-dimensional dense embedding.

[0034] Domain Adaptation: To improve accuracy in the gas sector, the M3E-Base model is fine-tuned using a corpus containing a large number of gas-related technical terms (such as "gas meter calibration", "suspected gas leak", and "riseer valve"), so that the generated vectors can more accurately capture the semantic core of gas services.

[0035] Vector storage: The generated sentence vectors are stored in a vector database (such as Milvus) and associated with the metadata of the original dialogue (such as session ID, customer service number, business type, etc.), laying the foundation for subsequent efficient similarity retrieval.

[0036] The M3E-base infrastructure is built upon the chinese-roberta-wwm-ext model. This base model itself is a Chinese RoBERTa model pre-trained using Whole Word Masking (WWM) technology. M3E-Base inherits its 12-layer Transformer structure, with 1024 hidden layers and 12 attention heads.

[0037] M3E-base embedding generates token embeddings: After word segmentation, the input text is converted into a series of token embedding vectors. Context encoding: The token sequence is processed by a Transformer encoder, which uses a self-attention mechanism to calculate the weight of each token relative to all other tokens in the sequence, generating a vector representation rich in contextual information.

[0038] Self-Attention mechanism: The hidden state of a word is managed through a learnable weight matrix. Generate query vector, key vector, and value vector respectively:

[0039] Attention weights are calculated using the dot product of Q and K, and then normalized using Softmax, representing the importance of other lexical units to the current lexical unit.

[0040] in The dimension of the key vector, scaling factor. This is used to prevent the gradient from vanishing due to an excessively large dot product.

[0041] M3E-Base's 12 attention heads allow the model to capture different types of semantic relations (such as syntax, coreference, etc.) in parallel across different representation subspaces.

[0042] Feed-Forward Network (FFN): The feed-forward network within each Transformer layer typically consists of two linear transformations and a non-linear activation function (such as GELU).

[0043] Pooling: This is the key step in aggregating a variable-length sequence of token vectors into a fixed-length sentence vector. M3E-Base uses mean pooling by default. Its calculation formula can be simplified as follows:

[0044] This involves taking the arithmetic mean of the token vectors of all non-special tokens in the sequence. This strategy balances the contributions of all tokens, generating a stable 1024-dimensional sentence vector that is semantically representative overall.

[0045] The core training objective of M3E-base is contrastive learning. The model learns and adjusts its semantic space during training, bringing semantically similar samples (positive sample pairs) closer together and semantically unrelated samples (negative sample pairs) further apart. InfoNCE Loss or its variants are commonly used as the loss function during training. The basic idea is to enable the model to distinguish between positive and negative samples. The specific formula is shown below:

[0046] Similarity calculation: sim(q, k+) and sim(q, k-) in the formula are used to calculate the similarity between the query sample and positive samples, and between the query sample and each negative sample. Commonly used similarity functions include dot product (inner product) or cosine similarity.

[0047] Temperature coefficient τ: The temperature parameter τ is an important hyperparameter that scales the similarity score. A smaller τ makes the Softmax distribution more "sharp," causing the model to focus more on hard negative samples that are difficult to distinguish, thus learning finer boundaries. Conversely, a larger τ makes the distribution smoother.

[0048] Exponentialization and Normalization: The exp function in the numerator and denominator converts the similarity score into a positive number. The entire fraction can be viewed as normalizing the sum of the similarity scores of positive sample pairs and the sum of the similarity scores of all sample pairs (positive samples + all negative samples). The value of this fraction can be understood as: given a query sample, the model predicts that its corresponding positive sample is... The probability of that. Ideally, this probability should be close to 1.

[0049] Negative logarithm: Apply -log at the end. Since the model's predicted probability values ​​are between 0 and 1, taking the negative logarithm makes the loss value positive. The higher the predicted probability, the smaller the final loss value. Therefore, minimizing the InfoNCE loss is equivalent to maximizing the probability that the model correctly identifies a positive sample.

[0050] After a similarity measurement model is trained, the most commonly used method to measure the semantic similarity between two sentence vectors (such as A and B) is cosine similarity. It calculates the cosine of the angle between the two vectors in space, without considering their absolute magnitudes. The formula is:

[0051] Its range is [-1, 1], and the closer the value is to 1, the more similar the semantics are.

[0052] M3E-Base model fine-tuning data preparation: Data quality directly determines the fine-tuning effect. You need to prepare a sentence pair dataset containing text in the gas industry.

[0053] 1. Data Format: Model fine-tuning typically supports three formats: Pair: Contains positive sample pairs, such as ("Emergency response plan for gas pipeline leaks", "How to handle PE gas pipeline leaks").

[0054] Triplet: Add a negative sample to the sentence pair, for example ("gas stove ignition failure", "reasons why the gas stove won't ignite", "unstable water temperature in the gas water heater"), where the third sentence is an irrelevant sample of the first two.

[0055] ScoredPair: Directly labels the similarity score of the sentence pairs, for example ("Difference between CNG and LNG", "Difference between compressed natural gas and liquefied natural gas", 2.0). The higher the score, the more similar the sentences are.

[0056] 2. Data construction in the gas industry involves extracting text from industry standards, safety regulations, equipment manuals, and customer service QA, and constructing the data through the following methods: manual annotation: ensuring high quality, but at a higher cost.

[0057] Generate using large models: Generate synthetic data based on a knowledge base in the gas industry using large models such as DeepSeek.

[0058] Model fine-tuning: Once you have your data ready, you can use FineTuner provided by the uniem library to simplify the fine-tuning process.

[0059] Key parameter analysis: epochs: usually 3 to 5 rounds to avoid overfitting.

[0060] Batch size: A larger batch size results in more stable training, but requires more GPU memory. Gas text is generally quite long, so you can start with 32.

[0061] learning_rate: 2e-5 is a common starting point, which can be fine-tuned.

[0062] After fine-tuning the model, it will be automatically saved to a specified directory (such as finetuned-model) by default. You can load it using the sentence-transformers library for later use.

[0063] Finally, through comparative testing, the improvement brought about by the fine-tuning was intuitively felt.

[0064] Expected results: After fine-tuning, the semantic similarity scores of "service life of indoor gas hoses" and "how often gas connection hoses need to be replaced" should be significantly higher than their scores with "cathode protection principle of gas pipeline network", indicating that the model can better distinguish between relevant and irrelevant professional concepts.

[0065] III. Configurable Quality Inspection Rule Engine: This is the core innovation of this invention. It deeply integrates semantic vector matching technology with business rules through a flexibly orchestratable four-level component.

[0066] 1. Quality Inspection Operator: The smallest indivisible computational unit. Each operator is responsible for performing a basic detection function. For example: Emotion Recognition Operator: Based on speech spectrum or text keywords, determine the emotional state (positive, neutral, negative) of customer service or users.

[0067] Service language operator: Detects whether standard greetings and closing phrases are used at specific nodes (such as the beginning and the end).

[0068] Vector similarity operator: Given a dialogue text, it is vectorized using M3E-Base and its similarity is calculated with a pre-defined standard answer vector or a typical question case library. For example, the similarity between the vector of the user's description "My gas stove won't light" and the vectors of questions in the case library such as "Gas stove cannot be ignited" and "Stove malfunction" is calculated. If the similarity exceeds a threshold (e.g., 0.85), it is considered a match. This solves the semantic gap problem of keyword matching.

[0069] 2. Quality Inspection Conditions: Complex judgment logic composed of multiple quality inspection operators combined through logical operators (AND / OR / NOT). One condition represents a complete business scenario judgment. For example: Condition A (Leakage Alarm Red Line Condition): {Sensitive words such as "air leak" appear in the dialogue}, {The user's emotional state is negative}, and {The vector similarity operator determines that the similarity between the dialogue semantics and the standard leakage handling process is <0.5}. This condition is used to identify serious violations where customer service fails to handle leakage alarms according to security specifications.

[0070] Condition B (Business Specification Condition): {Customer service mentions "ownership transfer service"} and {Vector similarity operator determines that the vector similarity between the customer service explanation and the standard ownership transfer instructions is >0.8}. This is used to verify the accuracy of the business explanation.

[0071] 3. Quality Inspection Rules: A dynamically maintainable rule base stores all predefined quality inspection operators and conditions. Administrators can add, modify, or disable quality inspection rules without coding through a graphical interface, thereby quickly responding to new safety regulations or service standards issued by the gas company and greatly improving the system's adaptability.

[0072] 4. Quality Inspection Task: This task is designed for a specific quality inspection objective and consists of a set of quality inspection rules and execution parameters. It defines all content to be inspected and its rules in a single quality inspection analysis. Parameters include: applicable business scenarios (e.g., "installation consultation," "fault reporting"), sampling ratio (100% full or partial), execution frequency (real-time or post-event), etc.

[0073] IV. System Workflow and Decision-Making Mechanism A complete quality inspection process is as follows: 1. Data Trigger: When a customer service call ends or an online chat is completed, the system automatically triggers a preset quality inspection task.

[0074] 2. Vectorization and Feature Extraction: The dialogue text is fed into the system and, after preprocessing, is transformed into semantic vectors by the M3E-Base model.

[0075] 3. Rule Engine Execution: The engine loads all quality inspection conditions associated with the task and executes the quality inspection operators contained in each condition in sequence.

[0076] 4. Multi-dimensional comprehensive decision-making: The engine does not rely solely on single vector similarity, but combines vector matching results, keyword hits, sentiment analysis, silence duration, and other multi-dimensional features to make a comprehensive judgment based on the logic of the conditions. For example, even if the vector matching degree is high, if the customer service representative engages in inappropriate silence for an extended period of time, they may still be judged as unqualified.

[0077] 5. Results Generation and Feedback: Generate Quality Inspection Report: The system automatically generates a quality inspection report that includes detailed deductions, location of violations, and improvement suggestions, and can also be rated by customer service personnel.

[0078] Real-time alerts: For triggered red line conditions (such as condition A), the system can send alerts to the supervisor in real time so that timely intervention can be carried out.

[0079] Model self-optimization: Quality inspection results and feedback from manual review are recorded and used to periodically perform incremental learning on the M3E-Base model, optimizing its performance in the specific business of the gas company, forming a closed loop that gets smarter the more it is used.

[0080] V. Summary of Technical Advantages This technical solution introduces M3E-Base vector matching technology and designs a highly flexible and configurable rule engine, enabling gas customer service quality inspection to leap from "keyword matching" to "semantic understanding", from "fixed rules" to "dynamic adaptation", and from "sampling inspection" to "full analysis". It effectively solves the various defects mentioned in the background technology and significantly improves the accuracy, efficiency and business value of quality inspection.

[0081] The table below clearly compares the differences between the traditional method and the solution of this invention in terms of key performance indicators, and intuitively demonstrates the beneficial effects of this invention.

[0082]

[0083] .

[0084] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0085] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0086] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the gas intelligent quality inspection methods based on vector matching and dynamic rule engine in the above embodiments.

[0087] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0088] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0089] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0090] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0091] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A gas intelligent quality inspection method based on vector matching and dynamic rule engine, characterized in that, The process includes the following steps: S1. Collecting and preprocessing multi-source data, including voice call data, online chat logs, work order texts, and user feedback forms; cleaning the collected audio and text data to complete data preprocessing; S2. Converting the preprocessed text data into high-dimensional vectors to support quality inspection rule matching based on semantic similarity operators; S3. Executing quality inspection tasks through a configurable quality inspection rule engine, which supports hierarchical configuration of quality inspection operators, quality inspection conditions, quality inspection rules, and quality inspection tasks. S4. Manually review the quality inspection results, correct erroneous data, and use the corrected data to fine-tune the semantic understanding model in order to improve the recognition accuracy of the semantic similarity operator.

2. The intelligent gas quality inspection method based on vector matching and dynamic rule engine according to claim 1, characterized in that: S2 specifically includes vectorization and semantic understanding steps. It uses the M3E-Base text embedding model optimized for Chinese to convert the preprocessed customer service text into a 1024-dimensional dense vector. The M3E-Base model is fine-tuned using a corpus containing a large number of gas industry terms. The generated sentence vectors are stored in a vector database and associated with the metadata of the original dialogue.

3. The intelligent gas quality inspection method based on vector matching and dynamic rule engine according to claim 2, characterized in that: The quality inspection operators in S3 include: keyword operator: performing precise word matching in the transcribed text of the dialogue; regular expression operator: performing string matching with complex rules and fuzzy patterns in the dialogue text; speech rate detection operator: analyzing the customer service representative's voice and measuring their average speaking speed; volume detection operator: analyzing the customer service representative's voice and measuring the average level and fluctuation of their voice loudness; and semantic similarity operator: determining whether the customer service representative's or customer's statements are similar to the target statements in meaning and intent, which is the core of intelligent quality inspection.

4. The intelligent gas quality inspection method based on vector matching and dynamic rule engine according to claim 2, characterized in that: The quality inspection conditions are: complex judgment logic composed of multiple quality inspection operators combined through logical operators; one condition represents a complete business scenario judgment.

5. The intelligent gas quality inspection method based on vector matching and dynamic rule engine according to claim 2, characterized in that: The quality inspection rules are: a dynamically maintainable rule base that stores all predefined quality inspection operators and conditions; administrators can add, modify, or disable quality inspection rules through a graphical interface without writing any code.

6. The intelligent gas quality inspection method based on vector matching and dynamic rule engine according to claim 2, characterized in that: The quality inspection task includes: targeting a specific quality inspection objective, consisting of a set of quality inspection rules and execution parameters; it defines all the contents and rules to be checked in a quality inspection analysis; the parameters include: applicable business scenarios, sampling ratio, and execution frequency.

7. The intelligent gas quality inspection method based on vector matching and dynamic rule engine according to any one of claims 1-6, characterized in that: The execution process includes the following steps: Data Triggering: When a customer service call ends or an online chat is completed, the system automatically triggers a preset quality inspection task; Vectorization and Feature Extraction: The dialogue text is sent to the system, preprocessed, and then converted into semantic vectors by the M3E-Base model; Rule Engine Execution: The engine loads all quality inspection conditions associated with the task and executes the quality inspection operators contained in each condition in sequence; Multi-Dimensional Comprehensive Decision Making: The engine does not only rely on a single vector similarity, but also combines multi-dimensional features such as vector matching results, keyword hits, sentiment analysis, and silence duration to make a comprehensive judgment according to the logic of the conditions.

8. The intelligent gas quality inspection method based on vector matching and dynamic rule engine according to claim 7, characterized in that: It also includes result generation and feedback: generating quality inspection reports: the system automatically generates quality inspection reports containing detailed deduction items, location of violations, and improvement suggestions, and can also score customer service personnel; Real-time alerts: When red line conditions are triggered, alerts are sent to quality inspectors in real time.

9. The intelligent gas quality inspection method based on vector matching and dynamic rule engine according to claim 8, characterized in that: Model self-optimization: Quality inspection results and feedback from manual review are recorded and used to periodically perform incremental learning on the M3E-Base model, optimizing its performance in the specific business of the gas company, forming a closed loop that gets smarter the more it is used.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 9.

Images