A method and system for real-time assistance and work order linkage in power customer service

Abstract

This invention discloses a method and system for real-time assistance and work order linkage in power customer service. The method includes: real-time acquisition of voice data from incoming calls and real-time transcription to generate translated text, simultaneously extracting voice emotion features; semantic recognition of the translated text to extract business requests, combining the translated text and voice emotion features to perform risk analysis using a multi-dimensional risk warning algorithm and outputting a risk level; automatic determination and labeling of customer type and risk tags based on the translated text and risk level; automatic generation and filling of work orders based on business requests, customer type, and risk tags, and association of work orders with the translated text and voice data. This invention, through the deep integration of real-time voice transcription and risk warning, solves the problems of high transcription latency, incomplete risk warning coverage, and low system collaboration efficiency in existing technologies, achieving fully automated closed-loop management and significantly improving service efficiency, risk control capabilities, and customer satisfaction.

Description

Technical Field

[0001] This invention relates to the field of digital data processing technology, and more specifically, to a method and system for real-time assistance and work order linkage in power customer service. Background Technology

[0002] As power grid services upgrade towards real-time response and predictive services, real-time voice translation and risk warning have become core technical requirements. Real-time voice translation enables customer service to quickly obtain requests, while risk warnings intercept service complaints in advance and standardize service behavior.

[0003] Currently, many power grid service-related systems only integrate basic call center functions, making it difficult to improve service efficiency and quality. This is specifically reflected in the following aspects:

[0004] First, the adaptability of real-time voice transcription technology is insufficient, making it difficult to support efficient service. Existing systems generally adopt an "offline batch transcription" mode, requiring an average wait of 3 to 5 minutes after the call ends to generate the transcribed text. This prevents customer service personnel from accessing the transcribed content in real time during the call. When faced with complex requests or poor call quality, customer service personnel have to repeatedly ask customers, increasing the duration of a single call by 40%, which in turn affects customer satisfaction and proactive service capabilities. In addition, there is a disconnect between the transcription results and the business systems. Real-time linkage with the work order system and customer profile system is not achieved, requiring customer service personnel to manually copy information. Manual data entry accounts for 60% of the work order filling workload and is prone to information omissions.

[0005] Secondly, the lack of a risk warning mechanism makes it difficult to manage service risks in a timely manner. Current risk warning systems, such as those targeting customer complaint tendencies and customer service attitudes, rely entirely on manual sampling of call recordings, with a sampling rate of less than 30%. This results in 60% of service risks being missed across all calls. This often leads to reactive handling of issues only after a customer complaint, increasing complaint processing costs. Furthermore, the warning system is limited in its scope and predictive capabilities, relying solely on text keywords (such as "complaint" and "dissatisfaction") to assess risk without considering emotional characteristics of the voice (such as sudden increases in speech rate, volume, and pause duration). This makes it impossible to identify potential risks where "the customer hasn't explicitly said 'complaint,' but is already angry." The warning lead time is less than 3 seconds, missing the optimal intervention window. Moreover, even when risks are identified, such as when customers explicitly express dissatisfaction, they are not automatically labeled "high-risk" in the work order. This results in the work order processing priority not being adjusted, leading to an average processing time of 2 hours for high-risk work orders, 50% slower than regular work orders.

[0006] Finally, existing technologies often focus on optimizing single aspects, such as work order allocation and semantic recognition, while lacking the construction of a complete technical system. Customer type classification relies on manual labeling, with an error rate greater than 15%; there is no unified basis for determining call priority; and work order submission lacks dynamic validation logic. These problems result in a learning curve of more than 3 months for new customer service personnel, and service quality fluctuations exceeding 20%. Due to the lack of a complete technical system encompassing "real-time voice translation, multi-dimensional risk warning, and intelligent work order linkage," existing technologies either neglect the real-time nature and terminology adaptability of translation or lack comprehensive risk warning coverage and real-time intervention capabilities. This leads to problems such as "slow response, error-proneness, and risk omissions" in power grid services, urgently requiring a targeted and systematic technical solution. Summary of the Invention

[0007] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method and manufacturing method for real-time assistance and work order linkage for power customer service.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] In a first aspect, the present invention provides a method for real-time assistance and work order linkage in power customer service, comprising:

[0010] The system collects voice data from incoming calls in real time, performs real-time translation of the voice data to generate translated text, and simultaneously extracts voice emotion features.

[0011] Semantic recognition is performed on the translated text to extract business requirements. Combining the translated text and voice emotion features, a multi-dimensional risk warning algorithm is used to perform risk analysis and output the risk level.

[0012] Based on the translated text and risk level, the customer type and risk label are automatically determined and labeled;

[0013] Based on the business requirements, customer type, and risk tags, work orders are automatically generated and populated, and the work orders are associated with the translated text and voice data.

[0014] This invention realizes the intelligent upgrade of the power customer service process. By integrating real-time voice translation and multi-dimensional risk warning, it improves the speed and accuracy of service response and effectively shortens the call and work order processing time. Through full-chain data linkage and automated work order processing, it reduces the reliance on manual labor and the risk of information omission, and comprehensively enhances the ability to provide early warning and proactive intervention for service risks, thereby improving the efficiency, quality and customer satisfaction of power grid services as a whole.

[0015] Preferably, the voice data of incoming calls is collected in real time, translated, and the translated text is generated. Simultaneously, voice emotion features are extracted from the voice data, including:

[0016] An end-to-end incremental transcription model is used to segment the speech data according to a preset frame length. After compressing the model using INT8 quantization technology, each frame of speech data is independently translated and the translated text is output in streaming.

[0017] Multiple acoustic features are extracted from the speech data simultaneously. After standardization, the acoustic features are associated with the translated text and stored according to timestamps. The acoustic features include speech rate, volume, fundamental frequency, and pause duration.

[0018] Preferably, the step of performing semantic recognition on the translated text to extract business requirements, combining the translated text and voice emotion features, performing risk analysis through a multi-dimensional risk warning algorithm, and outputting a risk level includes:

[0019] A deep learning-based semantic recognition model is used, combined with a knowledge base in the power industry, to extract business requirements from the translated text;

[0020] A risk warning algorithm integrating a rule engine and a deep learning model is constructed to match keywords and semantic rules in the translated text, and to perform deep learning analysis by integrating the semantic vector of the translated text with the speech emotion features.

[0021] The final risk level is determined by combining the rule matching results with the risk probabilities derived from deep learning analysis.

[0022] Preferably, the deep learning-based semantic recognition model is a BERT-BiLSTM-CRF model; the BERT-BiLSTM-CRF model is composed of BERT layers, BiLSTM layers, and CRF layers connected sequentially, wherein:

[0023] The BERT layer is a pre-trained language model layer used to convert translated text into fixed-dimensional word vectors, and then extract deep semantic features of the translated text through a pre-trained Transformer encoder. Furthermore, the pre-trained language model layer is fine-tuned using a power-specific corpus to enhance the semantic understanding of power-related professional terms.

[0024] The BiLSTM layer is a bidirectional long short-term memory network layer used to receive deep semantic features output by the BERT layer. It captures the sequence dependencies from left to right in the translated text through forward LSTM units and captures the sequence dependencies from right to left in the translated text through backward LSTM units. After fusing the bidirectional sequence dependency features, it outputs the contextual association features of the text sequence.

[0025] The CRF layer is a conditional random field layer, used to receive the contextual features output by the BiLSTM layer and the preset power business request label transition probability matrix. By calculating the globally optimal path of the sequence labels, it outputs the power business request entity label sequence corresponding to the translated text, thus completing the extraction of power business requests.

[0026] Preferably, based on the translated text and risk level, the customer type and risk label are automatically determined and labeled, including:

[0027] Obtain customer profile information and historical work order records;

[0028] A weighted decision model is established, which uses translated text, customer file information and historical work order records as multi-source input features, and assigns preset weights to each input feature for weighted fusion calculation;

[0029] Customer types are dynamically determined based on the weighted fusion calculation results, and corresponding risk labels are automatically applied based on the risk level.

[0030] Preferably, based on the business requirements, customer type, and risk tags, a work order is automatically generated and populated, and the work order is associated with the translated text and voice data, including:

[0031] Based on preset field mapping rules, work order field information is intelligently extracted from the translated text and automatically filled into the corresponding work order fields;

[0032] The customer type and risk label are automatically linked to the work order, and the work order processing priority is dynamically adjusted according to the risk level.

[0033] Establish a two-way traceability link between work orders and corresponding translated text and call voice data.

[0034] Preferably, the method further includes:

[0035] Regularly collect the accuracy statistics of the real-time translation and the early warning success rate statistics of the multi-dimensional risk early warning algorithm;

[0036] Based on the accuracy statistics and early warning success rate statistics, the model parameters and judgment rules of the real-time translation algorithm and the multi-dimensional risk early warning algorithm are iteratively optimized.

[0037] Secondly, the present invention provides a system for real-time assistance and work order linkage for power customer service, comprising:

[0038] The voice processing module is used to collect voice data from incoming calls in real time, translate the voice data in real time to generate translated text, and extract voice emotion features simultaneously.

[0039] The risk warning module is used to perform semantic recognition on the translated text to extract business requirements, combine the translated text and voice emotion features, perform risk analysis through a multi-dimensional risk warning algorithm, and output the risk level.

[0040] The label determination module is used to automatically determine and label customer type and risk label based on the translated text and risk level;

[0041] The work order linkage module is used to automatically generate and populate work orders based on the business requirements, customer type and risk tags, and associate the work orders with the translated text and voice data.

[0042] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein,

[0043] The memory is used to store programs;

[0044] The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the method for real-time assistance and work order linkage for power customer service as described in the first aspect embodiment of the present invention.

[0045] Fourthly, the present invention also provides a computer-readable storage medium for storing a computer-readable program or instruction, which, when executed by a processor, enables the implementation of the steps in the method for real-time assistance and work order linkage for power customer service as described in the first aspect embodiment of the present invention.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] (1) The present invention adopts an end-to-end incremental transcription model and INT8 quantization technology to realize real-time translation of voice data. Customer service can obtain accurate translation content in real time during the call, reduce the amount of manual entry of work orders and information omission, shorten the duration of a single call, and solve the problems of high latency, poor adaptability and disconnection from business systems in the existing technology.

[0048] (2) This invention constructs a risk warning algorithm that integrates a rule engine and a deep learning model. It combines keywords, semantic rules, and voice emotion features of the translated text for comprehensive analysis, achieving 100% full coverage of calls. The risk warning time is significantly increased, allowing customer service to intervene in a timely manner and avoiding passive handling of problems only after customer complaints. In addition, risk tags are automatically bound to work orders, improving the priority of high-risk work orders and greatly reducing the average processing time.

[0049] (3) By establishing a full-chain intelligent judgment and work order linkage mechanism, this invention realizes the full-process automated closed-loop management from voice translation, risk warning to work order generation and processing, solves the problem of low collaborative efficiency caused by the existing technology focusing on the optimization of a single link, effectively improves the efficiency of work order processing and the consistency of service quality, and reduces operating costs. Attached Figure Description

[0050] Figure 1 Flowchart of the method for real-time assistance and work order linkage for power customer service provided by the present invention;

[0051] Figure 2 A flowchart of the multi-dimensional risk warning process provided by this invention;

[0052] Figure 3 The following is a flowchart illustrating the logic of customer tag determination and profile generation provided by this invention;

[0053] Figure 4 The flowchart for work order generation and information entry provided by this invention;

[0054] Figure 5 The overall flowchart of intelligent judgment and work order linkage provided by the present invention;

[0055] Figure 6 A structural block diagram of the system for real-time assistance and work order linkage for power customer service provided by the present invention;

[0056] Figure 7 This is a structural block diagram of the electronic device provided by the present invention. Detailed Implementation

[0057] This invention aims to design a method for real-time assistance and work order linkage for power customer service, solving problems such as slow convergence speed, uncontrollable optimization effect, and low material utilization in existing pipe layout technology. It aims to meet the practical application needs of modern processing for improving material utilization and reducing layout time, and has strong practicality and application prospects. The following will be described and introduced in conjunction with the accompanying drawings through multiple embodiments.

[0058] The naming or numbering of steps in the embodiments of the present invention does not mean that the steps in the method flow must be executed in the time / logical order indicated by the naming or numbering. The execution order of the named or numbered process steps can be changed according to the technical purpose to be achieved, as long as the same or similar technical effect can be achieved.

[0059] Figure 1 A flowchart illustrating the method for real-time assistance and work order linkage for power customer service provided by this invention. (Refer to...) Figure 1 The method for real-time assistance and work order linkage for power customer service provided in this embodiment includes:

[0060] Step S1: Collect voice data of incoming calls in real time, translate the voice data in real time to generate translated text, and extract voice emotion features simultaneously.

[0061] Specifically, once a caller connects with customer service, the system activates the voice capture module to continuously capture voice data during the call. Simultaneously, it initiates real-time translation, converting the captured voice data into transcribed text during the call and pushing it to the customer service interface in real time. During this process, the system also extracts features reflecting emotional states from the voice data and associates these features with the corresponding transcribed text, ensuring data traceability.

[0062] Step S2: Semantic recognition is performed on the translated text to extract business requirements. Combining the translated text and voice emotion features, a multi-dimensional risk warning algorithm is used to perform risk analysis and output the risk level.

[0063] It should be noted that the existing system can only perform shallow keyword matching in terms of semantic recognition, and cannot accurately understand the complex demands of power business. In terms of risk warning, it relies entirely on manual sampling of some call recordings and makes judgments based on only a few text keywords, without fully considering the emotional factors of customers during the call. However, emotional state has a significant impact on service risk. Angry and impatient customers are more likely to cause conflicts and complaints, and relying solely on text analysis can easily miss such risks.

[0064] In view of this, the present invention combines translated text and voice emotion features, and employs a multi-dimensional risk warning algorithm for risk analysis. It utilizes text semantic recognition to clarify customer business needs, and leverages voice emotion features to understand customer emotional states, comprehensively assessing risks from both semantic and emotional dimensions. This improves the accuracy and comprehensiveness of risk assessment and effectively addresses the problem of one-sided risk analysis in existing technologies.

[0065] Step S3: Based on the translated text and risk level, automatically determine and label the customer type and risk tag.

[0066] In existing technologies, customer type classification typically relies on manual judgment and labeling by customer service personnel. This lacks unified standards, results in a high error rate, and fails to dynamically incorporate risk information from real-time calls, leading to unreasonable allocation of subsequent service resources. To address this issue, this invention automatically determines and labels customer types and risk tags based on translated text and risk levels. By comprehensively analyzing the semantic information and risk level of the current customer call, it dynamically adjusts customer type determination. Simultaneously, it accurately labels risk based on business needs and risk levels, improving the accuracy and relevance of customer classification and risk labeling, and providing strong support for subsequent differentiated services.

[0067] Step S4: Based on the business requirements, customer type, and risk tags, automatically generate and populate work orders, and associate the work orders with the translated text and voice data.

[0068] Existing technologies suffer from inefficiency and incomplete information in work order generation and association. On the one hand, work order generation relies on manual input, which is slow, error-prone, and unable to quickly respond to customer needs. For example, when customer service personnel manually fill out work orders, they may overlook important information, affecting subsequent processing. On the other hand, the lack of effective association between work orders and original voice data and transcribed text makes subsequent queries and tracing difficult, hindering service quality monitoring and problem analysis.

[0069] To address the aforementioned issues, this invention automatically generates and populates work orders based on business requirements, customer types, and risk tags, and associates these work orders with translated text and voice data. This automates and intelligently generates work orders, improving efficiency and accuracy. Simultaneously, it establishes complete data associations, facilitating subsequent queries, traceability, and statistical analysis, thereby enhancing overall service quality and management level.

[0070] This invention realizes the intelligent upgrade of the power customer service process. By integrating real-time voice translation and multi-dimensional risk warning, it improves the speed and accuracy of service response and effectively shortens the call and work order processing time. Through full-chain data linkage and automated work order processing, it reduces the reliance on manual labor and the risk of information omission, and comprehensively enhances the ability to provide early warning and proactive intervention for service risks, thereby improving the efficiency, quality and customer satisfaction of power grid services as a whole.

[0071] In some embodiments of this application, step S1 involves real-time acquisition of voice data from an incoming call, real-time transcription of the voice data to generate transcribed text, and simultaneous extraction of voice emotion features, specifically including:

[0072] S11 employs an end-to-end incremental transcription model, which segments speech data according to a preset frame length, compresses the model using INT8 quantization technology, and then independently translates each frame of speech data, outputting the translated text in a streaming manner.

[0073] Specifically, the ITT (Incremental Transformer Transducer) model is used as the core translation model. The real-time collected call audio data is first divided into independent audio segments with a preset frame length of 200 milliseconds per frame, avoiding the waiting delay caused by traditional whole-sentence translation. Here, the preset frame length refers to a pre-defined audio stream segmentation unit; each frame of audio data is translated independently, avoiding delays caused by waiting to process entire audio segments.

[0074] Next, the ITT model is compressed using INT8 quantization technology, reducing its size by 60% and ensuring that the translation time for a single frame of speech data is less than 5 milliseconds, with the overall inference latency controlled within 100 milliseconds per frame. INT8 quantization technology refers to 8-bit integer quantization, which converts model parameters from high-precision to 8-bit integers to compress the model size while maintaining translation accuracy.

[0075] During the translation process, a pre-built power-specific corpus is used to adapt the ITT model. This power-specific corpus is a collection of speech and text data built specifically for the power sector. It contains 150,000 data entries, 30,000 of which are power-related technical terms, categorized by power generation, transmission, distribution, and consumption. A terminology embedding layer is used to fine-tune the ITT model, enhancing its ability to recognize power-related technical terms such as photovoltaic inverters, distribution transformers, and feeder automation terminals, improving the accuracy of power terminology recognition to 98%. Simultaneously, the corpus is updated monthly, adding no fewer than 500 new terms each month (such as virtual power plant peak shaving) to ensure the model's adaptability to new power-related businesses.

[0076] The translated text is streamed to the customer service interface in real time, segment by segment, allowing customers to view it while on a call. Simultaneously, the translated text is synchronized in real time to the ticketing system and customer profile system via the WebSocket protocol and port 5050. The WebSocket 5050 port is a network interface that supports bidirectional real-time communication. Each translated text is bound to the caller ID, call timestamp, and area code extracted from the customer profile system, ensuring traceability and enabling seamless integration with the business systems without manual copying.

[0077] S12, Simultaneously extract multiple acoustic features from the speech data, standardize the acoustic features, and then associate and store them with the translated text according to timestamps; wherein, the acoustic features include speech rate, volume, fundamental frequency, and pause duration.

[0078] While performing speech translation in step S11, acoustic features are simultaneously extracted from the real-time acquired speech data based on the Librosa framework. These features include speech rate, volume, fundamental frequency, and pause duration. Librosa is an open-source framework for audio signal processing, used in this invention to extract acoustic features from the speech data. Speech rate is the number of effective words per second after removing pauses; volume is the short-time energy average; fundamental frequency is a feature reflecting pitch changes within the range of 5Hz-500Hz; and pause duration is the effective duration of a single pause exceeding 500 milliseconds.

[0079] The extracted acoustic features were processed using Z-score normalization to convert the feature data into standard data with a mean of 0 and a standard deviation of 1. This data was then further processed into a 32-dimensional vector to ensure a consistent feature data format for easy use by subsequent risk warning algorithms. Z-score normalization is a data standardization method that converts data into standard data with a mean of 0 and a standard deviation of 1 by calculating the deviation of the data from the mean and dividing by the standard deviation.

[0080] The processed acoustic feature vectors and translated texts are associated one-to-one with each other according to the call timestamp and stored together in the Milvus vector database to form a "text-emotion feature-timestamp" associated dataset, providing non-textual data support for subsequent multi-dimensional risk warning.

[0081] This embodiment addresses the issues of poor recognition of specialized terminology in general models and high latency in domain-specific adaptation schemes. Low-latency translation is achieved through the ITT incremental transcription model and INT8 quantization technology, allowing customer service representatives to simultaneously access the text. A dedicated corpus for the power industry is constructed and fine-tuned through terminology embedding layers, improving the accuracy of power terminology to 98%. Simultaneously, voice emotion features are extracted and standardized, providing comprehensive data support for subsequent services and effectively improving the quality and efficiency of power customer service.

[0082] In some embodiments of this application, step S2, which involves semantic recognition of the translated text to extract business requirements, combining the translated text and voice emotion features, performing risk analysis through a multi-dimensional risk warning algorithm, and outputting a risk level, includes:

[0083] S21 employs a deep learning-based semantic recognition model, combined with a knowledge base in the power sector, to extract business requirements from translated text.

[0084] The semantic recognition model based on deep learning is the BERT-BiLSTM-CRF model. After the translated text is generated in step S1, the BERT-BiLSTM-CRF model is started for semantic recognition. The BERT-BiLSTM-CRF model is composed of BERT layers, BiLSTM layers, and CRF layers connected sequentially, wherein:

[0085] The BERT (Bidirectional Encoder Representations from Transformers) layer is a pre-trained language model layer. It is used to convert translated text into fixed-dimensional word vectors and then extract deep semantic features of the translated text through a pre-trained Transformer encoder. Furthermore, the pre-trained language model layer is fine-tuned using a power industry-specific corpus to enhance the semantic understanding of power industry terminology.

[0086] The BiLSTM (Bidirectional Long Short-Term Memory) layer is a bidirectional long short-term memory network layer used to receive deep semantic features output by the BERT layer. It captures the sequence dependencies from left to right in the translated text through forward LSTM units and captures the sequence dependencies from right to left in the translated text through backward LSTM units. After fusing the bidirectional sequence dependency features, it outputs the contextual association features of the text sequence.

[0087] The CRF (Conditional Random Field) layer is a conditional random field layer used to receive the contextual features output by the BiLSTM layer and the preset power business request label transition probability matrix. By calculating the globally optimal path of the sequence labels, it outputs the power business request entity label sequence corresponding to the translated text, thus completing the extraction of power business requests.

[0088] S22, Construct a risk warning algorithm that integrates a rule engine and a deep learning model, perform keyword and semantic rule matching on the translated text, and integrate the semantic vector of the translated text with the voice emotion features for deep learning analysis.

[0089] In this embodiment, the rule engine is a system module that quickly filters risk conditions based on preset rules (such as keyword matching and regular expressions), and can achieve real-time matching of illegal terms and complaint keywords.

[0090] Figure 2 The flowchart of the multi-dimensional risk warning process provided by this invention is as follows: Figure 2As shown, after voice recognition, two branches are triggered simultaneously: semantic recognition and sentiment analysis. In the semantic recognition branch, the rule engine matches keywords with semantic rules in the translated text. Specifically, service attitude risk is assessed by matching customer service inappropriate phrases such as "I don't know" and "I can't do it" using regular expressions; a rule is triggered when such phrases appear twice or more. Customer complaint tendency risk is assessed by matching keywords such as "complaint" and "dissatisfaction" (triggered ≥1 time); business error risk is assessed by comparing the semantic similarity between customer requests and customer service responses using a cosine similarity algorithm (triggered <80%). Simultaneously, semantic vectors are extracted from the translated text using BERT. The sentiment analysis branch extracts acoustic features such as speech rate, volume, fundamental frequency, and pause duration from the speech data, standardizing them before using them as input. In the deep learning analysis phase, the service attitude risk is assessed using a CNN-LSTM model, which takes the customer service voice features and the translated text semantic vector as input and outputs the attitude risk probability (≥0.7 for confirmed risk). The customer complaint tendency risk is assessed using a LightGBM model, which integrates customer voice emotion features and semantic vectors; a probability ≥0.65 marks a high risk. The business error risk is assessed using a BiLSTM model, which identifies information deviations in customer service responses (such as incorrect station numbering); a probability ≥0.8 triggers a warning. Risk scores are updated synchronously with the voice frame every 200ms, and the results are combined with the rules.

[0091] S23. Combine the rule matching results with the risk probabilities derived from deep learning analysis to determine the final risk level.

[0092] Specifically, based on the combined rule matching results and deep learning risk probabilities, the system determines whether a risk condition has been triggered: if the risk score is greater than or equal to 0.7, the system locates the corresponding voice keyframes and sensitive words, labels them as high, medium, or low risk levels, and simultaneously displays a risk warning pop-up on the customer service interface, pushing targeted response strategies. For example, for service attitude risks, the system pushes reassuring script templates; for complaint risks, it pushes the issue escalation process, and synchronizes this information with the customer service supervisor's backend. If no risk condition has been triggered, the system marks the status as risk-free. Finally, the system generates a full summary report, which is a comprehensive summary report of the call content, identified risk situations, and handling suggestions, providing a basis for subsequent service optimization.

[0093] This invention, through the construction of a risk warning system integrating a deep-fusion rule engine and a multi-layered deep learning model, achieves accurate identification and real-time warning of service attitude risks, customer complaint tendency risks, and business error risks. By combining parallel processing semantic analysis and emotion recognition processes, it solves the problems of incomplete risk coverage, high warning delay, and high misjudgment rate of traditional methods, thereby improving the risk management capabilities and service quality of power customer service.

[0094] In some embodiments of this application, step S3, based on the translated text and risk level, automatically determines and labels the customer type and risk tag, including:

[0095] S31. Obtain customer profile information and historical work order records; establish a weighted judgment model, using translated text, customer profile information, and historical work order records as multi-source input features, and assign preset weights to each input feature for weighted fusion calculation.

[0096] Customer profile information is a structured data set recording customer electricity usage attributes, basic information, and historical service characteristics; historical work order records refer to work order data for past electricity services, including request type, processing result, and other information. The weighted judgment model assigns preset weights to multi-source input features such as translated text, customer profile information, and historical work order records, integrating multi-dimensional data to determine customer type. Customer profiles are comprehensive descriptions of customer characteristics formed by integrating multi-dimensional information such as customer type, service priority, work order type, and service risk.

[0097] Figure 3 The following is a flowchart illustrating the logic of customer tag determination and profile generation provided by this invention: Figure 3 As shown, after the call begins, scenario matching is performed first, and the system automatically identifies the service scenario type based on the call context. Then, the system enters the rule judgment stage, calling a pre-set multi-dimensional judgment rule library. In the rule judgment stage, the system executes three core judgments in parallel: customer type judgment, based on the weighted judgment model output, automatically labels the customer as such, for example, a new report customer, a reset customer, or a special customer, and adds corresponding status markers; task priority judgment, comprehensively considering risk level and business needs, divides into three priority levels: urgent, high, and routine, and associates them with work order types such as fault, consultation, and business processing, while also integrating service risk and complaint warning information; and anomaly cause judgment records key information such as fault details, workday recorder, and risk status.

[0098] In specific implementation, as described in step S31, the system first obtains customer profile information and historical work order records, establishes a weighted decision model, uses real-time translated text, customer profile information and historical work order records as multi-source input features, and assigns preset weights to each feature, such as translated text weight 0.5, customer profile information weight 0.3, and historical work order records weight 0.2. The confidence score of each customer type is calculated through a weighted fusion algorithm.

[0099] S32 dynamically determines the customer type based on the weighted fusion calculation results and automatically labels the corresponding risk level.

[0100] Specifically, customer types are dynamically determined based on weighted fusion calculation results, and corresponding risk labels are automatically assigned based on risk levels. This process can follow... Figure 3The customer tagging and profile generation logic flow shown is as follows: When a call begins, the system first performs scenario matching, then triggers the judgment rule mechanism. During the rule judgment phase, the system simultaneously conducts a comprehensive analysis across four dimensions: Customer type dimension: Based on the confidence score output by the weighted judgment model, the system accurately classifies customers into categories such as new landline customers, important customers, and special group customers, and simultaneously completes corresponding status marking; Service priority dimension: By comprehensively considering customer type and risk level, the system clearly divides service priority into three levels: urgent, high, and routine, while triggering different colored pop-up icons for visual reminders; Work order type dimension: Based on the business request characteristics extracted from the translated text, work orders are subdivided into specific types such as fault, consultation, and business processing, achieving automated entry of core work order fields; Service risk dimension: Combining the risk level data output in step S2, the system accurately labels service risks, complaint warnings, and other risk tags and fully records risk status information. Finally, by integrating the judgment results of these four dimensions, the system generates a complete customer profile that includes customer type characteristics, service priority settings, work order attribute descriptions, and risk status assessments.

[0101] This invention utilizes a weighted judgment model to fuse multi-source input features, achieving dynamic customer type determination and automatic risk labeling. The accuracy of customer type determination is improved to over 98%, and the false judgment rate for call priority is reduced to <5%. Combined with risk level labeling, the processing priority of high-risk work orders is increased by 50%, and the average processing time is reduced to 40 minutes. The customer profile generation logic covers multi-dimensional information, achieving end-to-end data linkage and automated processing, effectively reducing reliance on manual labor and improving the efficiency and quality of power grid services.

[0102] In some embodiments of this application, step S4 involves automatically generating and filling a work order based on the business request, customer type, and risk label, and associating the work order with the translated text and voice data, including:

[0103] S41, based on preset field mapping rules, intelligently extract work order field information from the translated text and automatically fill it into the corresponding work order field.

[0104] Specifically, the process begins by establishing a correspondence between "translated text keywords and work order segments." For example, "faulty equipment: photovoltaic inverter" corresponds to the "faulty equipment type" field in the work order, and "station number 102" corresponds to the "fault location" field. Then, a CRF entity recognition model is used to extract core work order fields from the translated text, including customer name, request type, and fault information. This model achieves an extraction accuracy of ≥95%. Automatic work order generation is triggered within 5 seconds of the call ending. In addition to the core fields, the filled content includes the customer type from the rule base, the request type from semantic recognition, a 50-100 word summary of key call content, and the risk level from risk rules. After automatic filling, customer service only needs to verify the <10% of "unrecognized fields." Customer service can modify the automatically filled content, and modification history is recorded in real time. Ultimately, this reduces manual data entry by ≥70% and reduces work order completion time from 5 minutes to 1.5 minutes.

[0105] S42, automatically bind the customer type and risk label to the work order, and dynamically adjust the work order processing priority according to the risk level.

[0106] Risk warning results include high risk of customer complaints and service attitude risks, which are written into the work order system in real time through application programming interfaces. The risk level is marked before the work order title, such as "High Risk - Fault Reporting". The priority of work order processing is dynamically adjusted according to the risk level. High-risk work orders are automatically put into the priority processing queue and dispatched 50% faster. If there is a complaint risk, a risk warning notification will be pushed simultaneously. At the same time, a high-risk work order processing time monitoring mechanism is established, with the normal processing time set at 40 minutes. A reminder is sent to the processing personnel 10 minutes before the timeout.

[0107] After a work order is completed, the system automatically sends a satisfaction survey SMS to the customer. The survey results are linked to the risk warning records to optimize the warning model. For example, if customer satisfaction is low after a certain type of warning, the model parameters will be adjusted. A risk warning report will also be generated every month, which includes the proportion of various risks, the success rate of warnings, and the handling effect, providing a basis for algorithm iteration. For example, if the proportion of service attitude risk is high, customer service training materials will be supplemented.

[0108] S43, establish a two-way traceability association between work orders and corresponding translated text and call voice data.

[0109] Specifically, in this embodiment, "caller ID (customer's unique identifier) ​​+ timestamp (real-time call marker)" is used as the dual primary key to achieve automatic association of translated text, voice emotion features, risk warning results, customer profiles, and historical work orders without manual intervention; a cross-system data interface standard is established to unify the data formats of translated text (JSON format), risk tags (encoded, 01 represents low risk, 02 represents medium risk, and 03 represents high risk), and work order sub-fields, ensuring real-time data interoperability between modules; for repeated call scenarios, the same customer is matched by caller ID, if 3 If a second inquiry with the same request is made within a day and the semantic similarity is ≥85%, the system automatically aggregates historical work orders from the CRM, work order system, and call center system, displaying the historical processing status in the new work order, including the personnel handling the inquiry, progress, and result. For the two-way traceability function, a "record playback" button is embedded in the work order details page, supporting jumps to the corresponding call time period with an error of ≤5 seconds. The recording playback page is associated with the corresponding work order number; clicking the number will jump to the work order details. Simultaneously, the work order details page also embeds a "risk warning segment," which allows jumps to the specific time period during the call when the warning was triggered, with an error of ≤5 seconds, facilitating subsequent review.

[0110] This embodiment achieves a fully closed-loop work order linkage, reducing manual data entry by ≥70%, significantly shortening work order completion time, and achieving a core field extraction accuracy rate of ≥95%. High-risk work order processing priority is increased by 50%, timeout warnings ensure timeliness, and satisfaction surveys and reports assist in model optimization. Repeat call identification aggregates historical data, bidirectional traceability enables work order and voice linkage, and dual primary keys and interface standardization break down data silos, improving the efficiency of power customer service work order processing, risk management capabilities, and customer satisfaction.

[0111] Figure 4 The flowchart for work order generation and information entry provided by this invention. Figure 4 This demonstrates the complete business process of work order generation and information entry according to the present invention. The process begins after the call is connected, when the system automatically triggers a pop-up window on the work order interface, prompting customer service personnel to begin the work order information entry process.

[0112] During the work order information entry stage, the system first performs a caller ID matching test. If the caller ID matches successfully, the system automatically retrieves the customer's basic information from the customer profile system and fills it into the corresponding fields of the work order, including basic data such as customer name, electricity address, and historical electricity usage. If the caller ID does not match successfully, the system switches to manual entry mode, where customer service personnel manually enter the necessary customer identification information.

[0113] After completing the customer information processing, the system automatically performs the linking operation between historical work orders and recordings. Using a dual primary key consisting of the caller ID and timestamp, the system can accurately link the customer's historical service records, including past work order processing details, historical call recordings, and related service records.

[0114] In the core information filling stage of the work order, the system automatically extracts key business information from the real-time translated text based on preset field mapping rules and a CRF entity recognition model. This information includes essential work order fields such as faulty equipment type, specific fault symptoms, and request classification, achieving an extraction accuracy rate of over 95%. Simultaneously, the system automatically fills in customer type identifiers from the rule base, request type classifications generated based on semantic recognition, a 50-100 character summary of the call content, and a risk level assessment provided by the risk warning system.

[0115] During the final confirmation stage before saving a work order, customer service personnel need to review the automatically populated content, focusing on checking unrecognized fields that account for less than 10%. They can also make necessary modifications to the filled content, and all modifications are recorded by the system in real time. Once the customer service representative confirms the work order information is complete and accurate, they can save the work order. The system then generates a comprehensive report containing work order quality assessment and call analysis data, providing data support for subsequent service optimization. If the work order information needs further improvement, the customer service representative can choose to fill out the work order again and re-enter the information entry process.

[0116] In some embodiments of this application, the method for real-time assistance and work order linkage for power customer service provided by the present invention further includes:

[0117] The accuracy statistics of the real-time translation and the success rate statistics of the multi-dimensional risk warning algorithm are collected periodically; based on the accuracy statistics and the success rate statistics, the model parameters and judgment rules of the real-time translation algorithm and the multi-dimensional risk warning algorithm are iteratively optimized.

[0118] Specifically, based on the collected accuracy and early warning success rate statistics, the system initiates a model optimization process. For the real-time translation algorithm, the optimization process includes, but is not limited to, adjusting the parameters of the end-to-end incremental transcription model, particularly focusing on the continuous improvement of the power industry-specific corpus and the optimization of the terminology embedding layer. By analyzing translation error cases, the system focuses on improving the accuracy of professional terminology recognition and the ability to understand complex business requirements. For the multi-dimensional risk early warning algorithm, based on the early warning success rate statistics, the system simultaneously optimizes the matching rules in the rule engine and the parameter configuration of the deep learning model, including adjusting the network structure of the CNN-LSTM fusion model, optimizing the feature weights of the LightGBM model, and improving the recognition logic of the BiLSTM model.

[0119] Regarding the optimization of judgment rules, the system dynamically adjusts risk judgment thresholds and warning trigger conditions by analyzing the proportion of various risks in the risk warning reports. For example, when the proportion of service attitude risk increases abnormally, the system automatically supplements the corresponding customer service training materials and adjusts the relevant judgment rules; when the customer satisfaction survey results after a certain type of warning are unsatisfactory, the model parameters will be corrected accordingly. At the same time, the system has also established an optimization mechanism based on customer feedback, which conducts in-depth correlation analysis between the satisfaction survey results collected after the work order is completed and the risk warning records, in order to verify and calibrate the accuracy and practicality of the warning model.

[0120] This iterative optimization process operates on a monthly cycle, generating a detailed performance analysis report for each cycle to guide algorithm improvements in the next cycle. This data-driven, continuous optimization mechanism ensures that the system's translation accuracy and risk warning capabilities continuously improve with usage time, creating a virtuous cycle of self-improvement.

[0121] This embodiment establishes a data-driven closed-loop optimization mechanism to continuously improve the real-time translation algorithm and risk warning system, effectively solving the technical problems of traditional systems being rigid, inflexible, and unable to adaptively optimize. By regularly collecting performance statistics and conducting targeted algorithm iterations, the translation accuracy has maintained a continuous upward trend, and the warning success rate has steadily improved. At the same time, by continuously calibrating the optimization direction with customer feedback data, the system's adaptability and long-term service stability have been significantly improved, providing a reliable technical guarantee for the continuous optimization of power grid services. Ultimately, it achieves the system evolution goal of becoming more accurate and intelligent with use.

[0122] In some embodiments of this application, Figure 5 This invention demonstrates the complete business process of a method for real-time assistance and work order linkage in power customer service. The process begins with an incoming call from a customer; the system first determines whether the call was successfully answered. If not answered, a missed call pop-up is triggered, and the process terminates or continues waiting depending on whether the pop-up is closed.

[0123] Upon successful answering of the call, the system immediately activates the scenario matching module, automatically identifying the service scenario type by analyzing the call context. It then proceeds to the rule judgment stage, invoking a pre-defined multi-dimensional rule base to make a preliminary determination of core elements such as customer type and service priority.

[0124] In the intelligent analysis phase, the system performs real-time speech transcription and multi-dimensional risk analysis in parallel. It achieves millisecond-level transcription of speech data through an end-to-end incremental transcription model, while simultaneously extracting emotional features from the speech. Combined with semantic recognition and risk warning algorithms, it accurately analyzes business needs and assesses service risk levels, enabling real-time risk monitoring of all calls.

[0125] During the work order linkage phase, work order information is automatically generated and populated based on the preliminary analysis results. The system automatically extracts core fields of the work order through intelligent field mapping, automatically binds customer type and risk tag to the work order, and dynamically adjusts the processing priority according to the risk level. At the same time, a two-way traceability link is established between the work order and the translated text and voice data to ensure data integrity and traceability.

[0126] Finally, in the report optimization phase, the system generates a comprehensive report containing work order quality assessment and call volume analysis data, providing data support for subsequent algorithm iterations and optimizations. By regularly collecting statistical data on translation accuracy and early warning success rates, the algorithm model and judgment rules are continuously optimized, forming a virtuous cycle of self-improvement.

[0127] The entire process achieves intelligent processing across the entire chain from answering the call to generating the work order. Through the close connection of five core links—scenario matching, rule judgment, intelligent analysis, work order linkage, and report optimization—a complete closed loop for power customer service has been constructed.

[0128] This invention, through the construction of a fully intelligent service system, achieves a transformation and upgrade of power customer service from passive response to proactive prediction. The system deeply integrates real-time voice translation, multi-dimensional risk warning, and intelligent work order linkage, improving service response speed and accuracy, effectively reducing reliance on manual labor and operational errors. A data-driven continuous optimization mechanism ensures the long-term stable operation of the system and the continuous improvement of service quality, providing efficient and reliable technical support for power grid services.

[0129] Figure 6 The structural block diagram of the system for real-time assistance and work order linkage for power customer service provided by the present invention is shown below. Figure 6 The 600-system for real-time customer service assistance and work order linkage for the power industry includes:

[0130] The voice processing module 601 is used to collect voice data of incoming calls in real time, translate the voice data in real time to generate translated text, and extract voice emotion features simultaneously.

[0131] The risk warning module 602 is used to perform semantic recognition on the translated text to extract business requirements, combine the translated text and voice emotion features, perform risk analysis through a multi-dimensional risk warning algorithm, and output the risk level.

[0132] The label determination module 603 is used to automatically determine and label customer type and risk label based on the translated text and risk level;

[0133] The work order linkage module 604 is used to automatically generate and fill work orders based on the business requirements, customer type and risk tags, and associate the work orders with the translated text and voice data.

[0134] The system for real-time assistance and work order linkage of power customer service provided by the present invention is used to execute the methods for real-time assistance and work order linkage of power customer service provided in the foregoing embodiments. The methods for real-time assistance and work order linkage of power customer service have been described in detail in the foregoing embodiments, and will not be repeated here.

[0135] Figure 7 A structural block diagram of the electronic device provided by the present invention, such as Figure 7 As shown, the present invention also provides an electronic device 700, which can be a mobile terminal, desktop computer, laptop, handheld computer, server, or other computing device. The electronic device 700 includes a processor 701 and a memory 702, wherein the memory 702 stores a program 703 for real-time assistance and work order linkage for power customer service.

[0136] In some embodiments, memory 702 may be an internal storage unit of a computer device, such as a hard disk or memory. In other embodiments, memory 702 may be an external storage device of a computer device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, memory 702 may include both internal and external storage units of the computer device. Memory 702 is used to store application software and various types of data installed on the computer device, such as program code for installing the computer device. Memory 702 can also be used to temporarily store data that has been output or will be output. In one embodiment, when the program 703 for real-time assistance and work order linkage of the power customer service is executed by processor 701, the following steps are implemented:

[0137] The system collects voice data from incoming calls in real time, performs real-time translation of the voice data to generate translated text, and simultaneously extracts voice emotion features.

[0138] Semantic recognition is performed on the translated text to extract business requirements. Combining the translated text and voice emotion features, a multi-dimensional risk warning algorithm is used to perform risk analysis and output the risk level.

[0139] Based on the translated text and risk level, the customer type and risk label are automatically determined and labeled;

[0140] Based on the business requirements, customer type, and risk tags, work orders are automatically generated and populated, and the work orders are associated with the translated text and voice data.

[0141] In some embodiments, processor 701 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 702 or process data, such as executing programs for real-time assistance and work order linkage for power customer service.

[0142] This embodiment also provides a computer-readable storage medium storing a program for real-time assistance and work order linkage for power customer service. When the program is executed by a processor, it performs the following steps:

[0143] The system collects voice data from incoming calls in real time, performs real-time translation of the voice data to generate translated text, and simultaneously extracts voice emotion features.

[0144] Semantic recognition is performed on the translated text to extract business requirements. Combining the translated text and voice emotion features, a multi-dimensional risk warning algorithm is used to perform risk analysis and output the risk level.

[0145] Based on the translated text and risk level, the customer type and risk label are automatically determined and labeled;

[0146] Based on the business requirements, customer type, and risk tags, work orders are automatically generated and populated, and the work orders are associated with the translated text and voice data.

[0147] The embodiments described above are merely illustrative of several implementations of the present invention, and while the 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 modifications and improvements 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.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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; and these 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 method for power customer service real-time assistance and work order linkage, characterized in that: The method comprises the following steps: Real-time collection of voice data of an incoming call, real-time translation of the voice data to generate translated text, and synchronous extraction of voice emotion features; including: using an end-to-end incremental transcription model, dividing the voice data according to a preset frame length, independently transcribing each frame of voice data after compressing the model using INT8 quantization technology, and outputting the translated text in a streaming manner; synchronously extracting multiple acoustic features from the voice data, and storing the acoustic features after standardization processing and the translated text according to timestamps; wherein the acoustic features include speech rate, volume, fundamental frequency, and pause duration; Performing semantic recognition on the translated text to extract business demands, combining the translated text and voice emotion features, performing risk analysis through a multi-dimensional risk early warning algorithm, and outputting a risk level; including: using a semantic recognition model based on deep learning to extract business demands from the translated text in combination with a power domain knowledge base; constructing a risk early warning algorithm that integrates a rule engine and a deep learning model, performing keyword and semantic rule matching on the translated text, and performing deep learning analysis by combining the semantic vector of the translated text and the voice emotion features; determining the final risk level by comprehensively considering the rule matching result and the risk probability obtained by deep learning analysis; The semantic recognition model based on deep learning is a BERT-BiLSTM-CRF model; the BERT-BiLSTM-CRF model is composed of a BERT layer, a BiLSTM layer, and a CRF layer connected in sequence, wherein: The BERT layer is a pre-trained language model layer, which is used to convert the translated text into a fixed-dimensional word vector, extract the deep semantic features of the translated text through a pre-trained Transformer encoder, and fine-tune the pre-trained language model layer using a power-specific corpus to enhance the semantic understanding ability of power-specific terms; The BiLSTM layer is a bidirectional long short-term memory network layer, which is used to receive the deep semantic features output by the BERT layer, capture the sequence dependency from left to right in the translated text through a forward LSTM unit, capture the sequence dependency from right to left in the translated text through a backward LSTM unit, and output the context association features of the text sequence after fusing the bidirectional sequence dependency features; The CRF layer is a conditional random field layer, which is used to receive the context association features output by the BiLSTM layer and a preset power business demand label transition probability matrix, calculate the globally optimal path of the sequence label, output the power business demand entity label sequence corresponding to the translated text, and complete the extraction of the power business demand; Based on the translated text and the risk level, automatically determine and label the customer type and the risk label; According to the business demand, the customer type and the risk label, automatically generate and fill in a work order, and associate the work order with the translated text and the voice data. 2.The method of claim 1, wherein: Based on the translated text and the risk level, automatically determine and label the customer type and the risk label, including: Obtain customer profile information and historical work order records; A weighted judgment model is established, and the translated text, customer profile information and historical work order records are taken as multi-source input features, and each input feature is assigned a preset weight for weighted fusion calculation; The customer type is dynamically judged according to the weighted fusion calculation result, and the corresponding risk label is automatically labeled combined with the risk level. 3.The method of claim 1, wherein: According to the business appeal, customer type and risk label, the work order is automatically generated and filled, and the work order is associated with the translated text and voice data, including: Based on the preset field mapping rule, the work order field information is intelligently extracted from the translated text and automatically filled into the corresponding work order field; The customer type and risk label are automatically bound to the work order, and the work order processing priority is dynamically adjusted according to the risk level; The bidirectional traceability association of the work order and the corresponding translated text and call voice data is established. 4.The method of claim 1, wherein: It also includes: Periodically collect the accuracy rate statistical data of the real-time translation and the early warning success rate statistical data of the multi-dimensional risk early warning algorithm; Based on the accuracy rate statistical data and the early warning success rate statistical data, the model parameters and judgment rules of the real-time translation algorithm and the multi-dimensional risk early warning algorithm are iteratively optimized.

5. A system for power customer service real-time assistance and work order linkage, which is used to perform the power customer service real-time assistance and work order linkage method of any one of claims 1-4, characterized in that, The system includes: A voice processing module for real-time collection of voice data of incoming call, real-time translation of the voice data to generate translated text, and synchronous extraction of voice emotion features; A risk early warning module for semantic recognition of the translated text to extract business appeal, combining the translated text and voice emotion features, and performing risk analysis through a multi-dimensional risk early warning algorithm and outputting a risk level; A label judgment module for automatically judging and labeling customer type and risk label based on the translated text and risk level; A work order linkage module for automatically generating and filling a work order according to the business appeal, customer type and risk label, and associating the work order with the translated text and voice data.

6. An electronic device, characterized in that including a memory and a processor, wherein the memory is used to store programs; the processor is coupled with the memory and is used to execute the programs stored in the memory to realize the steps in the method of power customer service real-time assistance and work order linkage according to any one of claims 1 to 4.

7. A computer readable storage medium characterized by A computer readable program or instruction is used to store the steps in the method of power customer service real-time assistance and work order linkage according to any one of claims 1 to 4.

Images