A power marketing mobile operation assisting method

Abstract

This invention discloses a method for assisting mobile operations in power marketing, belonging to the field of power marketing operation assistance technology. The method includes: receiving voice commands from the power marketing mobile operation site; preprocessing the voice commands and extracting the command text using a hybrid recognition model; analyzing the operation matched by the command text using a semantic understanding model based on a power knowledge graph; generating an operation plan matching the operation based on a power operation knowledge base using a RAG retrieval enhancement engine, and matching the operation plan to the inspection work order corresponding to the voice command, generating a task list corresponding to the inspection work order. Furthermore, it uses a genetic algorithm to schedule work orders, combines GIS maps and a multi-dimensional intelligent assistance model to assist operators in generating work orders, and optimizes and iterates the model and knowledge graph based on feedback to achieve intelligent operations that support voice-driven, intelligent retrieval, dynamic scheduling, automatic recognition, and safe evolution, comprehensively improving the efficiency, standardization, and intelligence level of on-site operations.

Description

Technical Field

[0001] This invention relates to the field of power marketing operation assistance technology, and in particular to a mobile operation assistance method for power marketing. Background Technology

[0002] With the advancement of digital transformation in the power industry, the demand for intelligent support in power marketing field operations is becoming increasingly urgent. Traditional mobile operations mainly rely on manual review of paper documents or logging into multiple business systems to query information, resulting in low information acquisition efficiency, non-standardized operations, and slow response to professional issues. When carrying out tasks such as surveying, inspection, work order processing, and power supply scheme preparation, field personnel often face challenges such as data fragmentation, difficulty in knowledge retrieval, and insufficient decision support, affecting the quality of operations and service efficiency. Existing technologies lack deep integration and intelligent application of domain knowledge, making it difficult to achieve proactive guidance and decision support for business processes.

[0003] In recent years, artificial intelligence technology has developed rapidly, with large-scale models, retrieval-enhanced generative algorithms (RAGs), deep learning, and intelligent optimization algorithms providing new pathways for intelligent field operations. Meanwhile, Geographic Information Systems (GIS) have been widely applied in the spatial management of power equipment.

[0004] Therefore, there is an urgent need to build a mobile operation assistance system that integrates multimodal interaction, domain knowledge base, dynamic retrieval and intelligent generation to improve the intelligence, standardization and efficiency of power marketing field operations. Summary of the Invention

[0005] The purpose of this invention is to provide a mobile operation assistance method for power marketing that supports intelligent operations with voice-driven, intelligent retrieval, dynamic scheduling, automatic identification, and safety evolution, thereby comprehensively improving the efficiency, standardization, and intelligence level of on-site operations.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: a mobile operation assistance method for electricity marketing, comprising:

[0007] Receive voice commands from mobile power marketing work sites;

[0008] After preprocessing the voice commands, the command text is extracted using a hybrid recognition model;

[0009] The semantic understanding model based on the power knowledge graph is used to analyze the operation matched by the instruction text;

[0010] Based on the power operation knowledge base, the RAG retrieval enhancement engine is used to generate operation plans that match the operation, and the operation plans are matched to the inspection work orders corresponding to the voice commands to generate a task list corresponding to the inspection work orders.

[0011] Optionally, the mobile operation assistance method for electricity marketing of the present invention further includes:

[0012] Within each work order scheduling cycle, the system receives work order information to be processed and obtains information on schedulable personnel. The work order information to be processed includes the exception type, urgency, exception location coordinates, personnel skill requirements, and estimated time consumption. The information on schedulable personnel includes the personnel's current location coordinates, personnel skills, and current personnel workload.

[0013] Based on the information of the work orders to be processed and the information of the schedulable personnel, an improved genetic algorithm is used to solve the load balancing problem and obtain a matching scheme between new work orders and schedulable personnel.

[0014] The matching scheme and pending work order information are pushed to the mobile work assistance terminals of dispatchable personnel.

[0015] Optionally, the step of using an improved genetic algorithm to solve for load balancing based on the work order information to be processed and the schedulable personnel information, to obtain a matching scheme between new work orders and schedulable personnel, includes:

[0016] (1) Data modeling is performed on the work orders to be processed and the personnel to be dispatched based on the information of the work orders to be processed and the personnel to be dispatched;

[0017] (2) Modeling the load balancing problem and encoding individual chromosomes, including:

[0018] Let the set of work orders to be assigned be... , For the Mth work order, each work order is associated with a work order feature vector. ,in The number of feature dimensions for the work order;

[0019] Let the set of dispatchable operators be . , For the Nth work order, each schedulable personnel is associated with a personnel feature vector. ,in The number of dimensions for personnel characteristics;

[0020] In a genetic algorithm, a chromosome represents a complete scheduling scheme.

[0021] ,

[0022] in This indicates that pending work orders will be processed. Assign to dispatchable personnel;

[0023] The quality of chromosome individual X is determined by the fitness function. The fitness function, used for evaluation, reflects the scheduling scheme's performance in terms of path cost, capacity matching, and timeliness compliance. Its expression is:

[0024] ,

[0025] in, For path cost items, Penalty for skill mismatch Penalties for breach of contract by the deadline The weight of the path cost item. The weighting of the penalty for skill mismatch. The weighting of penalties for breach of contract by the deadline. ,and All are greater than 0, and each weight reflects the degree of importance the business attaches to path efficiency, skill matching, and timeliness.

[0026] (3) Initialize the population based on the pending work order data model, the schedulable personnel data model and the load balancing model, and calculate the fitness function of each chromosome in the initial population;

[0027] (4) Based on the fitness function calculation results, perform selection, crossover and mutation operations to obtain a new generation of better populations and calculate the fitness function until the calculation results meet the iteration termination condition. Then select the individual with the highest fitness value as the final load balancing scheme. If the calculation results do not meet the iteration termination condition, repeat step (4).

[0028] Optionally, in the fitness function, the path cost term The expression is:

[0029] ,

[0030] Where N is the total number of operators that can be scheduled. Indicates the feature vector from the personnel The first in The current location of each person. Indicates the feature vector from the work order The first in Work order The execution of the geographical location, For Starting from the point of origin, traverse all the operators assigned to the task. A path planning function for the total time and / or distance to the said geographic location;

[0031] Skill Matching Penalty The expression is as follows:

[0032]

[0033] Where M is the total number of work orders. Work order feature vector The skill requirements extracted from the i-th operation order. To extract from personnel feature vectors The extracted first The skill matrix of the aforementioned operators, Denotes the L1 norm of the skill matrix. To prevent division by zero errors for extremely small positive numbers, ;

[0034] Penalties for breach of deadline The expression is:

[0035]

[0036]

[0037] in, This indicates taking the maximum value. Indicates the first The estimated completion time for each of the aforementioned work orders. Indicates the first The deadline for each of the aforementioned work orders. Indicates the time when scheduling starts. This represents the function representing the estimated travel time. Indicates the feature vector from the personnel No. Operators The current position, Indicates the feature vector from the work order The first in Work order The geographical location of the execution, Indicates the first The operation duration of each of the aforementioned work orders. Indicates the first The historical efficiency of an operator is defined as the ratio of the historical average standard time to the actual time spent. >1 indicates high efficiency.

[0038] The above technical solutions can ensure the processing of emergency work orders and the optimal path, enabling the algorithm to accurately respond to business needs.

[0039] Building upon the above, load balancing optimization can be implemented as a supplementary adjustment mechanism to the genetic algorithm results. Based on the initial scheduling (load balancing scheme) output by the genetic algorithm, the allocation of personnel and work orders can be iteratively adjusted using a maximum-minimum fairness algorithm. This addresses the issue of uneven personnel workload that may arise in the initial scheduling phase of the algorithm.

[0040] Optionally, the mobile operation assistance method for electricity marketing of the present invention further includes:

[0041] The server generates on-site operation assistance information for work orders to be processed based on GIS maps and on-site operation assistance models, which can be downloaded by operators through mobile operation assistance terminals; the on-site operation assistance information includes at least inspection path information and equipment defect identification and cause analysis information;

[0042] The server generates and updates keys for each mobile work auxiliary terminal according to a set rotation cycle, based on the unique identifier and dynamic password of the mobile work auxiliary terminal device. The mobile work auxiliary terminal uses the key to transmit on-site work order processing data to the server via an encrypted transmission channel. This encrypted transmission channel uses the national cryptographic SSL protocol and the SM2 elliptic curve algorithm for key exchange between the mobile work auxiliary terminal and the server, employs the SM4 symmetric encryption algorithm for data encryption, and attaches a message digest generated by the SM3 algorithm, as well as using the SM3 algorithm for communication integrity protection. Upon receiving the encrypted data, the data receiver recalculates the message digest and compares it to verify whether the data has been tampered with during transmission. This implementation method can completely replace the traditional TLS protocol and meets the security requirements of Level 3 of the Power Industry Network Security Protection 2.0.

[0043] Optionally, the on-site work order processing data includes defect identification records, knowledge query logs, and work order processing results;

[0044] The mobile operation assistance method for power marketing also includes: the server-side performs performance evaluation and model update of the field operation assistance model based on the field work order processing data accumulated within the current period window, according to the set knowledge evolution cycle, and the knowledge recommendation priority of the power operation knowledge base.

[0045] Regularly updating the knowledge recommendation priority ensures that frequently used knowledge is placed at the top.

[0046] Optionally, preprocessing the voice commands includes:

[0047] The high-frequency and low-frequency coefficients of the speech command signal are obtained by five-layer discrete wavelet decomposition, and the threshold is calculated based on the VisuShrink criterion. :

[0048]

[0049]

[0050] in, The threshold value is... The standard deviation of noise. The length of the voice command signal. It is a logarithmic function. These are the detail coefficients obtained from the first-level wavelet decomposition. A function to obtain the median;

[0051] Based on the threshold The detail coefficients of each layer are subjected to soft thresholding, expressed as follows;

[0052]

[0053] in, These are the detail coefficients after soft thresholding. , , This indicates taking the larger value. is the sign function, indicating the positive or negative direction of the coefficients to be preserved;

[0054] The speech command signal is reconstructed based on the results of soft thresholding, using the following formula:

[0055]

[0056] In the formula, For the original voice command signal The reconstructed voice command signal, These are the approximate coefficients obtained from the 5th level wavelet decomposition;

[0057] The reconstructed voice command's voice frame is converted into a Mel spectrogram.

[0058] The above technical solution can filter out periodic device noise while preserving effective speech features, ensuring the recognizability of key commands. The specific process of converting the reconstructed speech command's speech frame into a Mel spectrogram can be found in existing technologies.

[0059] Optionally, the hybrid recognition model includes a CNN layer, an RNN layer, and an output layer. The CNN layer is used to extract spectral features from the Mel spectrogram of the speech command. The RNN layer is used to capture long-term dependencies between speech frames to obtain dynamic temporal features. The output layer is used to map the dynamic temporal features output from the RNN layer to a character sequence to obtain the command text.

[0060] Optionally, the semantic understanding model based on the power knowledge graph includes a semantic parsing module, a knowledge fusion module, an attention fusion module, and a parsing output module;

[0061] The step of analyzing the job operation matched by the instruction text using a semantic understanding model based on power knowledge graphs includes:

[0062] The instruction text is cleaned, segmented, and tokenized to generate a standardized token sequence;

[0063] The token sequence is input into the semantic parsing module implemented using the BERT-BiGRU-CRF semantic understanding model. The BERT layer performs deep contextual semantic encoding on the token sequence to generate a semantic vector. The BiGRU layer performs bidirectional sequence dependency modeling on the semantic vector. The CRF layer performs global optimal label decoding and outputs an intent classification label and a structured entity label sequence containing job object entities and job type entities.

[0064] The structured entity tag sequence is input into the knowledge fusion module and linked with the power knowledge graph. The text entities are mapped to standard equipment nodes and standard operation nodes in the power knowledge graph. The standardized operation process under the standard equipment node is queried from the power knowledge graph to obtain the standardized operation process corresponding to the standard operation node and generate the corresponding subgraph embedding vector.

[0065] The semantic vector, subgraph embedding vector, and historical task context vector are input into the attention mechanism module. Through attention weight calculation, a knowledge context vector, a dialogue context vector, and a fusion representation vector containing the current semantics, domain knowledge, and historical task status are generated.

[0066] Based on the fusion representation vector, the task object, task type and standard operation process are bound together, and the standardized task object, standardized task type and corresponding standardized operation process corresponding to the instruction text are output.

[0067] In the above technical solutions, the knowledge context vector is used to carry the equipment attributes, operation specifications and safety constraints information in the power knowledge graph; the dialogue context vector is used for referential resolution, omission information completion and intent clarification in multi-turn dialogue scenarios; and the fusion representation vector is used to integrate current semantics, domain knowledge and historical state information to generate a structured output of standardized operation objects, standardized operation types and operation processes.

[0068] The power knowledge graph designed in this invention includes entities such as equipment, procedures, and operating steps, and relationships such as "includes," "based on," and "preconditions," which can provide background knowledge for intent parsing; the BERT model encodes the contextual semantics of the text and outputs deep feature vectors; BiGRU captures the sequence dependencies of the text, and CRF decodes entity labels, including entity labels and job type labels; by matching the job type label subtree under the entity label node in the knowledge graph through entity links, the specific operating steps are located.

[0069] Optionally, the attention weight calculation performed by the attention mechanism module includes calculating the dynamic attention weight of the dialogue context based on the standard device node and standard job node information in the subgraph embedding vector, using the following formula:

[0070]

[0071]

[0072] in, For the current command and the history number Similarity of turn-based dialogue commands, The corresponding normalized dialogue context dynamic attention weights; For An exponential function with base 0; The query vector represents the job object and job type labels output by the BERT-BiGRU-CRF semantic understanding model for the current instruction. The key vector is the historical [number]. The round-robin dialogue instructions are identified and output by the BERT-BiGRU-CRF semantic understanding model, which identifies the job object and job type labels. For the total number of rounds of historical dialogue, This indicates the calculation of the similarity score.

[0073] The objective of the above implementation is to automatically associate context when inputting contextual instructions with sequential logic, by calculating and comparing the dynamic attention weights of the dialogue context, thus avoiding the need for staff to repeatedly input entity and operation information.

[0074] Optionally, the power operation knowledge base adopts a hybrid storage mode of structured and unstructured data. Structured data containing equipment parameters and formulas is stored in a relational database, while unstructured data containing regulations, documents, and cases is stored in the vector database Milvus after extracting key sentences using TextRank. A three-level index is established, whereby the first-level index contains an inverted index of the professional domain thesaurus; the second-level index contains a 768-dimensional semantic vector generated by Sentence-BERT; and the third-level index contains a document chapter structure tree, supporting hierarchical retrieval of "specification → clause → example".

[0075] The method of generating operation plans that match the operation based on the power operation knowledge base and using the RAG retrieval enhancement engine includes:

[0076] Enhanced query statements are constructed by integrating information about the job object, job type, and job process.

[0077] The enhanced query statement is input into the encoder to generate a query vector. Based on the query vector, a similarity retrieval is performed in a pre-built power operation knowledge base to recall a set of associated document fragments that match the operation object and operation type.

[0078] Based on at least one of the factors of equipment matching degree, job type matching degree, and document timeliness, the set of related document fragments is reordered to filter highly relevant document fragments.

[0079] The highly relevant document fragments are segmented and spliced ​​according to process stages to form a reference context;

[0080] The reference context, job object, job type, and job process are input into a generator implemented using a large language model to generate an operation plan matching the job operation. The operation plan is organized by job process stages, including key points for each stage, safety risk warnings, emergency response measures, and a compliance checklist. The output format can be restricted to: clause reference + key points + precautions.

[0081] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0082] 1. This invention combines on-site voice command input with an artificial intelligence model and power knowledge graph to achieve accurate voice recognition and intelligent automatic operation generation. The RAG retrieval enhancement engine obtains operation orders by matching operation attention schemes according to the operation. The genetic algorithm and load balancing algorithm are used to automatically schedule operation orders, achieving fully intelligent operation order sorting and operation content generation, thus improving the work efficiency on-site. Furthermore, based on GIS maps and operation orders, on-site operation assistance is generated to provide a more operator-friendly manual on the basis of the work orders. This not only improves the convenience and efficiency of operation, but also ensures the safety of operators and the applicability of operation scenarios. Subsequent iterative optimization through federated learning and LoRA algorithms can also improve the robustness and scenario applicability of the various models and knowledge graphs in this invention, and improve the accuracy of operation operations. At the same time, it can also provide simple and convenient intelligent assistance functions for new employees, reducing reliance on senior personnel and alleviating the pressure of personnel training.

[0083] 2. The automated matching and scheduling of new work orders with available personnel can significantly reduce manual operation time, greatly reduce the manpower and financial resources spent on work order processing, and significantly improve the response speed of operations.

[0084] 3. The generation of operation plans and on-site operation auxiliary information in operation work orders can reduce the time spent by operators in the workflow, speed up the response speed of emergency repairs and other businesses, and at the same time provide on-site operators with accurate and efficient defect identification technology, improve the accuracy of defect identification and the rationality of the solution, reduce rework caused by human judgment errors, and reduce business risks.

[0085] 4. Encrypted data transmission between the server and the terminal ensures the security of on-site work order processing data during transmission, meets the confidentiality requirements of the power industry, and avoids the risk of data leakage. Attached Figure Description

[0086] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0087] Figure 1 This is a schematic diagram illustrating the technical concept of an intelligent auxiliary method for mobile operations in power marketing, provided by an embodiment of the present invention.

[0088] Figure 2 This is an encryption flowchart provided in an embodiment of the present invention. Detailed Implementation

[0089] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use.

[0090] The following describes the specific implementation process of the intelligent auxiliary method for mobile power marketing operations in this embodiment of the invention. (Refer to...) Figure 1 .

[0091] A mobile operation assistance method for power marketing includes: receiving voice commands from the mobile operation site of power marketing; preprocessing the voice commands and extracting the command text using a hybrid recognition model; analyzing the operation matched by the command text using a semantic understanding model based on a power knowledge graph; generating an operation plan matching the operation using a RAG retrieval enhancement engine based on a power operation knowledge base, and matching the operation plan to the inspection work order corresponding to the voice command to generate a task list corresponding to the inspection work order.

[0092] The above method can be implemented based on a system architecture of server and mobile operation terminal: the terminal receives voice commands at the work site, and then the terminal or server performs data analysis corresponding to the voice commands to obtain a list of on-site work tasks corresponding to the voice commands, i.e., the operation plan.

[0093] Taking the routine inspection of electricity metering boxes in a residential community as an example, the operator uses a mobile marketing terminal to issue voice commands after entering the metering box area. The terminal first preprocesses the voice command signal, and then extracts the core text of the preprocessed voice command through a hybrid recognition model. Based on the semantic understanding model and the power knowledge graph, the text vector of the core text is extracted and matched with the corresponding operation.

[0094] I. Preprocessing of Voice Commands

[0095] In some possible embodiments, the voice command is decomposed into high-frequency and low-frequency coefficients of the voice command signal using a 5-layer discrete wavelet decomposition method, and a threshold is calculated based on the VisuShrink criterion. :

[0096]

[0097]

[0098] in, The threshold value is... The standard deviation of noise. The length of the voice command signal. It is a logarithmic function. These are the detail coefficients obtained from the first-level wavelet decomposition. A function to obtain the median;

[0099] Based on the threshold The detail coefficients of each layer are subjected to soft thresholding, expressed as follows;

[0100]

[0101] in, These are the detail coefficients after soft thresholding. , , This indicates taking the larger value. is the sign function, indicating the positive or negative direction of the coefficients to be preserved;

[0102] The speech command signal is reconstructed based on the results of soft thresholding, using the following formula:

[0103]

[0104] In the formula, For the original voice command signal The reconstructed voice command signal, These are the approximate coefficients obtained from the 5th layer wavelet decomposition.

[0105] The acquired speech signals are often mixed with environmental noise such as vehicle traffic noise and wind noise. By enabling the wavelet noise reduction module, the signal is decomposed into five levels of discrete wavelet decomposition using the db8 wavelet basis function, and the high-frequency wavelet coefficients are processed using a soft thresholding method based on the VisuShrink criterion, effectively filtering out background noise. This noise reduction process can improve the speech signal-to-noise ratio by approximately 10 to 15 dB, significantly improving signal quality.

[0106] After the command signal is reconstructed, the processed voice frame of the voice command is converted into a Mel spectrogram to complete the preprocessing of the voice command.

[0107] The specific processing steps and algorithms for Mel spectrograms are as follows:

[0108] (1) Frame segmentation and windowing processing

[0109] For sampling rate The original audio signal is divided into frames. Let each frame have N = 400 sampling points and a frame shift of M = 160 sampling points. For the t-th frame signal... Apply a Hamming window, the window function is defined as:

[0110]

[0111] in, For the nth window coefficient, It is a cosine function. Pi is the constant of π.

[0112] The signal of the t-th frame after windowing is:

[0113]

[0114] (2) Fast Fourier Transform (FFT)

[0115] For the windowed frame Zero padding to length And perform a discrete Fourier transform to obtain a complex sequence in the frequency domain:

[0116]

[0117] in, Let be the complex spectral value of the t-th frame at the k-th frequency bin; j is the imaginary unit. For FFT points, Pi is the constant of a circle. The total number of frames. It is a natural constant. For the kth frequency, This is the coefficient of the nth window.

[0118] (3) Power spectrum calculation

[0119] Calculate the one-sided power spectrum, taking only the first side. Frequency:

[0120]

[0121] in, This represents the power spectral density of the t-th frame at the k-th frequency bin. This indicates the number of frequencies considered.

[0122] (4) Mel filter bank design and mapping

[0123] Build A triangular Mel filter. First, the linear frequency range is... Transformed to the Mel scale, where the highest frequency , Sampling rate. The conversion relationship between Mel scale and linear frequency. for:

[0124]

[0125] in, It is a logarithmic function with base 10. For frequency;

[0126] Uniformly generated on the Mel domain Each point is then converted back to a linear frequency through an inverse transform, and finally mapped to the corresponding FFT frequency index. The frequency response of the m-th triangular filter Defined as:

[0127]

[0128] in, For FFT frequency index, satisfying 0 ≤ <K, Let m be the position of the center of the m-th filter on the FFT frequency bin; This represents the position of the center of the (m-1)th filter on the FFT frequency bin; This represents the position of the (m+1)th filter center on the FFT frequency bin; This represents the judgment condition in other situations.

[0129] (5) Mel energy calculation and logarithmic compression

[0130] Calculate the energy of each frame in each Mel channel, and then take the natural logarithm or common logarithmic compression:

[0131]

[0132]

[0133] in, Let m be the energy of the m-th Mel channel in the t-th frame. The log-Mel energy forms the feature vector of this frame. To prevent extremely small positive numbers from taking the logarithm of zero.

[0134] (6) Construction of Mel spectrogram

[0135] T consecutive frames By arranging the components in chronological order, a two-dimensional matrix is ​​formed.

[0136]

[0137] in, For size The matrix, which is the Mel spectrogram, serves as the input feature for subsequent speech recognition or semantic understanding models;

[0138] The aforementioned preprocessing method can filter out periodic equipment noise while preserving effective voice features, ensuring the recognizability of key commands. The preprocessing of voice command signals is preferably performed through a mobile terminal, followed by subsequent data analysis based on the preprocessing results, or the preprocessing results can be transmitted to a server for further data analysis.

[0139] II. Extraction of Instruction Text

[0140] In some possible embodiments, after obtaining the preprocessed result of the voice command, the mobile terminal or server extracts the command text using a hybrid recognition model. Specifically:

[0141] The hybrid recognition model includes a CNN layer, an RNN layer, and an output layer. The CNN layer is used to extract spectral features from the Mel-spectrum of the speech command. The RNN layer is used to capture the long-term dependencies between speech frames and obtain dynamic temporal features. The output layer is used to map the dynamic temporal features output by the RNN layer to a character sequence to obtain the command text.

[0142] The CNN consists of four convolutional layers with a kernel size of 3×3, a stride of 1, and padding of 1. It processes the Mel spectrogram, with layers 1-2 capturing the spectral details of monosyllabic words and layers 3-4 aggregating the combined features of multisyllabic words. The RNN temporal modeling includes two layers of bidirectional LSTM to process the temporal dependencies of speech, with 256 hidden units. It distinguishes the tonal differences of similar speech commands and solves the problem of recognizing homophones with different meanings. For example, it can handle cases where the individual characters are different but the overall text structure is the same, such as "check current" and "check voltage," but the meanings are different. This improves the accuracy of speech recognition and ensures the reliability of subsequent text generation.

[0143] The loss function and vocabulary of the CNN-RNN hybrid recognition model adopt the CTC loss function to handle the misalignment problem between speech and text. More than 200 special instruction vocabularys are customized for power business, including equipment names, operation verbs, and procedure terms. The training data includes 500 hours of speech in power operation scenarios, covering different dialect accents. Through large-scale speech data training, it can cover a variety of usage scenarios and user groups to closely resemble real speech usage scenarios and improve the accuracy of speech recognition.

[0144] The CNN-RNN hybrid model can convert the speech "Start meter box inspection process" into the text "Start meter box inspection process".

[0145] III. Analysis of the corresponding operation of voice commands

[0146] In some possible embodiments, after obtaining the corresponding instruction text of the voice command, the terminal or server extracts the text vector of the core text based on a semantic understanding model and a power knowledge graph to match the corresponding job operation. Specifically:

[0147] The semantic understanding model based on the power knowledge graph includes a semantic parsing module, a knowledge fusion module, an attention fusion module, and a parsing output module;

[0148] The step of analyzing the job operation matched by the instruction text using a semantic understanding model based on power knowledge graphs includes:

[0149] The instruction text is cleaned, segmented, and tokenized to generate a standardized token sequence;

[0150] The token sequence is input into the semantic parsing module implemented using the BERT-BiGRU-CRF semantic understanding model. The BERT layer performs deep contextual semantic encoding on the token sequence to generate a semantic vector. The BiGRU layer performs bidirectional sequence dependency modeling on the semantic vector. The CRF layer performs global optimal label decoding and outputs an intent classification label and a structured entity label sequence containing job object entities and job type entities.

[0151] The structured entity tag sequence is input into the knowledge fusion module and linked with the power knowledge graph. The text entities are mapped to standard equipment nodes and standard operation nodes in the power knowledge graph. The standardized operation process under the standard equipment node is queried from the power knowledge graph to obtain the standardized operation process corresponding to the standard operation node and generate the corresponding subgraph embedding vector.

[0152] The semantic vector, subgraph embedding vector, and historical task context vector are input into the attention mechanism module. Through attention weight calculation, a knowledge context vector, a dialogue context vector, and a fusion representation vector containing the current semantics, domain knowledge, and historical task status are generated.

[0153] Based on the fusion representation vector, the task object, task type and standard operation process are bound together, and the standardized task object, standardized task type and corresponding standardized operation process corresponding to the instruction text are output.

[0154] For example, the text "Start meter box inspection process" is fed into the semantic understanding model. The power knowledge graph contains entities such as equipment, procedures, and operating steps, and relationships such as "includes," "based on," and "preconditions," providing background knowledge for intent parsing. The BERT model encodes the contextual semantics of "check meter box inspection steps" and outputs a deep feature vector. Then, BiGRU captures the sequence dependency between "meter box" and "inspection steps," and CRF decodes the entity labels, including the entity "meter box" + the action "inspection steps." Finally, by matching entity links to the "standard inspection process" subtree under the "meter box" node in the knowledge graph, the specific operating steps are located. Multi-step recognition prevents the lack of contextual and semantic relationships during model recognition, which could lead to accurate text recognition but inaccurate semantic recognition. This further improves the accuracy of text recognition, identifying the core entities of the instruction text "Start meter box inspection process": "meter box" and "inspection."

[0155] In some possible embodiments, the attention weight calculation performed by the attention mechanism module includes calculating the dynamic attention weight of the dialogue context based on the standard device node and standard job node information in the subgraph embedding vector, using the following formula:

[0156]

[0157]

[0158] in, For the current command and the history number Similarity of turn-based dialogue commands, The corresponding normalized dialogue context dynamic attention weights; For An exponential function with base 0; The query vector represents the job object and job type labels output by the BERT-BiGRU-CRF semantic understanding model for the current instruction. The key vector is the historical [number]. The round-robin dialogue instructions are identified and output by the BERT-BiGRU-CRF semantic understanding model, which identifies the job object and job type labels. For the total number of rounds of historical dialogue, This indicates the calculation of the similarity score.

[0159] Dynamic attention weights Its main function is to, in context instructions with sequential logical relationships, [use]... By automatically associating the context with weights, the repeated input of the action features and entity label features is avoided. Keyword capture through the attention mechanism can further improve the accuracy of voice command recognition and matching.

[0160] Based on the above, entity links are established between the attention mechanism and the power knowledge graph. Based on the core entities "metering box" and "inspection", the technology can match the "standard inspection process" operation path under the "metering box" node.

[0161] IV. Generation of Work Operation Plans

[0162] In some possible embodiments, based on a power operation knowledge base, an operation plan matching the operation is generated using a RAG retrieval enhancement engine, including:

[0163] Enhanced query statements are constructed by integrating information about the job object, job type, and job process.

[0164] The enhanced query statement is input into the encoder to generate a query vector. Based on the query vector, a similarity retrieval is performed in a pre-built power operation knowledge base to recall a set of associated document fragments that match the operation object and operation type.

[0165] Based on at least one of the factors of equipment matching degree, job type matching degree, and document timeliness, the set of related document fragments is reordered to filter highly relevant document fragments.

[0166] The highly relevant document fragments are segmented and spliced ​​according to process stages to form a reference context;

[0167] The reference context, job object, job type, and job process are input into a generator implemented using a large language model to generate an operation plan matching the job operation. The operation plan is organized by job process stages, including key points for each stage, safety risk warnings, emergency response measures, and a compliance checklist. The output format can be restricted to: clause reference + key points + precautions.

[0168] In some possible embodiments, the knowledge base also includes a relational database and a vector database. The relational database stores structured data, while unstructured data is stored in the vector database after key sentences are extracted using the TextRank algorithm. The vector database includes a three-level index, and the RAG retrieval engine performs retrieval based on the three-level index.

[0169] The knowledge base is constructed using a hybrid storage approach of "structured + unstructured". Structured data such as equipment parameters and formulas are stored in a relational database; unstructured data such as procedures, documents, and cases are processed and stored in a vector database. The vector database has a three-level index: the first-level index contains an inverted index of a professional domain thesaurus; the second-level index contains 768-dimensional semantic vectors generated by Sentence-BERT; and the third-level index contains a document chapter structure tree, supporting hierarchical retrieval of "specification → clause → example". This three-level index not only improves matching accuracy but also facilitates orderly data storage and subsequent management. Furthermore, it allows for index matching according to the rules used in different scenarios, enhancing scenario adaptability and the accuracy of subsequent document generation.

[0170] In some possible embodiments, the aforementioned RAG retrieval enhancement engine involves algorithms such as the BM25 algorithm, cosine similarity algorithm, maximum marginal relevance algorithm, and Llama2-7B model. Specifically: the BM25 algorithm can be used to filter and calculate the text similarity between the text vectors of job operations and the text of the knowledge base. Based on cosine similarity, the text vectors of job operations with high text similarity and the knowledge base are finely ranked. The results of the BM25 algorithm and the cosine similarity algorithm are then weighted and calculated.

[0171]

[0172] in, For weighted scores, The result of the BM25 algorithm, This is the result of cosine similarity;

[0173] Based on the weighted calculation results, obtain the initial selection operation plan for the knowledge base;

[0174] The expression for calculating text similarity using the BM25 algorithm:

[0175]

[0176] in, For vocabulary In the document Similarity score in For vocabulary total, Total number of documents For including the i-th word The number of documents, where log is an exponential function. To control the word frequency saturation parameter, The normalized intensity parameter is the document length. It is 1.5. It is 0.75. For the i-th word In the document word frequency, This represents the average length of all documents.

[0177] The expression for cosine similarity is:

[0178]

[0179] in, Scoring the cosine similarity. For vocabulary , For document ;

[0180] By combining cosine similarity and the BM25 algorithm with weighted calculation, the matching degree between words and documents can be further improved, avoiding omissions and incorrect matches. At the same time, multiple similarity algorithms can also avoid the local optima problem of a single algorithm, further improving accuracy.

[0181] The long document of the initial operation consideration scheme is cropped by using a sliding window and maximum marginal relevance.

[0182] The sliding window size is 300 characters, with a step size of 100 characters;

[0183] The expression for maximum marginal correlation is:

[0184]

[0185] in, For the i-th document Maximum marginal relationality, For weighted balance coefficients, , Represents the i-th document and vocabulary similarity, For the i-th document In document collection The j-th most similar document Similarity;

[0186] By pruning the length of the text, we can ensure that the computational load is not too large, which would increase the computational burden and increase the investment in hardware costs. At the same time, small-scale data can also improve computational efficiency. Furthermore, based on the principle of maximum marginal relevance, we can also ensure that the pruned text does not destroy semantics and improve the accuracy of the text.

[0187] Based on the Llama2-7B model, the text format of the initially selected operational note schemes after trimming is sorted to generate operational note schemes. Llama2-7B uses LoRA technology to freeze 99% of the parameters, and the training data consists of 50,000+ power-related question-and-answer pairs, including standard explanations and case analyses. During generation, the output format is restricted to "clause citation + operational points + precautions" to output standard text.

[0188] Based on the aforementioned three-level index vector library, the operation path of the "standard inspection process" under the "metering box" node is matched;

[0189] Level 1 Search: Based on the BM25 algorithm, quickly filter documents containing the keywords "measuring box" and "inspection" in the inverted index;

[0190] Secondary retrieval: The query is encoded into a 768-dimensional semantic vector using Sentence-BERT, and the cosine similarity with the document fragment is calculated in the Milvus vector library to accurately locate the position of the management regulations;

[0191] Third-level structure positioning: Based on the document chapter tree structure, locate sub-clauses such as "Appearance inspection of metering box" and "Temperature measurement of wiring terminals".

[0192] The search results will be input into the finely tuned Llama2-7B large-scale model, and the final meter box inspection steps are as follows: 1. Check whether the box is damaged, rusted, and the door lock is intact; 2. Check whether the internal wiring is loose or shows signs of burning; 3. Verify the meter's operating status and its matching with the user's load. Precautions: Before opening the box, confirm there are no warning signs indicating liveness, and wear insulated gloves during operation.

[0193] The generated content is recognized and converted, and the system automatically queries for inspection work orders that meet the requirements in the background. The above steps are then embedded as "standard operating instructions" into the APP task list, transmitted to the mobile terminal, and the log recording function is enabled to support subsequent backtracking.

[0194] V. Load balancing and matching between work orders and inspection personnel

[0195] In some possible embodiments, the mobile operation assistance method for power marketing further includes: receiving work order information to be processed and obtaining dispatchable personnel information within each work order scheduling cycle. The work order information to be processed includes anomaly type, urgency, anomaly location coordinates, personnel skill requirements, and estimated time consumption information. The dispatchable personnel information includes the personnel's current location coordinates, personnel skills, and current personnel workload rate.

[0196] Based on the information of the work orders to be processed and the information of the schedulable personnel, an improved genetic algorithm is used to solve the load balancing problem and obtain a matching scheme between new work orders and schedulable personnel.

[0197] The matching scheme and pending work order information are pushed to the mobile work assistance terminals of dispatchable personnel.

[0198] The mobile operation assistance method for power marketing of the present invention can be implemented by a server and a mobile terminal. The scheduling and matching of work orders can be achieved by the server, while field operators obtain scheduling information such as matching schemes, pending work orders, routes, and on-site operation plans through the mobile terminal to assist in executing on-site operations. The mobile terminal is also used to receive and identify on-site instructions, transmit the identified instructions to the server, and perform work operation matching and generate an operation plan task list. Alternatively, the mobile terminal can independently perform work operation matching and operation plan generation after instruction recognition.

[0199] The server can receive pending work order information and obtain schedulable personnel information within each set work order scheduling cycle. Then, based on the pending work order information and schedulable personnel information, it uses an improved genetic algorithm to solve for load balancing, obtaining a matching scheme between pending work orders and schedulable personnel. The matching scheme and pending work order information are then pushed to the mobile work assistance terminals of the schedulable personnel. The pending work order information includes the exception type, urgency, exception location coordinates, personnel skill requirements, and estimated processing time. The schedulable personnel information includes the personnel's current location coordinates, skills, and current workload rate.

[0200] In some possible embodiments, the step of using an improved genetic algorithm to solve for load balancing based on the information of the work order to be processed and the information of the schedulable personnel to obtain a matching scheme between the new work order and the schedulable personnel includes: performing data modeling on the work order to be processed and the schedulable personnel based on the information of the work order to be processed and the information of the schedulable personnel; then scheduling the operation work orders and operators after data modeling based on the genetic algorithm; and using a load balancing optimization algorithm to optimize the unbalanced load of the schedule to generate the scheduled operation work orders.

[0201] In some possible embodiments, the data modeling includes multi-dimensional operation order core features and multi-dimensional worker features. A worker-operator matching matrix is ​​calculated using encoding and decoding of the multi-dimensional operation order core features and the multi-dimensional worker features. Then, the fitness function of the genetic algorithm is calculated using the multi-dimensional operation order core features and the multi-dimensional worker features. Based on the genetic algorithm, the worker-operator matching matrix and the fitness function are substituted into the scheduling optimization calculation to obtain the scheduled operation orders. The scheduled operation orders are then optimized for personnel load differences using a max-min fairness algorithm to generate the scheduled operation orders. The fitness function includes at least the path cost term, the skill matching penalty term, and the deadline violation penalty term. The multi-dimensional operation order core features include at least geographical location, skill requirements, operation duration, and deadline. The multi-dimensional worker features include at least current location, skill matrix, and historical efficiency.

[0202] The multidimensional core features of operation work orders include urgency, operation duration, geographical location, skill requirements, and deadline. Urgency is divided into 1-5 levels, operation duration is based on historical averages, geographical location is displayed using latitude and longitude coordinates, and skill requirements are represented by binary vectors. Personnel features include 8-dimensional parameters, such as current location, skill matrix, load rate, and historical efficiency. The load rate is the ratio of allocated working hours to maximum working hours, historical efficiency is the processing speed of similar work orders, urgency is directly related to scheduling priority, geographical location provides the basis for path cost calculation, and skill requirements and personnel skill matrix jointly determine the matching degree. These feature parameters will be directly integrated into the objective function of subsequent algorithms to ensure a close correlation between data and algorithms, while also accurately matching the operation scenario. This ensures data accuracy and improves the scenario adaptability of the data, guaranteeing that the data fits the operation process of the operators and the on-site situation.

[0203] The core features of the multi-dimensional work order include urgency, work duration, geographical location, skill requirements, and deadline. Urgency is divided into 1-5 levels, work duration is based on historical averages, geographical location is displayed using latitude and longitude coordinates, and skill requirements are represented by binary vectors. Multi-dimensional worker features include current location, skill matrix, workload rate, and historical efficiency. The workload rate is the ratio of allocated working hours to maximum working hours, historical efficiency is the processing speed of similar work orders, urgency is directly related to scheduling priority, geographical location provides the basis for path cost calculation, and skill requirements and the worker skill matrix jointly determine the matching degree. These feature parameters will be directly integrated into the objective function of the subsequent algorithm to ensure a close correlation between data and the algorithm. The specific process is as follows:

[0204] (1) Problem modeling and individual coding

[0205] Let the set of work orders to be assigned be... , For the Mth work order, each work order is associated with a work order feature vector. ,in The number of dimensions for work order characteristics, such as urgency, work duration, geographical location, skill requirements, and deadline;

[0206] Let the set of dispatchable operators be . , For the Nth work order, each person is associated with a person feature vector. ,in The number of personnel characteristic dimensions, such as current location, skill level, load status, historical efficiency, etc.

[0207] In genetic algorithms, an individual-chromosome representation represents a complete scheduling scheme:

[0208]

[0209] in Indicates that the work order Assigned to personnel .

[0210] (2) The role mechanism of multidimensional features in fitness calculation

[0211] The quality of individual X is determined by the fitness function. The evaluation function comprehensively reflects the performance of the scheduling scheme in terms of three core business objectives: path cost, capacity matching, and timeliness compliance. Its form is a weighted linear expression:

[0212]

[0213] in, For path cost items, Penalty for skill mismatch Penalties for breach of contract by the deadline The weight of the path cost item. The weighting of the penalty for skill mismatch. The weighting of penalties for breach of contract by the deadline. ,and All are greater than 0, and each weight reflects the degree of importance the business attaches to path efficiency, skill matching, and timeliness.

[0214] Path cost item The expression is as follows:

[0215]

[0216] Where N is the total number of operators that can be scheduled. Indicates the feature vector from the personnel The first in Individuals The current location of [the location]. Indicates the feature vector from the work order The first in Work order The execution of the geographical location, For Starting from the point of origin, traverse all the operators assigned to the task. A path planning function for the total time and / or distance to the said geographic location.

[0217] Skill Matching Penalty The expression is as follows:

[0218]

[0219] Where M is the total number of work orders. Work order feature vector The skill requirements extracted from the i-th operation order. To extract from personnel feature vectors The extracted first The skill matrix of the aforementioned operators, Denotes the L1 norm of the skill matrix. To prevent division by zero errors for extremely small positive numbers, A smaller value indicates a better match.

[0220] Penalties for breach of deadline The expression is as follows:

[0221]

[0222]

[0223] in, Calculate for the maximum value. This represents the estimated completion time for the i-th operation order. This represents the deadline for the i-th operation order. Indicates the time when scheduling starts. This represents the function representing the estimated travel time. Indicates the feature vector from the personnel No. Operators The current position, Indicates the feature vector from the work order The first in Work order The geographical location of the execution, This represents the operation duration of the i-th operation order. Indicates the first The historical efficiency of an operator is defined as the ratio of the historical average standard time to the actual time spent. >1 indicates high efficiency.

[0224] (3) Feature constraint treatment in genetic operations

[0225] Initialization and mutation phase: For any assignment, such as the i-th work order → For the j-th operator, check if hard constraints are met, such as skill coverage requirements. If not, resample for valid operators;

[0226] Crossover operation: After generating offspring, a feasibility check is also performed to ensure that all assignments are valid;

[0227] Elite retention: Retaining the lowest fitness, i.e. The smallest individual directly enters the next generation, accelerating convergence;

[0228] Load balancing optimization, as a supplementary adjustment mechanism to the algorithm results, iteratively adjusts the allocation of personnel and work orders based on the initial scheduling output by the genetic algorithm, using a maximum-minimum fairness algorithm. This addresses the issue of uneven personnel workload that may occur in the initial scheduling of the algorithm.

[0229] For example, the server receives 3 new work orders:

[0230] Work Order A: User has a low-voltage fault, urgency level 4, coordinates 118.38, 31.90, low-voltage skills required, estimated time 1.5 hours;

[0231] Work Order B: Meter burned out, urgency level 5, coordinates 118.39, 31.91, requires low-voltage skills, estimated time 2 hours;

[0232] Work Order C: Abnormal noise from the line, urgency level 3, coordinates 118.40, 31.89, requires high voltage skills, estimated time 3 hours.

[0233] Information on 3 dispatchable personnel:

[0234] Personnel A: Coordinates 118.38, 31.89, possesses high-voltage and low-voltage skills, workload rate 50%;

[0235] Personnel B: Coordinates 118.38, 31.92, possesses only low-voltage skills, workload rate 40%;

[0236] Personnel C: Coordinates 118.40, 31.90, possesses high-voltage and low-voltage skills, with a workload rate of 42%.

[0237] The solution is obtained using an improved genetic algorithm.

[0238] Encoding: Chromosome , Assign weights to work order A. Assign weights to work order B. Assign weights to work order C;

[0239] Initialize a population of 50 individuals and iterate for 50 generations;

[0240] Fitness function calculation:

[0241] Path cost, calculated based on the Haversine formula to determine the distance from the worker to the work order point;

[0242] Skill penalty: +10 if skills do not match;

[0243] Deadline penalty: +5 for exceeding the deadline;

[0244] ;

[0245] After iterative optimization, the optimal allocation scheme is output:

[0246] Personnel A → Work Order C: Short path, skill match;

[0247] Personnel B → Work Order B, highest urgency, unique skill match;

[0248] Personnel C → Work Order A, assigned based on proximity, load balancing.

[0249] Load balancing verification, calculating the load rate after allocation:

[0250] Personnel A: 50% + 3h / 8h = 87.5%;

[0251] Personnel B: 40% + 2h / 8h = 65%;

[0252] Personnel C: 42% + 1.5h / 8h = 60.75%.

[0253] The maximum difference is 26.75%, exceeding the 15% threshold. This triggers the maximum-minimum fairness algorithm, adjusting work order A to be executed by employee C, and work order C to be executed by employee A, thus optimizing load balancing.

[0254] Personnel A: 50% + 1.5h / 8h = 68.75%;

[0255] Personnel B: 40% + 2h / 8h = 65%;

[0256] Personnel C: 42% + 3h / 8h = 79.5%;

[0257] The maximum difference in load rate is 14.5%, which is lower than the 15% threshold, thus satisfying the constraint.

[0258] The optimal scheduling plan is pushed to the terminals of each operator via the APP, and the task markers on the map are updated simultaneously.

[0259] VI. Regarding on-site operation assistance

[0260] The server can also generate on-site operation assistance information for work orders to be processed based on GIS maps and on-site operation assistance models, which can be downloaded by operators through mobile operation assistance terminals; the on-site operation assistance information includes at least inspection route information and equipment defect identification and cause analysis information.

[0261] In some possible embodiments, the server generates an operation-to-be-executed map based on a GIS map, according to the scheduled work orders. The on-site operation support models involved include equipment defect identification models and path planning algorithm models.

[0262] The GIS map can be a GIS spatiotemporal map based on the OGC standard tile pyramid with a level of 0-19, and a custom power thematic layer, including transformer area boundaries, line routes, and equipment locations. It is stored through a PostGIS spatial database and supports spatial queries, range queries based on R tree indexes, and spatiotemporal queries. It supports map rendering, accelerated by WebGL, and supports dynamic labeling of equipment status, with green for normal status and red for abnormal status. The offline map package supports loading in offline environments.

[0263] The path planning algorithm uses the A* algorithm, and the cost function expression is:

[0264]

[0265]

[0266]

[0267] in, Let cost function be This represents the actual path cost incurred from the starting point to the current node n. For the heuristic cost estimation from the current node n to the target point, , , and for The weighting coefficients, This is the actual road length. This is a restricted area. Penalties for traffic congestion Penalties during peak electricity usage periods The straight-line distance from the current node to the target point. To estimate travel time, This is the device priority offset. , and This is the priority planning coefficient.

[0268] The defect identification model for equipment such as metering boxes is implemented based on the YOLOv8 model. The input image is enhanced with Mosaic, the backbone uses CSPDarknet, and a CBAM attention mechanism is added to the neck. Root cause reasoning is implemented using a three-layer Bayesian network structure, following the path from defect → direct cause → root cause. The conditional probability table is trained on over 50,000 historical cases, and the belief propagation algorithm is used to calculate the posterior probability. The model outputs the top two root causes and their processing steps to improve the accuracy of defect identification. It also provides operators with a reliable and accurate defect identification method, further improving work efficiency, reducing operator errors and omissions, and assisting new operators. It provides algorithmic and data support for new operator training and experience accumulation.

[0269] For example, taking a customer manager's daily field work as an example, the app opens and automatically locates the current position. The target address is entered, and the GIS engine is invoked; power distribution transformers and pole-mounted switches within 500 meters are queried; an improved A* algorithm is used to plan the route, with the cost function considering road grade, restricted areas, and peak electricity consumption times; the optimal route is output and navigation is provided.

[0270] To achieve equipment defect identification, after the inspection personnel arrive at the site, they use a terminal camera to photograph the metering box: the image is preprocessed by Mosaic data augmentation; it is input into the YOLOv8 model and CSPDarknet backbone + CBAM attention is used; two defects, "damaged box door" and "loose wiring", are detected, along with their corresponding confidence levels.

[0271] Furthermore, the server can continue to perform root cause analysis and processing suggestion analysis: the defect information is fed into a pre-trained Bayesian network: the first layer observation is a broken box door → the second layer direct cause is an external impact with a probability of 0.7 and a corrosion probability of 0.3; the third layer root cause is inadequate inspection, with a posterior probability of 0.65 and insufficient protection of 0.58; the output processing suggestion is "replace the metering box door, strengthen the surrounding protective railing, and include it in the monthly inspection list".

[0272] Then, the compliance verification of the power supply scheme can be carried out. The "DL / T 5220-2023" standard in the knowledge base is called, and combined with the user load of 80kW and voltage level of 10kV, the scheme is automatically generated to use a single-circuit overhead line with conductor type JKLYJ-70.

[0273] In addition, the RAG engine verification states: "According to Article 6.2.3 of DL / T 5220-2023, an 80kW load should be equipped with ≥50mm² conductors, and the current solution meets the requirements."

[0274] Finally, output a structured power supply scheme document, including clause references and verification results.

[0275] VII. Regarding encrypted information transmission between the server and the terminal

[0276] To ensure information transmission security, in some possible embodiments, the server generates and updates keys for each mobile work assistance terminal based on its unique identifier and dynamic password, according to a set rotation cycle. The mobile work assistance terminal uses these keys to transmit on-site work order processing data to the server via an encrypted transmission channel. The encrypted transmission channel uses the national cryptographic SSL protocol and the SM2 elliptic curve algorithm for key exchange between the mobile work assistance terminal and the server, employs the SM4 symmetric encryption algorithm for data encryption, and attaches a message digest generated by the SM3 algorithm, as well as using the SM3 algorithm for communication integrity protection. After receiving the encrypted data, the data receiver recalculates the message digest and compares it to verify whether the data has been tampered with during transmission.

[0277] During the key derivation process, the encryption keys for work orders and on-site operation auxiliary data are derived from the key materials using the SM3 hash algorithm, ensuring key uniqueness and unpredictability.

[0278] For encryption of on-site operational auxiliary data, the SM4-ECB mode can be used, which is suitable for short structured data such as work order status and equipment readings, with each record encrypted independently.

[0279] For scheduled work orders, SM4-CBC mode can be used for encryption, which is suitable for long and complex work order data and provides anti-replay and semantic obfuscation.

[0280] To ensure integrity, regardless of the encryption mode used, an SM3 message authentication code is added.

[0281] The key management system is built upon the national cryptographic protocol KMIP, covering the entire lifecycle management of keys. Keys are generated by a hardware security module, securely distributed to various terminal devices via an encrypted channel, and subject to a periodic automatic rotation mechanism with a seven-day rotation cycle. After expiration, the keys are securely destroyed.

[0282] The above method establishes an encrypted transmission channel by adopting the national cryptographic SSL protocol, combines the SM4 symmetric encryption algorithm to achieve data encryption, and uses the SM3 algorithm to provide communication integrity protection. It can completely replace the traditional TLS protocol and meet the security requirements of Level 3 of the Power Industry Network Security Protection 2.0.

[0283] The data verification mechanism adds a 256-bit message digest generated by the SM3 algorithm to the encrypted data. After receiving the data, the receiver recalculates the digest and compares it to verify that the data has not been tampered with during transmission, thus effectively ensuring the integrity of power data transmission.

[0284] Taking a customer manager's daily field work as an example, field data is encrypted and uploaded. After the work is completed, the work order data, including defect photos and processing records, is encrypted using SM4-ECB mode. The encrypted data is appended with a message digest generated by the SM3 algorithm and uploaded to the cloud via a national cryptographic SSL channel. This transmission channel uses the SM2 asymmetric algorithm for key exchange and the SM4 symmetric encryption algorithm to encrypt the communication content.

[0285] VIII. Regarding Model Updates

[0286] The mobile operation assistance method for power marketing of the present invention also includes: the server performing performance evaluation and model update of the field operation assistance model based on the field work order processing data accumulated within the current period window according to the set knowledge evolution cycle, and prioritizing knowledge recommendation of the power operation knowledge base.

[0287] In some possible embodiments, model deviation data and on-site operation feedback of the scheduled work orders can be collected periodically, and high-frequency abnormal data can be obtained by clustering the model deviation data and operation feedback using the DBSCAN clustering algorithm.

[0288] The projection matrix of BERT in the BERT-BiGRU-CRF semantic understanding model and the output projection layer of BiGRU and the output projection layer of the CNN-RNN hybrid recognition model are injected into and stored in the LoRA parameter set; the LoRA parameter set includes LoRA parameters. and header parameters ;

[0289] High-frequency outlier data is introduced using a federated learning algorithm to perform multiple optimization training iterations on the BERT-BiGRU-CRF semantic understanding model and the CNN-RNN hybrid recognition model, which store LoRA parameter sets. Each optimization training involves multiple participants. After each optimization training, the LoRA parameter sets of multiple participants are weighted and globally updated. Based on the updated LoRA parameter sets, the optimization iterations of the BERT-BiGRU-CRF semantic understanding model and the CNN-RNN hybrid recognition model are completed. The weighted expression for the LoRA parameter set is:

[0290]

[0291]

[0292] Where t represents the t-th training iteration, and K represents the number of participants in the training. Let be the number of valid high-frequency anomaly samples in the i-th power. Let j be the number of valid high-frequency anomaly samples. Let LoRA be the parameter of the i-th power. For the i-th side, is the header parameter.

[0293] The data analysis process employs a sliding window approach, with a seven-day cycle to continuously collect multi-dimensional information. This includes deviations in model predictions (e.g., cases of incorrect defect identification), user feedback during interactions (e.g., marking invalid query results), and trends in key business metrics (e.g., fluctuations in work order processing times). The DBSCAN clustering algorithm is used to group and analyze the collected anomaly data, setting reasonable neighborhood distances and minimum sample thresholds to identify frequently occurring problem scenarios, such as low accuracy in identifying specific types of defects.

[0294] Model fine-tuning is implemented through a federated learning architecture. Each terminal node uses locally accumulated anomalous data to train the model individually, employing the FedProx optimization algorithm to control the regularization strength during training, with a regularization parameter μ=0.01. After training, only gradient update information containing Gaussian noise is uploaded, with Gaussian noise σ=0.1 added to meet differential privacy protection requirements with relatively small random perturbations. The central server assigns appropriate weights based on the data scale of each terminal, performs weighted average aggregation of the uploaded gradient information, and updates the global model parameters using the Adam optimizer. The learning rate and momentum parameters are fine-tuned, and a complete model fine-tuning process is executed every three days. Through iterative optimization of the model, not only can the accuracy of speech recognition, work order generation, and job assistance be improved, but also the robustness and applicability of the model in different scenarios can be enhanced.

[0295] For example: the local terminal detects a decrease in the accuracy of the "loose wiring" identification and marks it as an abnormal sample;

[0296] Locally, the FedProx algorithm was used with μ=0.01 to fine-tune the YOLOv8 model, the gradient was uploaded, and Gaussian noise σ=0.1 was added.

[0297] The central node aggregates gradients from multiple terminals, performs a weighted average, and then updates the global model.

[0298] In some possible embodiments, operator feedback is collected, including but not limited to the number of clicks, favorites / tags, actual number of times applied to work order processing, dwell time, and the relevant business scenario. Based on this feedback, a relevance score for the corresponding power knowledge graph is calculated to generate ranking tags; for example:

[0299]

[0300] in, This represents the relevance score of the i-th knowledge item in the power knowledge graph. This represents the number of clicks on the i-th knowledge entry in the electricity knowledge graph. This represents the number of times the i-th knowledge entry in the power knowledge graph is actually applied to work order processing. For the i-th knowledge entry in the power knowledge graph, add it to your favorites / mark it. This represents the negative business feedback for the i-th knowledge entry in the power knowledge graph. Preset weights for click counts. The preset weights for the number of times the work order is actually processed. Preset weights for favorites / tags Preset weights for negative business feedback, such as , , , ;

[0301] Then, the relevance scores are sorted in descending order to obtain the sorting labels.

[0302] Training samples are generated from the power knowledge graph based on the ranking labels, and multi-dimensional features are extracted, for example:

[0303] Static characteristics: knowledge type, creation time, device category, etc.;

[0304] Dynamic features: click-through rate over the past 7 days, application scenario matching degree, such as semantic similarity with the current work order calculated by BERT, etc.;

[0305] User profile features: historical preference vectors of the current operator, etc.;

[0306] Contextual features: current work order type, geographical location, urgency, etc.;

[0307] The feature vectors are normalized and then input into the knowledge word ranking model.

[0308] A knowledge word ranking model is established based on the Adam optimizer, and a loss function for the knowledge word ranking model is constructed based on the LambdaRank algorithm. The multidimensional features of the power knowledge graph are then optimized and iterated based on the knowledge word ranking model.

[0309] The core metrics of the prompt quality assessment model include task completion rate, which is whether the generated content conforms to business specifications, and user correction rate, which reflects the frequency of human intervention in prompts.

[0310] Knowledge recommendation rule re-ranking involves collecting behavioral data on knowledge clicks, favorites, and applications; constructing multi-dimensional feature vectors such as click counts and application scenario matching; and iteratively optimizing these vectors to improve usability for subsequent users. The LambdaRank algorithm constructs the loss function as follows:

[0311]

[0312]

[0313] in, The loss value measures the deviation between the ranking model's predicted results and the actual ranking. For knowledge i, For knowledge j, To determine the true difference in relevance between knowledge i and knowledge j, Indicators of ranking quality after swapping knowledge i and knowledge j ( The difference in impact reflects the severity of the sorting error; the larger the value, the greater the severity. For the ranking level of knowledge i, The higher the ranking of knowledge j, the more necessary the model is to be revised. For model-based predicted scores and The Logistic function is modified to output probability values ​​between 0 and 1. Knowledge recommendation priorities are updated weekly to ensure frequently used knowledge is featured prominently.

[0314] The overall loss function is the sum of the losses for all valid sample pairs:

[0315]

[0316] in, The sum of losses for the effective sample pairs. For all knowledge i and knowledge j in knowledge graph P, Let be the loss value for knowledge i and knowledge j.

[0317] For example: A decrease in the completion rate of the "Power Supply Scheme Generation" task and an increase in the user correction rate were detected.

[0318] Optimize the prompt template: "Please generate a 10kV power supply scheme for the {load} kW user according to DL / T 5220-2023, including the output conductor type and safety verification";

[0319] Based on the LambdaRank algorithm, frequently used knowledge such as "metering box inspection process" and "low-pressure fault handling" are recommended at the top.

[0320] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0321] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0322] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0323] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0324] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A mobile operation assistance method for electricity marketing, characterized in that, include: Receive voice commands from mobile power marketing work sites; After preprocessing the voice commands, the command text is extracted using a hybrid recognition model; The semantic understanding model based on the power knowledge graph is used to analyze the operation matched by the instruction text; Based on the power operation knowledge base, the RAG retrieval enhancement engine is used to generate operation plans that match the operation, and the operation plans are matched to the inspection work orders corresponding to the voice commands to generate a task list corresponding to the inspection work orders.

2. The mobile operation assistance method for electricity marketing according to claim 1, characterized in that, it also... include: Within each work order scheduling cycle, the system receives work order information to be processed and obtains information on schedulable personnel. The work order information to be processed includes the type of abnormality, urgency, coordinates of the abnormality location, personnel skill requirements, and estimated time consumption. The information on schedulable personnel includes the personnel's current location coordinates, personnel skills, and current personnel workload information. Based on the information of the work orders to be processed and the information of the schedulable personnel, an improved genetic algorithm is used to solve the load balancing problem and obtain a matching scheme between the work orders to be processed and the schedulable personnel. The matching scheme and pending work order information are pushed to the mobile work assistance terminals of dispatchable personnel.

3. The mobile operation assistance method for electricity marketing according to claim 2, characterized in that, The step of using an improved genetic algorithm to solve for load balancing based on the information of work orders to be processed and the information of schedulable personnel, and obtaining a matching scheme between work orders to be processed and schedulable personnel, includes: (1) Data modeling is performed on the work orders to be processed and the personnel to be dispatched based on the information of the work orders to be processed and the personnel to be dispatched; (2) Modeling the load balancing problem and encoding individual chromosomes, including: Let the set of work orders to be assigned be... , For the Mth work order, each work order is associated with a work order feature vector. ,in The number of feature dimensions for the work order; Let the set of dispatchable operators be . , For the Nth work order, each schedulable personnel is associated with a personnel feature vector. ,in The number of dimensions for personnel characteristics; In a genetic algorithm, a chromosome represents a complete scheduling scheme. in This indicates that pending work orders will be processed. Assign to dispatchable personnel; The quality of chromosome individual X is determined by the fitness function. The fitness function, used for evaluation, reflects the scheduling scheme's performance in terms of path cost, capacity matching, and timeliness compliance. Its expression is: in, For path cost items, Penalty for skill mismatch Penalties for breach of contract by the deadline The weight of the path cost item. The weighting of the penalty for skill mismatch. The weighting of penalties for breach of contract by the deadline. ,and All are greater than 0, and each weight reflects the degree of importance the business attaches to path efficiency, skill matching, and timeliness. (3) Initialize the population based on the pending work order data model, the schedulable personnel data model and the load balancing model, and calculate the fitness function of each chromosome in the initial population; (4) Based on the fitness function calculation results, perform selection, crossover and mutation operations to obtain a new generation of better populations and calculate the fitness function until the calculation results meet the iteration termination condition. Then select the individual with the highest fitness value as the final load balancing scheme. If the calculation results do not meet the iteration termination condition, repeat step (4).

4. The mobile operation assistance method for electricity marketing according to claim 3, characterized in that, In the fitness function, the path cost term The expression is: Where N is the total number of operators that can be scheduled. Indicates the feature vector from the personnel The first in The current location of each person. Indicates the feature vector from the work order The first in Work order The execution of the geographical location, For Starting from the point of origin, traverse all the operations assigned to the operators. A path planning function for the total time and / or distance to the said geographic location; Skill Matching Penalty The expression is as follows: Where M is the total number of work orders. Work order feature vector The skill requirements extracted from the i-th operation order. To extract from personnel feature vectors The extracted first The skill matrix of the aforementioned operators, Denotes the L1 norm of the skill matrix. To prevent division by zero errors for extremely small positive numbers, ; Penalties for breach of deadline The expression is: in, This indicates taking the maximum value. This represents the estimated completion time for the i-th operation order. This represents the deadline for the i-th operation order. Indicates the time when scheduling starts. This represents the function indicating the estimated travel time. Indicates the feature vector from the personnel No. One operator The current position, Indicates the feature vector from the work order The first in Work order The geographical location of the execution, This represents the operation duration of the i-th operation order. Indicates the first The historical efficiency of an operator is defined as the ratio of the historical average standard time to the actual time consumed.

5. The mobile operation assistance method for electricity marketing according to claim 1, characterized in that it further... include: The server generates on-site operation assistance information for work orders to be processed based on GIS maps and on-site operation assistance models, which can be downloaded by operators through mobile operation assistance terminals. The on-site operation auxiliary information includes at least inspection route information and equipment defect identification and cause analysis information; According to the set rotation cycle, the server generates and updates the key of each mobile operation auxiliary terminal based on the unique identifier and dynamic password of the mobile operation auxiliary terminal device. The mobile operation auxiliary terminal uses the key to transmit the on-site work order processing data to the server through an encrypted transmission channel. The encrypted transmission channel uses the national cryptographic SSL protocol and the SM2 elliptic curve algorithm for key exchange between the mobile operation auxiliary terminal and the server. It uses the SM4 symmetric encryption algorithm to encrypt data and attaches a message digest generated by the SM3 algorithm. It also uses the SM3 algorithm to protect communication integrity. After receiving the encrypted data, the data receiver recalculates the message digest and compares it to verify whether the data has been tampered with during transmission.

6. The mobile operation assistance method for electricity marketing according to claim 5, characterized in that, The on-site work order processing data includes defect identification records, knowledge query logs, and work order processing results. The mobile operation assistance method for power marketing also includes: the server-side performs performance evaluation and model update of the field operation assistance model based on the field work order processing data accumulated within the current period window, according to the set knowledge evolution cycle, and the knowledge recommendation priority of the power operation knowledge base.

7. The mobile operation assistance method for electricity marketing according to claim 1, characterized in that, Preprocessing the voice commands includes: The high-frequency and low-frequency coefficients of the speech command signal are obtained by five-layer discrete wavelet decomposition, and the threshold is calculated based on the VisuShrink criterion. : in, The threshold value is... The standard deviation of noise. The length of the voice command signal. It is a logarithmic function. These are the detail coefficients obtained from the first-level wavelet decomposition. A function to obtain the median; Based on the threshold The detail coefficients of each layer are subjected to soft thresholding, expressed as follows; in, These are the detail coefficients after soft thresholding. , , This indicates taking the larger value. is the sign function, indicating the positive or negative direction of the coefficients to be preserved; The speech command signal is reconstructed based on the results of soft thresholding, using the following formula: In the formula, For the original voice command signal The reconstructed voice command signal, These are the approximate coefficients obtained from the 5th level wavelet decomposition; The reconstructed voice command's voice frame is converted into a Mel spectrogram.

8. The mobile operation assistance method for electricity marketing according to claim 1, characterized in that, The semantic understanding model based on the power knowledge graph includes a semantic parsing module, a knowledge fusion module, an attention fusion module, and a parsing output module; The step of analyzing the job operation matched by the instruction text using a semantic understanding model based on power knowledge graphs includes: The instruction text is cleaned, segmented, and tokenized to generate a standardized token sequence; The token sequence is input into the semantic parsing module implemented using the BERT-BiGRU-CRF semantic understanding model. The BERT layer performs deep contextual semantic encoding on the token sequence to generate a semantic vector. The BiGRU layer performs bidirectional sequence dependency modeling on the semantic vector. The CRF layer performs global optimal label decoding and outputs an intent classification label and a structured entity label sequence containing job object entities and job type entities. The structured entity tag sequence is input into the knowledge fusion module and linked with the power knowledge graph. The text entities are mapped to standard equipment nodes and standard operation nodes in the power knowledge graph. The standardized operation process under the standard equipment node is queried from the power knowledge graph to obtain the standardized operation process corresponding to the standard operation node and generate the corresponding subgraph embedding vector. The semantic vector, subgraph embedding vector, and historical task context vector are input into the attention mechanism module. Through attention weight calculation, a knowledge context vector, a dialogue context vector, and a fusion representation vector containing the current semantics, domain knowledge, and historical task status are generated. Based on the fusion representation vector, the task object, task type and standard operation process are bound together, and the standardized task object, standardized task type and corresponding standardized operation process corresponding to the instruction text are output.

9. The mobile operation assistance method for electricity marketing according to claim 8, characterized in that, The attention mechanism module performs attention weight calculations including calculating dynamic attention weights for the dialogue context based on the standard device node and standard job node information in the subgraph embedding vector, using the following formula: in, For the current command and the history number Similarity of turn-based dialogue commands, The corresponding normalized dialogue context dynamic attention weights; For An exponential function with base 0; The query vector represents the job object and job type labels output by the BERT-BiGRU-CRF semantic understanding model for the current instruction. The key vector is the historical [number]. The round-robin dialogue instructions are identified and output by the BERT-BiGRU-CRF semantic understanding model, which identifies the job object and job type labels. For the total number of rounds of historical dialogue, This indicates the calculation of the similarity score.

10. The mobile operation assistance method for electricity marketing according to claim 1, characterized in that, The power operation knowledge base adopts a hybrid storage mode of structured and unstructured data. Structured data containing equipment parameters and formulas is stored in a relational database, while unstructured data containing regulations, documents, and cases is stored in the Milvus vector database after extracting key sentences using TextRank. A three-level index is also established, where the first-level index contains an inverted index of the professional domain thesaurus; the second-level index contains a 768-dimensional semantic vector generated by Sentence-BERT; and the third-level index contains a document chapter structure tree. The method of generating operation plans that match the operation based on the power operation knowledge base and using the RAG retrieval enhancement engine includes: Enhanced query statements are constructed by integrating information about the job object, job type, and job process. The enhanced query statement is input into the encoder to generate a query vector. Based on the query vector, a similarity retrieval is performed in a pre-built power operation knowledge base to recall a set of associated document fragments that match the operation object and operation type. Based on at least one of the factors of equipment matching degree, job type matching degree, and document timeliness, the set of related document fragments is reordered to filter highly relevant document fragments. The highly relevant document fragments are segmented and spliced ​​according to process stages to form a reference context; The reference context, job object, job type, and job process are input into a generator implemented using a large language model to generate an operation plan matching the job operation. The operation plan is organized by job process stages, including key points for each stage, safety risk warnings, emergency response measures, and a compliance checklist. The output format can be restricted to: clause reference + key points + precautions.

Images