A dispatch telephone scene recognition and log generation method and system based on large language model technology

Abstract

The application provides a dispatch telephone scene recognition and log generation method and system based on a large language model technology, which is applied to a cloud-edge-end collaborative processing system and includes the following steps: through a communication gateway corresponding to a public switched telephone network and a session initiation protocol, two-way dispatch telephone voice streams are synchronously collected and pretreated; the pretreated voice data is input into a large language model adapted to power dispatch professional corpus, speech transcription is performed to obtain dialogue text with speaker differentiation; cloud-edge collaborative hierarchical scene recognition is performed on the dialogue text, preliminary scene judgment is completed by an edge side, deep semantic analysis and final scene recognition are completed by a cloud end based on the preliminary judgment result, and a scene label is output; based on the scene label and the dialogue text, key dispatch information is extracted through a dispatch log structured analysis model, and a structured standard log conforming to a power dispatch specification is generated according to a mapping relationship between a dispatch scene and a standard log.

Description

Technical Field

[0001] This invention belongs to the field of power grid dispatching and artificial intelligence technology, specifically relating to a method and system for dispatching telephone scene recognition and log generation based on large language model technology. Background Technology

[0002] As the core command and communication tool for power companies and dispatching departments, including the State Grid's municipal and county-level dispatching agencies, the dispatch telephone is a key link to ensure the daily operation and emergency response of the power system; the dispatch log serves as an important basis for tracing dispatching operations, assigning responsibilities, and reviewing work processes, and its accuracy and completeness directly affect the standardization and security of dispatching work.

[0003] In recent years, several solutions based on speech recognition and natural language processing technologies have been proposed for the automated processing of dispatch call records. A typical example is the real-time transcription and archiving scheme based on control speech recognition disclosed in patent document CN115567637A. This scheme constructs a control speech recognition model that includes acoustic and language models, transcribes the call audio in real time, and displays it in a chat dialog box format. Simultaneously, it uses a text convolutional neural network intent recognition model and a bidirectional long short-term memory network combined with a conditional random field slot-filling model to archive the transcribed content. Patent document CN121567809A proposes a log generation scheme that integrates dispatch recordings and networked command issuance. By fusing key information from dispatch recordings and networked command issuance, and after speech recognition and natural language processing, it generates structured dispatch logs based on preset templates, and provides log storage, query, and auditing functions. Patent document CN112420025A designs a structured data generation scheme oriented towards dispatch log specifications. It employs a recurrent neural network-derived speech recognition model to convert call audio into text, uses word vector tools for text optimization and deduplication, and then extracts key information through a backpropagation neural network to form structured data conforming to the specifications of the power grid dispatch management system. Patent document CN118609552A focuses on information extraction for power grid fault scenarios. It collects dispatch telephone voice in real time and converts it into text, constructs a power grid-specific natural language processing model, and uses an algorithm based on a pre-trained language model combined with bidirectional long short-term memory networks and conditional random fields to achieve entity annotation, extract descriptive information related to power grid faults, and then performs fault type judgment and cause analysis.

[0004] The aforementioned existing technical solutions have improved the automation level of dispatch call records to a certain extent, but they still have many shortcomings in the actual operation scenario of power dispatch and cannot fully meet the industry's strict requirements for high reliability and high compliance.

[0005] Existing solutions all employ a combination of traditional automatic speech recognition and general natural language processing techniques, without incorporating large language model technology. Their semantic understanding capabilities are limited by model size and training paradigm. When faced with the colloquial expressions, ambiguous instructions, inverted and elliptical sentences, and numerous technical terms unique to the power dispatching field, traditional models relying on keyword matching and shallow semantic features struggle to achieve deep semantic parsing. This results in a significant drop in recognition accuracy in non-standard expression scenarios and a severely insufficient generalization ability for intent classification.

[0006] Existing technologies for semantic understanding and information extraction are largely limited to specific sub-scenarios. For example, the model in CN118609552A is specifically designed for power grid fault scenarios, and its vocabulary recognition system and relation extraction logic are built around fault descriptions; the key information extraction in CN112420025A is also mainly geared towards equipment anomaly and fault reporting scenarios. For routine business scenarios that occur frequently in daily power dispatching work, such as parameter queries, inspection reports, issuance of planned instructions, and notifications, existing technologies have not yet provided dedicated scenario classification and structured parsing capabilities, resulting in a high reliance on manual processing for the generation of logs for non-fault-related calls.

[0007] Existing solutions mostly employ a shallow generation mode of entity recognition plus template filling, which can only extract simple entity information and fill it into preset fields, without performing event classification, operational logic relationship analysis, or causal reasoning on the call content. None of the above solutions have established a deep structured parsing mechanism compliant with the State Grid power dispatch log management specifications, and therefore cannot automatically generate the six standard dispatch logs conforming to the specifications. The generated logs still have significant gaps compared to manually written standard logs in terms of field completeness, semantic coherence, and format compliance, and cannot be directly integrated with power production platforms such as production management systems and operation management systems.

[0008] Existing systems all use a single communication protocol and do not support dual-mode PSTN and SIP communication. They are also incompatible with the State Grid's existing traditional circuit-domain PSTN dispatch consoles and modern IP-domain SIP dispatch terminals. Deploying the existing system would require completely overhauling the original dispatch communication architecture, which would be costly, involve long downtime for the modification, and pose significant risks to power grid operation.

[0009] Existing solutions are all designed based on the assumption of continuous network connectivity, with each stage—voice acquisition, cloud processing, and log generation—highly dependent on a stable network connection. When faced with extreme conditions such as network outages or communication fluctuations, the voice recognition and log generation services will completely fail, failing to guarantee the basic requirement of 24 / 7 uninterrupted operation of scheduling services. Current technologies have not yet disclosed a fault-tolerant mechanism for maintaining local voice processing and log storage offline, and for automatically synchronizing to the cloud after network recovery.

[0010] While existing solutions achieve the transcription and archiving of call content or the generation of logs, they lack a structured association mechanism between the original call recordings, the transcribed text, and the final dispatch logs. Call records and log entries are independent at the data level. When post-event verification, responsibility determination, or operation tracing is required, it is difficult to quickly locate the corresponding call recording segment from the logs, and it is also difficult to reverse-engineer the corresponding log generation results from the call records, affecting the audit efficiency and tracing accuracy of dispatch operations. Summary of the Invention

[0011] To address the shortcomings and deficiencies of existing technologies, such as limited semantic understanding in dispatch telephone voice recognition, poor system compatibility, insufficient adaptability to offline conditions, low standardization of log generation, and loose correlation between calls and logs, this invention provides a method and system for dispatch telephone scene recognition and log generation based on a large language model. This invention simultaneously collects two dispatch telephone voice streams through a PSTN and SIP dual-mode communication gateway without modifying the original dispatch communication architecture. A large language model, fine-tuned with low-rank adaptation of power dispatch professional corpus, performs speech transcription and deep semantic understanding to obtain speaker-distinguished dialogue text and key dispatch information. A cloud-edge collaborative hierarchical scene recognition mechanism is adopted, where a lightweight model on the edge side completes preliminary scene judgment, and the large language model on the cloud performs accurate semantic parsing based on the preliminary results and outputs dispatch business scene labels. Furthermore, a structured parsing model for dispatch logs, combined with preset log field importance weights, performs a completeness quantification assessment, and automatically generates structured standard logs conforming to power dispatch specifications according to the mapping relationship between dispatch scenarios and standard logs. When a network anomaly is detected, this invention caches voice data, dialogue text, and generated logs in local storage on the edge side. After the network is restored, it automatically synchronizes them to the cloud according to timestamps and priority order, and supports breakpoint resumption. At the same time, it establishes a unique association between call voice data, dialogue text, and standard dispatch logs to support bidirectional traceability. It also automatically writes standard logs into the production management system or operation management system through a preset interface, and automatically creates pending tasks or dispatches work orders based on key dispatch information. Overall, it realizes intelligent, standardized, highly available, and fully traceable power dispatch call processing.

[0012] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0013] A method for identifying and generating logs for dispatching calls based on large language model technology, applied to a cloud-edge-device collaborative processing system, includes:

[0014] Two dispatch telephone voice streams are simultaneously collected and preprocessed through the communication gateway corresponding to the public switched telephone network and the session initiation protocol.

[0015] The preprocessed speech data is input into a large language model adapted from a power dispatch professional corpus, and speech is transcribed to obtain dialogue text with speaker differentiation.

[0016] The dialogue text is subjected to cloud-edge collaborative hierarchical scene recognition. The edge side completes the initial scene judgment, and the cloud side completes deep semantic parsing and final scene recognition based on the initial judgment result, and outputs scene labels.

[0017] Based on the scenario tags and dialogue text, key scheduling information is extracted through a scheduling log structured parsing model, and a structured standard log conforming to power dispatching specifications is generated according to the mapping relationship between scheduling scenarios and standard logs.

[0018] Furthermore, the communication gateway interfaces with the existing dispatch telephone system via a 2M trunk interface.

[0019] Furthermore, the large language model adopts a low-rank adaptation approach, performs multi-task joint fine-tuning based on the power dispatching professional corpus, and optimizes three tasks: professional terminology correction, scene classification, and key information extraction.

[0020] Furthermore, the generated structured standard logs are automatically written into the corresponding log module of the production management system or operation management system through a preset interface, and pending tasks or work orders are automatically created based on the key scheduling information.

[0021] Furthermore, the cloud-based large language model uses structured prompts to guide reasoning and outputs scene labels and key information in a standard format; it compares the confidence difference between the initial scene judgment on the edge side and the final scene recognition on the cloud side, and triggers a review when the difference exceeds a preset threshold.

[0022] Furthermore, the system detects network status and caches voice data, dialogue text, and generated logs in local storage on the edge when the network is abnormal. After the network is restored, the data is automatically synchronized to the cloud according to timestamp and data priority, supporting breakpoint resume.

[0023] Furthermore, based on the preset importance weights of log fields, a log integrity score is calculated, and logs that do not meet the score are automatically completed or manually intervened; a unique association is established between voice data, dialogue text and standard scheduling logs, supporting bidirectional traceability.

[0024] Furthermore, the data is synchronized in the cloud using national cryptographic algorithms to perform end-to-end encryption.

[0025] Furthermore, a dispatch call scenario recognition and log generation system based on large language model technology, employing a cloud-edge-device collaborative architecture, includes:

[0026] The dual-mode voice access module is used to simultaneously acquire and preprocess two dispatch telephone voice streams through the communication gateway of the corresponding public switched telephone network and session initiation protocol.

[0027] The speech-to-text module is used to input preprocessed speech data into a large language model adapted from power dispatch professional corpus to obtain dialogue text with speaker differentiation.

[0028] The cloud-edge collaborative hierarchical scene recognition module is used to perform preliminary scene judgment on the dialogue text by the edge side, and the cloud side performs deep semantic analysis and final scene recognition based on the preliminary judgment results, and outputs scene labels.

[0029] The standard log generation module is used to extract key scheduling information based on the scenario tags and dialogue text, and generate a structured standard log that conforms to the power dispatching specifications according to the mapping relationship between scheduling scenarios and standard logs.

[0030] And a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0031] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0032] Compared to existing technologies, this invention and its preferred solution can complete system deployment without modifying the original dispatch communication architecture, while being compatible with both traditional circuit domain and modern IP domain dispatch terminals, significantly reducing deployment costs and the grid operation risks caused by downtime modifications. It employs a large language model adapted for the power dispatch field combined with a cloud-edge collaborative hierarchical recognition mechanism, effectively improving the semantic understanding accuracy of power dispatch terminology, colloquial expressions, and complex instructions, achieving coverage of all business scenarios and balancing real-time processing with recognition accuracy. It can automatically generate structured standard logs conforming to power dispatch specifications, ensuring the integrity and format compliance of log content, and can directly connect to power production and operation-related platforms, significantly reducing the workload of manual recording and organization. It possesses comprehensive offline high availability assurance capabilities, effectively solving the problem of business interruption under network anomalies and meeting the requirements of 24 / 7 uninterrupted operation of power dispatch. It establishes a unique association relationship for the entire call data chain, enabling bidirectional rapid traceability of dispatch operations and improving the efficiency of business auditing and responsibility identification. Simultaneously, it ensures data security through secure encrypted transmission, forming a closed loop from call access to business processing, comprehensively improving the standardization and operational safety management capabilities of power dispatch work. Attached Figure Description

[0033] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0034] Figure 1 This is a schematic diagram of the cloud-edge-device hardware architecture according to an embodiment of the present invention. Detailed Implementation

[0035] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0036] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0037] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0038] The purpose of this invention is to address the core shortcomings of existing technologies, such as limited semantic understanding capabilities, narrow scenario coverage, poor system compatibility, low standardization of log generation, insufficient adaptability to extreme operating conditions, and loose correlation between calls and logs. This invention provides a method and system for dispatching telephone scenario recognition and log generation based on large language model technology. This solution can achieve intelligent, standardized, and secure processing of the entire dispatching call process without modifying the existing dispatching communication architecture, fully meeting the State Grid power dispatching system's actual requirements for system compatibility, operational reliability, business compliance, and security control.

[0039] This invention proposes a method and system for dispatch call scene recognition and log generation based on large language model technology, applicable to intelligent terminal scenarios for dispatch and command at the State Grid city and county levels. The overall technical solution is built around a closed-loop process encompassing voice acquisition, semantic parsing, scene recognition, log generation, and system linkage. Based on the open-source OpenHarmony operating system, it adopts a PSTN and SIP dual-mode communication architecture and a three-layer collaborative deployment architecture (cloud, edge, and terminal). It integrates the large language model capabilities of the control cloud AI platform with the local processing capabilities of the edge side to achieve intelligent processing of dispatch calls and automatic generation of standardized logs.

[0040] In terms of system architecture design, this invention adopts a three-tier distributed architecture of cloud, edge, and device. For example... Figure 1As shown, the endpoint is the OpenHarmony smart terminal deployed at the dispatch site, responsible for dual-mode voice stream acquisition and preprocessing; the edge is the edge computing node with local computing and large-capacity caching capabilities, undertaking local semantic parsing, offline caching, and lightweight scene recognition tasks; the cloud is the control cloud artificial intelligence platform, which relies on a large language model fine-tuned by power dispatch professional corpus to complete high-precision semantic understanding, full-scene intent classification, and structured log generation. The three work together to complete the intelligent processing of the entire dispatch call process.

[0041] In terms of communication access, this invention adopts a dual-mode communication access method of PSTN and SIP. It is compatible with both traditional circuit domain PSTN dispatch consoles and modern IP domain SIP dispatch terminals through dispatch telephone gateways and voice gateways. It connects to the existing dispatch telephone system through 2M trunks, without replacing the original dispatch consoles and telephone equipment, and realizes unified collection and processing of voice streams without changing the original dispatch communication architecture.

[0042] Regarding offline high availability assurance, this invention configures a local solid-state drive cache and offline processing module on the edge side. When a network interruption is detected, the system automatically switches to offline operation mode, supporting local calls, recording caching, and preliminary log storage; after the network is restored, the cached data is automatically synchronized to the cloud according to timestamp and priority order, realizing breakpoint resume and data consistency assurance, ensuring that the scheduling service can run uninterrupted.

[0043] In the voice acquisition and preprocessing stage, the system acquires two voice streams, PSTN and SIP, through the scheduling telephone gateway and voice gateway respectively. The raw voice data is then processed sequentially with noise reduction, gain equalization and silence removal to effectively filter out background interference such as room noise and current noise. The voice volume is normalized to the standard range and invalid silence segments are removed, providing stable and reliable voice data for subsequent speech transcription.

[0044] The speech-to-text process employs an automatic speech recognition module that integrates a large language model. This module has been fine-tuned and trained with professional corpus including substation and line names, voltage levels, and dispatch instruction terms. It can deeply adapt to the language characteristics of the power dispatch field, accurately recognize various professional terms, colloquial expressions, and complex instructions, and output dialogue text with speaker identifiers, timestamps, and dual-track speech information, providing a high-quality text foundation for subsequent semantic analysis and scene recognition.

[0045] In terms of semantic parsing and scene recognition, this invention adopts a cloud-edge collaborative two-stage recognition mechanism. The lightweight model on the edge side completes local rapid semantic parsing and preliminary scene judgment, while the large language model in the cloud completes the accurate classification of twelve types of scheduling business scenarios. Finally, it outputs complete scene labels covering fault reporting, scheduling instructions, parameter query, emergency response, equipment operation, inspection reporting, wildfire, icing, unit management, power grid defects, etc., taking into account both real-time recognition and classification accuracy.

[0046] Regarding log generation and standardization, this invention constructs a structured parsing model for scheduling logs based on a pre-trained language model combined with a bidirectional long short-term memory network and a conditional random field architecture. Employing a joint loss function training method, it can simultaneously perform call event classification, operational logic analysis, and key information extraction. The system conducts a completeness assessment based on the importance weights of log fields as shown in Table 1, performing differentiated identification and filling verification of mandatory core fields, secondary core fields, and optional fields to ensure log content is complete, highlights key points, and has a standardized structure. Based on this, the system automatically generates six types of standard scheduling logs conforming to State Grid specifications according to the mapping relationship between scenarios and standard log types shown in Table 2, and generates a visual intermediate state page for dispatchers to verify, modify, and confirm.

[0047] Table 1 Importance Weight Table of Power Dispatch Log Fields

[0048]

[0049] Table 2. Mapping rules from 12 scenario types to 6 standard log types

[0050]

[0051] Regarding data security and cloud synchronization, this invention establishes a dedicated network channel with full-link encryption using a scheduling log cloud synchronization module deployed in Security Zone III. This supports multiple working modes, including real-time synchronization, offline retransmission, and breakpoint resumption, ensuring secure, reliable, and traceable data transmission. After receiving data, the cloud performs integrity verification and deep compliance processing, generating a permanent and unique identifier for each log entry.

[0052] In terms of data association and traceability, this invention establishes a unique two-level association mechanism between the original call recording, the transcribed text, and the final scheduling log. Through log identifiers, it enables bidirectional traceability, allowing for reverse lookup of corresponding call recording segments from the log and forward tracing of log generation results from the call record. This significantly improves the efficiency of scheduling operation auditing, responsibility identification, and historical verification.

[0053] In terms of business collaboration and security control, this invention achieves automatic integration with relevant business platforms such as operation management systems and production management systems through preset interfaces and standard log templates. It supports the automatic creation of to-do tasks and work order dispatch based on call content, forming a closed loop from call access to business processing. The system also integrates voiceprint and facial recognition dual authentication functions to verify the identity of dispatch operators, further enhancing the security control capabilities of dispatch operations.

[0054] By organically combining the above technical solutions, this invention breaks through the technical bottleneck of traditional speech recognition and natural language processing rule matching. Relying on the collaborative work of a large-scale language model specifically for power dispatching and a structured parsing model, it achieves accurate identification of twelve types of dispatching scenarios, adaptive understanding of professional terminology, and semantic parsing of complex instructions, significantly improving the accuracy and reliability of scenario identification and information extraction. The system can automatically generate six types of standard dispatch logs conforming to State Grid specifications and establish a unique secondary association between call records and dispatch logs, achieving fully structured, standardized, and traceable logs, fundamentally improving the problems of unstructured logs, difficult retrieval, and missing associations in traditional logs.

[0055] This solution effectively reduces operational risks caused by human error, instruction delays, and information omissions through fully automated processing, improving the standardization and efficiency of dispatching operations. Its highly reliable architecture and offline protection mechanism ensure stable and continuous operation of dispatching services even under complex conditions. Furthermore, its standardized log generation and end-to-end traceability design meet the stringent requirements of safety supervision, compliance auditing, and accountability in the power dispatching field, thereby comprehensively enhancing the intelligence level and safety management capabilities of power grid dispatching operations.

[0056] The implementation process of the present invention will be further demonstrated and described below through more specific embodiments:

[0057] The implementation process of this embodiment mainly includes four steps: dual-mode voice acquisition and preprocessing, voice transcription based on a dedicated large model for power dispatching, cloud-edge collaborative scene recognition, and automatic generation of structured logs, so as to realize intelligent processing of the entire dispatching call process.

[0058] Step 1: Dual-mode speech acquisition and preprocessing

[0059] This step establishes dual-channel voice acquisition via PSTN and SIP through a dispatch telephone gateway and a voice gateway, respectively. A 2M trunk is configured to connect to the existing dispatch telephone system. The sampling rate is set to 16kHz and the bit depth to 16bit. Real-time acquisition of bidirectional voice streams generates raw voice data sequences, where each element is the voice sample value of the corresponding frame. The frame length is 20ms and the frame shift is 10ms. The system assigns a unique identifier to each voice stream, associating it with caller and called numbers, call start time, and other call metadata, which is stored in a temporary buffer on the edge side. The raw voice data is then preprocessed. First, Wiener filtering is used to remove background noise such as equipment room noise and current noise. The denoised voice frame is obtained by multiplying the original voice frame by the Wiener filter coefficients, where the Wiener filter coefficients are the ratio of the original voice power spectrum to the original voice power spectrum plus the noise power spectrum. The noise power spectrum is estimated using silent frames. Next, the voice volume is normalized to the range [-1, 1] using automatic gain control. Then, the short-time energy of each voice frame is calculated, a threshold of 0.01 is set, and silent frames with short-time energy below the threshold are removed, resulting in the denoised voice sequence. To ensure business continuity under extreme conditions, the system monitors network status in real time. When a network interruption is detected, the preprocessed voice data is stored in a local solid-state drive cache with a capacity of no less than 256GB on the terminal, using a unique voice stream identifier, generating a cache index table containing the voice stream identifier, cache time, and network status. After the network is restored, the system batch synchronizes the data to the cloud recording library according to the cache index table. After synchronization is complete, the index table status is marked as synchronized. Specifically, this includes:

[0060] 1.1 Dual-mode voice stream access: A dual-channel voice acquisition system is established through a PSTN (Public Switched Telephone Network) gateway and a SIP (Speed-In-Place) gateway. A 2M trunk is configured to connect to the existing dispatch telephone system. The sampling rate is set to 16kHz and the bit depth to 16bit. Bidirectional voice streams are acquired in real time to generate raw voice data sequences. ( (The i-th frame's audio sample value, frame length 20ms, frame shift 10ms); assign a unique identifier to each audio stream. Associate call metadata (caller / called number, call start time) ), stored in the edge-side temporary cache area.

[0061] 1.2 Speech Preprocessing:

[0062] Noise reduction processing: Wiener filtering is used to remove background noise (such as computer room noise and current noise), and the resulting audio frames are noise-reduced. Calculation formula:

[0063]

[0064] Among them, Wiener filter coefficients , The original speech power spectrum, The noise power spectrum (obtained through silent frames);

[0065] Gain equalization: Normalizes the voice volume to the [-1,1] range using automatic gain control (AGC);

[0066] Silence Removal: Calculating the short-time energy of each speech frame. Set threshold =0.01, excision < The silent frame;

[0067] The final denoised speech sequence is as follows: (m≤n, after removing silent frames).

[0068] 1.3 Offline caching and synchronization:

[0069] Detect network status (1 = online, 0 = offline), if =0, the preprocessed speech data is then... Store the data in the terminal's local SSD cache (capacity ≥ 256G) and generate a cache index table. .

[0070] After the network is restored ( =1), press Batch sync to the cloud recording library, and mark it after sync is complete. The status is "synchronized".

[0071] Step 2: Speech transcription based on a large-scale power dispatching model

[0072] First, a corpus of over 100,000 power dispatching terminology entries was constructed, including substation and line names, voltage levels, and dispatching instruction terms. The basic automatic speech recognition model was then fine-tuned to optimize the accuracy of terminology recognition. The preprocessed speech sequence was input into the fine-tuned automatic speech recognition module, which transcribed the speech frame by frame, outputting an initial text sequence. Simultaneously, based on voiceprint features, the system distinguished between different speakers, such as dispatchers and on-site personnel, and labeled each text segment with a corresponding speaker tag and the timestamp of the corresponding speech frame. After transcription, the system calculated the transcription confidence score, i.e., the average posterior probability of each word in the transcribed text. Text segments with a transcription confidence score below 0.85 were filtered and marked for manual review. To further improve the accuracy of terminology recognition, the system incorporates a terminology correction model. This model calculates the Jaccard coefficient between the term to be corrected and the terms in the power terminology database to obtain the terminology matching degree. If the matching degree is greater than 0.7, the term to be corrected is replaced with the standard term with the highest matching degree, ultimately generating accurate dialogue text. Specifically, this includes:

[0073] 2.1 Fine-tuning of the power industry corpus: Constructing a power dispatching corpus (Including over 100,000 samples such as substation / line names, voltage levels, and dispatch instruction terms), the basic ASR model was fine-tuned to optimize the accuracy of professional terminology recognition.

[0074] 2.2 Large Model ASR Inference: The output of step 1 The finely tuned ASR module is input, and speech is transcribed frame by frame, outputting the initial text sequence. Based on voiceprint features, the speaker (dispatcher / on-site personnel) is distinguished, and each text segment is labeled with a speaker tag. timestamp (Corresponding to the audio frame time).

[0075] 2.3 Text Post-processing:

[0076] 2.3.1 Calculate the transcription confidence level :

[0077]

[0078] Where L is the length of the transcribed text, P Let be the posterior probability of the j-th word output by the large model.

[0079] filter Text fragments with a value <0.85 are manually reviewed.

[0080] 2.3.2 Terminology Correction Model: Utilizing Terminology Matching Degree Correction of terminology errors:

[0081]

[0082] Where T is the electrical terminology library, and Jaccard coefficient is... ,like >0.7, will Replace with the term 't' with the highest matching degree to generate the final dialogue text. .

[0083] As a preferred embodiment, the power dispatch-specific large language model adopted is based on an open-source general-purpose large language model with a Transformer architecture, and its parameters are efficiently fine-tuned to suit the language characteristics and business requirements of the power dispatch scenario. Considering the limited scale of labeled data in the power dispatch field, in order to avoid overfitting and reduce training costs, a low-rank adaptation (LoRA) fine-tuning paradigm is adopted, which trains only the low-rank matrix of the model's attention layer and freezes all other parameters of the model.

[0084] The fine-tuning process employs a multi-task joint training approach, simultaneously optimizing three core tasks: first, the task of correcting professional terminology in speech-to-text transcription; second, the task of classifying twelve types of scheduling business scenarios; and third, the task of extracting key information from scheduling logs. Through multi-task learning, the model simultaneously possesses semantic understanding, scene recognition, and information extraction capabilities, avoiding the insufficient generalization ability problem caused by single-task fine-tuning.

[0085] Step 3: Cloud-edge collaborative scene recognition:

[0086] A two-stage processing mechanism combining initial edge-side judgment and refined cloud-side judgment is adopted. First, keywords such as overload, load limiting, and wildfire are extracted from the final dialogue text to generate a keyword list and calculate local semantic feature vectors. These are then matched against a lightweight scene library on the edge side to output preliminary scene labels. Next, the final dialogue text and the local semantic feature vectors are uploaded to the control cloud AI platform. A dedicated large language model for power dispatching is input to generate global semantic feature vectors. Based on a Softmax scene classification model, the probability distribution of twelve dispatching scenarios is calculated to determine the final scene label and its corresponding confidence level. The system compares the scene recognition confidence levels on the cloud and edge sides. If the difference is less than 0.1, the cloud recognition result is directly adopted; if the difference is greater than or equal to 0.1, a manual review process is triggered, where dispatchers confirm the final scene type to ensure the accuracy of scene recognition. Specifically, this includes:

[0087] 3.1 Extraction Generate a keyword list using keywords (such as "overload", "load limiting", "wildfire"). Calculate the local semantic feature vector Match the edge-side lightweight scene library (containing 6 core scene categories) and output preliminary scene labels. .

[0088] 3.2 will and Uploaded to the control cloud AI platform, input into the power dispatching-specific large model, and a global semantic feature vector is generated. Based on the Softmax scene classification model:

[0089]

[0090] Where k corresponds to 12 types of scheduling scenarios, / To adjust the training weights of the cloud computing model, the multi-classifier outputs the probability distribution of 12 scene categories, thus determining the final scene labels. and confidence level .

[0091] 3.3 If The cloud-based results will be used directly; otherwise, manual review will be triggered.

[0092] As a preferred embodiment, the cloud-based large language model uses structured prompts to guide reasoning. The prompt template uniformly includes four parts: role definition, task requirements, output format, and constraints. The role definition clearly defines the model as "State Grid Power Dispatch Professional Assistant," the task requirements clearly define the specific work to be completed (such as scenario classification and key information extraction), the output format is specified as standard JSON format, and the constraints clearly prohibit the generation of irrelevant content and require the use of standard power dispatch terminology.

[0093] Taking a scheduling instruction scenario as an example, the typical prompt word structure is as follows:

[0094] You are an experienced State Grid dispatcher. Please analyze the following dispatch call text and complete two tasks: 1. Determine which of the twelve dispatch scenarios this call belongs to; 2. Extract key information from the caller, called party, involved equipment, operation content, and execution time. Output must strictly adhere to JSON format and must not contain any additional interpretations.

[0095] Call text: {Enter text}

[0096] The system will automatically load corresponding dedicated prompt word templates based on different initial scene labels, further improving the accuracy and consistency of large model inference.

[0097] Step 4: Automatic generation of structured logs

[0098] The automatic generation of structured logs is the core component of this invention's embodiment. First, a structured log parsing model trained on massive power dispatching business data is loaded. This model employs a pre-trained language model combined with a bidirectional long short-term memory network and a conditional random field architecture. During model training, the massive dispatching business dataset is divided into training, validation, and test sets in an 8:1:1 ratio. The dataset includes call text, scene labels, and annotations of key log information such as equipment, operations, and calling / called parties. The training process uses a joint loss function combining cross-entropy loss and conditional random field loss. Cross-entropy loss is used for scene classification tasks, and conditional random field loss is used for entity extraction tasks, with a weight coefficient set to 0.4. Training is performed based on the AdamW optimizer, and a power industry terminology vocabulary is introduced to fine-tune the word embedding layer, enabling the model to adapt to dispatching business scenarios in different regions. After training, the complete model is deployed on the control cloud AI platform, while a lightweight version is generated and deployed at the edge, supporting cloud-edge collaborative parsing and low-latency processing.

[0099] As a preferred embodiment, during model training, the core hyperparameters are set within the following general range: learning rate from 1e-5 to 5e-5, batch size from 8 to 32, number of training epochs from 10 to 30, and weight decay coefficient from 1e-4. These hyperparameters can be fine-tuned according to the scale and distribution of the actual training data to obtain optimal model performance.

[0100] For the lightweight model deployed at the edge, knowledge distillation technology is used for compression. The complete model is used as the teacher model and the lightweight model is used as the student model. The distillation loss function guides the student model to learn the output distribution of the teacher model. While ensuring that the accuracy loss does not exceed 2%, the number of model parameters is compressed to one-tenth of the original, and the inference speed is increased by more than 5 times, meeting the requirements of low latency processing at the edge.

[0101] During model inference, the system combines cloud-edge collaborative scene recognition results to categorize call events across six power business dimensions, including fault handling and instruction issuance. It also analyzes the causal and operational logic relationships between events, performs secondary verification of the intent of twelve business scenarios, and calculates the fusion confidence score. This fusion confidence score is obtained by weighting the cloud scenario confidence score and the model intent confidence score, with a weighting coefficient of 0.6. Based on the model's entity annotation capabilities, the system selectively extracts essential log fields such as caller ID, called party ID, involved devices, and core operations. After evaluating the confidence score of each field, it integrates them into a structured information set, automatically marking low-confidence fields as requiring manual review.

[0102] The system loads the corresponding State Grid general log template based on the final scenario label, establishes a one-to-one mapping relationship between the structured information set and the template fields, automatically fills in core fields, and completes basic fields such as call number and handling status from call metadata. At the same time, it completes standardization corrections such as professional terminology and time format. Subsequently, it calculates a preliminary log integrity score based on the importance weight of the log fields. Those with a qualified score generate a preliminary structured log in JSON format, which is temporarily stored in the edge-side log local cache module and marked with a synchronization status. Those with a qualified score automatically attempt to complete missing fields. If the score is still not qualified after completion, manual intervention is triggered.

[0103] After the initial logs are generated, a dedicated network channel using the scheduling log cloud synchronization module deployed in Security Zone III is established, employing full-link encryption with national cryptographic algorithms SM2, SM3, and SM4. At the edge, the initial structured logs and parsed call recording text are encapsulated into synchronization data packets, supporting three synchronization modes: real-time synchronization, offline retransmission, and breakpoint resumption. Upon receiving the data packets, the cloud performs integrity verification using the SM3 hash value. If the verification passes, the data is decrypted, stored, and a synchronization receipt is returned. If the verification fails, a limited number of retransmissions are triggered. After synchronization is complete, the synchronization status flag on the edge side is updated.

[0104] After receiving data in the cloud, the initial logs are converted into corresponding standard log types according to the predefined mapping rules from twelve types of dispatch scenarios to six types of State Grid standard dispatch logs. Deep compliance verification and standardization score calculation are then performed. Logs with passing standardization scores are loaded with the corresponding standard log template, generating State Grid standard structured logs and assigned a permanent and unique log identifier. The system automatically generates a visual confirmation page for intermediate log states, displaying the generated standard dispatch logs in a complete form with highlighted core fields for quick dispatcher review. The page provides entry points for modification, confirmation, and rejection. After verifying the content, the dispatcher clicks "confirm," and the log enters the reporting process. If discrepancies are found, field content can be modified directly on the page, and the system automatically updates the log data. If the log is completely incorrect, clicking "reject" marks it as abnormal, retains the original call data, and supports log regeneration. The system fully records the dispatcher's confirmation, modification, and rejection operations, simultaneously storing the operation time and operator information to meet traceability requirements. Finally, a one-to-one two-level association between call records and dispatch logs is established, an association table is generated and uniqueness is verified, and standard logs and associations are stored in the control cloud structured log library and association library respectively. Multi-dimensional retrieval indexes such as device, scenario, and time are established, and logs are automatically written to the corresponding log module of the production management system through preset interfaces and log templates, completing the closed loop of log parsing and archiving.

[0105] This step, based on the scenario tags output in step 3, the call metadata obtained in step 1, and the key call information extracted in steps 2 and 3, automatically generates six types of standard dispatch logs according to the State Grid power dispatching specifications. A visual intermediate state page is then generated for dispatchers to verify and confirm. Finally, the logs are automatically written to the PMS system log module through a preset interface and standard log template, achieving full automation of the log generation, confirmation, and reporting process. The specific steps are as follows:

[0106] 4.1 Structured parsing of dispatch logs: Loading a BERT-BiLSTM-CRF deep learning parsing model trained with massive amounts of power dispatch business data:

[0107] 1) Dataset Construction: This involves constructing a massive scheduling business dataset. The training set was divided into two parts in a ratio of 8:1:1. Validation set Test set It includes call text, scene tags, and key log information annotations (device, operation, caller and callee, etc.);

[0108] 2) Model architecture construction: The BERT-BiLSTM-CRF architecture is adopted. The BERT layer extracts deep semantic features of power dispatch text, the BiLSTM layer captures contextual logical relationships, and the CRF layer realizes accurate annotation and extraction of entities (key information).

[0109] 3) Model Training and Fine-tuning: A joint loss function combining cross-entropy loss (classification loss) and CRF loss (entity extraction loss) is used to ensure that the model has both scene recognition and key information extraction capabilities. The formula is as follows:

[0110]

[0111] Where L is the joint loss, Cross-entropy loss (for scene classification). For CRF loss (entity extraction). =0.4 is the weighting coefficient;

[0112] The model is trained based on the AdamW optimizer and fine-tuned by introducing a vocabulary of power industry terms for word embedding layer, so that the model can be adapted to local scheduling business scenarios in different regions.

[0113] 4) Model Deployment: The trained model is deployed in a lightweight manner on the cloud AI platform (cloud), and a lightweight version of the model is generated and deployed on the edge side, supporting cloud-edge collaborative parsing and low-latency processing.

[0114] Subsequently, based on the cloud-edge collaborative scenario recognition results from step 3, the call events are categorized and analyzed for causal / operational logic across six major power business dimensions, including fault handling and instruction issuance. The intent of 12 business scenarios is then verified a second time, and a fusion confidence score is generated. :

[0115]

[0116] in, ∈[0,1] represents the fusion confidence level. For step 3, the confidence level of the cloud scenario. To determine the confidence level of the analytical model's intent, β = 0.6 is used as the weighting coefficient.

[0117] Based on the entity annotation capability of the model, the system extracts essential core fields from logs, such as caller, called party, involved device, and core operation. After evaluating the confidence level of each field, the data is integrated into a structured information set. Fields with low confidence are marked for manual review, providing directly usable structured data for log filling.

[0118] 4.2 Preliminary Generation of Structured Logs: Based on the final scenario tags, the State Grid's general log template is loaded. A one-to-one mapping rule between the structured information set and the template is established. Core fields are automatically populated, and basic fields such as call number and handling status are supplemented from call metadata. Standardization corrections are also completed for professional terminology and time formats. A preliminary log completeness score is calculated according to the State Grid's standard field business importance weights. The scoring model is as follows:

[0119]

[0120] in, ∈[0,1] represents the initial log integrity score. This is a set of required core fields. For optional field set, The importance weights for field f are shown in Table 1. Populate the status of the field. =1 (already filled and correctly formatted), =0 (Unfilled / Formatting error);

[0121] Those who pass the test generate preliminary structured logs in JSON format, which are temporarily stored in the local log cache module on the edge side and marked as synchronized. Those who fail the test are automatically completed or trigger manual intervention.

[0122] 4.3 Cloud Synchronization of Scheduling Logs: A dedicated network channel encrypted with national cryptographic standards SM2 / SM4 is established through the cloud synchronization module of scheduling logs deployed in Security Zone III. The edge side encapsulates the preliminary structured logs and call recording parsed text into synchronization data packets, supporting three modes: real-time synchronization, offline retransmission, and breakpoint resumption. After receiving the data packets, the cloud performs integrity verification using SM3 hash values. If the verification passes, the data is decrypted, stored, and a synchronization receipt is returned. If the verification fails, a limited number of retransmissions are triggered. After synchronization is completed, the synchronization status flag on the edge side is updated to ensure the security, efficiency, and reliability of data transmission.

[0123] 4.4 Standard Log Generation and Secondary Association: The final standardization and association construction are completed on the control cloud platform. First, according to the rules uniformly defined by the State Grid, 12 scenario tags are mapped to 6 standard log types, such as daily work and temporary scheduling instructions (see Table 2). A deep compliance check is performed on the initial logs received from the cloud, and a standardization score is calculated. The standardization score model is as follows:

[0124]

[0125] in, ∈[0,1] represents the log standardized score. A set of required fields specific to 6 types of standard logs (as per...) Changes, such as wildfires (Including the location of the fire and the area affected) Istd(f) represents the standardized weight of a specific field, and Istd(f) represents the field's standardized state. =1 (compliant) =0 (non-compliant / missing), qualified logs load the corresponding standard log template and generate State Grid standard structured logs:

[0126]

[0127] And assign a unique permanent log. .

[0128] 4.5 Intermediate State Confirmation Page Generation and Scheduler Verification

[0129] 4.5.1 The system automatically generates a visual confirmation page for intermediate log states, which fully displays the six types of standard scheduling logs in form, with key fields highlighted for easy review by schedulers.

[0130] Page 4.5.2 provides entry points for modification, confirmation, and rejection:

[0131] After the dispatcher verifies that the information is correct, click "Confirm," and the log will begin the reporting process.

[0132] If any discrepancies are found, you can directly modify the field content on the page, and the system will automatically update the log data after the modification.

[0133] If the log is completely incorrect, you can click "Reject". The system will mark it as abnormal and retain the original call data, supporting the regeneration of the log.

[0134] 4.5.3 The system records the dispatcher's confirmation / modification / rejection operation traces, and synchronously retains the operation time and operator information to meet the traceability requirements. Then, it establishes a one-to-one and two-level association relationship between call records and dispatch logs, generates an association relationship table and performs uniqueness verification. Finally, the standard logs and association relationships are stored in the control cloud structured log library and association relationship library, respectively, and a multi-dimensional retrieval index is established for devices, scenarios, time and other dimensions. At the same time, the system automatically writes logs to the corresponding log module of the PMS system through preset interfaces and log templates to complete the closed-loop of the entire link from log parsing to archiving.

[0135] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0136] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0137] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0138] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0139] This invention is not limited to the above-described preferred embodiments. Anyone inspired by this invention can derive various other forms of a dispatch telephone scene recognition and log generation method and system based on large language model technology. All equivalent changes and modifications made within the scope of the claims of this invention shall fall within the scope of this invention.

Claims

1. A method for identifying and generating logs for dispatch telephone scenarios based on large language model technology, characterized in that, Applications include cloud-edge-device collaborative processing systems, including: Two dispatch telephone voice streams are simultaneously collected and preprocessed through the communication gateway corresponding to the public switched telephone network and the session initiation protocol. The preprocessed speech data is input into a large language model adapted from a power dispatch professional corpus, and speech is transcribed to obtain dialogue text with speaker differentiation. The dialogue text is subjected to cloud-edge collaborative hierarchical scene recognition. The edge side completes the initial scene judgment, and the cloud side completes deep semantic parsing and final scene recognition based on the initial judgment result, and outputs scene labels. Based on the scenario tags and dialogue text, key scheduling information is extracted through a scheduling log structured parsing model, and a structured standard log conforming to power dispatching specifications is generated according to the mapping relationship between scheduling scenarios and standard logs.

2. The method for dispatch call scene recognition and log generation based on large language model technology according to claim 1, characterized in that: The communication gateway connects to the existing dispatch telephone system via a 2M trunk interface.

3. The method for dispatch call scene recognition and log generation based on large language model technology according to claim 1, characterized in that: The large language model adopts a low-rank adaptation approach and performs multi-task joint fine-tuning based on the power dispatch professional corpus, while optimizing three tasks: professional terminology correction, scene classification, and key information extraction.

4. The method for dispatch call scene recognition and log generation based on large language model technology according to claim 1, characterized in that: The generated structured standard logs are automatically written into the corresponding log module of the production management system or operation management system through a preset interface, and pending tasks or work orders are automatically created based on the key scheduling information.

5. The method for dispatch call scene recognition and log generation based on large language model technology according to claim 1, characterized in that: The cloud-based large language model uses structured prompts to guide reasoning and outputs scene tags and key information in a standard format. The confidence difference between the initial scene judgment at the edge and the final scene recognition in the cloud is compared, and a review is triggered when the difference exceeds a preset threshold.

6. The method for dispatch call scene recognition and log generation based on large language model technology according to claim 1, characterized in that: The system detects network status and caches voice data, dialogue text, and generated logs on the edge for local storage when the network is abnormal. After the network is restored, the data is automatically synchronized to the cloud according to timestamp and data priority, and supports resuming interrupted transmission.

7. The method for dispatch call scene recognition and log generation based on large language model technology according to claim 1, characterized in that: Log integrity scores are calculated based on the preset importance weights of log fields. Logs that do not meet the score are automatically completed or manually intervened. A unique association is established between voice data, dialogue text and standard scheduling logs, supporting bidirectional traceability.

8. The method for dispatch call scene recognition and log generation based on large language model technology according to claim 1, characterized in that: Data synchronization in the cloud employs national cryptographic algorithms for end-to-end encryption.

9. A dispatching call scenario recognition and log generation system based on large language model technology, characterized in that, Adopting a cloud-edge-device collaborative architecture, including: The dual-mode voice access module is used to simultaneously acquire and preprocess two dispatch telephone voice streams through the communication gateway of the corresponding public switched telephone network and session initiation protocol. The speech-to-text module is used to input preprocessed speech data into a large language model adapted from power dispatch professional corpus to obtain dialogue text with speaker differentiation. The cloud-edge collaborative hierarchical scene recognition module is used to perform preliminary scene judgment on the dialogue text by the edge side, and the cloud side performs deep semantic analysis and final scene recognition based on the preliminary judgment results, and outputs scene labels. The standard log generation module is used to extract key scheduling information based on the scenario tags and dialogue text, and generate a structured standard log that conforms to the power dispatching specifications according to the mapping relationship between scheduling scenarios and standard logs.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images