Electric power communication maintenance ticket approval method and system

Abstract

The invention discloses a power communication maintenance ticket approval method and system. The method comprises the steps of obtaining to-be-approved electric power communication maintenance ticket data; performing semantic understanding and information extraction on to-be-approved electric power communication maintenance ticket data by using a neural network module, and outputting structured semantic information; inputting the structured semantic information into a symbol reasoning module, performing logical reasoning and compliance verification by the symbol reasoning module based on a differentiable symbol rule base, and outputting an approval result; the method comprises the following steps: converting an electric power communication maintenance ticket approval rule into a symbolic logic expression with a weight by adopting a differentiable logic operator, and constructing a differentiable symbolic rule base; and constructing a double-layer optimization framework, and jointly optimizing parameters of the neural network module and weights of rules in the differentiable symbol rule base based on upper-layer approval task loss and lower-layer symbol rule logic consistency loss. According to the method, the examination and approval efficiency, accuracy and auditing performance are remarkably improved, and the method is suitable for intelligent processing of power grid communication maintenance business.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence application technology in power systems, specifically relating to a method and system for approving power communication maintenance tickets, which is applicable to the intelligent processing of power grid communication maintenance business. Background Technology

[0002] As a core business document ensuring the safe and stable operation of the power grid, the communication maintenance ticket is a key node in the entire maintenance process, and its approval process plays a vital role in ensuring the safe and compliant conduct of maintenance work.

[0003] Currently, the approval process for maintenance tickets mainly relies on manual labor. Approving personnel must use their professional knowledge and experience, combined with complex procedural documents, to verify each item on the ticket. This model has significant limitations: manual approval is inefficient and cannot meet the high-efficiency operation and maintenance needs of modern power grids; the rule system is complex, covering multiple constraints, making it difficult for humans to fully grasp, which can easily lead to misjudgments or missed inspections; and approval decisions heavily depend on the subjective judgment of personnel, as different approving personnel have different understandings of the ticket content and different grasps of the rule standards, directly affecting the objectivity and consistency of the approval conclusions.

[0004] Existing technologies include several intelligent approval schemes based on large language models, attempting to improve approval efficiency through the semantic understanding capabilities of neural networks. However, these methods suffer from fundamental flaws in practical applications: First, the models lack symbolic reasoning capabilities, failing to transform typical maintenance ticket templates, business ledgers, and other approval criteria into executable logical rules, resulting in insufficient accuracy in verifying the compliance of maintenance ticket content. Second, the model's decision-making process lacks interpretability, making it difficult to establish a traceability path between approval conclusions and specific rules and regulations, thus failing to meet the requirements of power communication security audits. Third, the models are insufficient in handling uncertainties during the approval process, prone to overconfidence or model illusion, affecting their reliability in real operation and maintenance environments. These flaws prevent substantial improvements in maintenance ticket approval efficiency and fail to effectively control operational risks caused by misjudgments and omissions, thus hindering large-scale application in power communication operation and maintenance systems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides a method and system for approving power communication maintenance tickets, enabling efficient, accurate, and interpretable intelligent approval of maintenance tickets.

[0006] The present invention adopts the following technical solution.

[0007] The first aspect of the present invention provides a method for approving power communication maintenance tickets, comprising: Obtain data on pending power and communication maintenance tickets; The neural network module is used to perform semantic understanding and information extraction on the power communication maintenance ticket data to be approved, and output structured semantic information. Structured semantic information is input into the symbolic reasoning module, which performs logical reasoning and compliance verification based on a differentiable symbol rule base, and outputs the approval result. Among them, differentiable logical operators are used to transform the power communication maintenance ticket approval rules into weighted symbolic logical expressions, and a differentiable symbolic rule library is constructed. A two-layer optimization framework is constructed, which jointly optimizes the parameters of the neural network module and the weights of the rules in the differentiable symbolic rule base based on the upper-layer approval task loss and the lower-layer symbolic rule logic consistency loss.

[0008] Optionally, the neural network module can be pre-trained using specialized texts in the communications field through a masked language modeling task; Acquire multi-source data from the power communication operation and maintenance system, and annotate the multi-source datasets to construct a training set; The neural network module is fine-tuned using the training set, and the fine-tuned neural network module is used as the initial policy for reinforcement learning. Based on the initial policy, the neural network module is optimized through a compound reward function.

[0009] Optionally, the composite reward function includes a first reward function, a second reward function, and a third reward function. The first reward function is used to evaluate the performance of the neural network module in the approval task, the second reward function is used to evaluate the rationality of the logical reasoning path of the neural network module, and the third reward function is used to encourage the neural network module to explore diverse reasoning paths.

[0010] Optionally, the loss function in the pre-training process includes cross-entropy loss and symbolic rule consistency loss. The symbolic rule consistency loss function is introduced to enhance the logical output. The symbolic rule consistency loss is calculated based on whether the structured semantic information output by the neural network module is consistent with the logical reasoning result of the symbolic reasoning module.

[0011] Optionally, labeling multi-source datasets includes: Based on typical communication maintenance ticket templates and safety procedures, formalized symbol rules are defined. For each historical maintenance ticket, mark the approval result; mark the satisfaction status of each historical maintenance ticket for each symbol rule; mark the key semantic entities extracted from the historical maintenance ticket text, and establish the mapping relationship between key semantic entities and symbol rules.

[0012] Optionally, calculating the lower-level symbolic rule logical consistency loss includes: Calculate the probability distribution of the output with respect to the symbolic variable given the input text; Calculate the degree to which the rules corresponding to the symbolic variables are satisfied based on the probability distribution; The complement of the satisfaction level is used as a measure of the distance between the probability distribution of the symbolic variable and the distribution of the corresponding rule; The expected value of the sum is calculated by weighting and summing the distribution distance metric and the corresponding rule weights, and then the logical consistency loss of the lower-level symbolic rule is obtained.

[0013] Optionally, in the two-layer optimization framework, a dynamically decaying entropy regularization term is added to the objective function consisting of the upper-layer approval task loss and the lower-layer symbolic rule logic consistency loss. Here, the information entropy of the probability distribution of the output with respect to the symbolic variables under the given input text is calculated, the expected value of the information entropy is calculated and multiplied by the dynamic decay coefficient to obtain the entropy regularization term, and the final total loss function is obtained by subtracting the entropy regularization term from the objective function.

[0014] Optionally, the method further includes: A multi-agent collaborative structure is established to output approval results through multi-agent collaboration. The multi-agent collaborative structure includes a central control agent and specialized agents. The central control agent distributes approval tasks to the corresponding specialized agents. After each specialized agent executes the approval task, it outputs the review results including confidence assessment and the inference path and sends them to the central control agent. The central control agent outputs the final approval result based on the approval results of each agent. Intelligent agents are constructed based on optimized neural network modules and symbolic reasoning modules. Each specialized intelligent agent corresponds to a differentiable symbol rule base for different tasks. According to different review tasks, the corresponding differentiable symbol rule base and prompt word template are loaded so that each specialized intelligent agent can perform different approval tasks.

[0015] Optional, specialized intelligent agents include: The basic information checking agent is used to check the basic information of the maintenance ticket, including the title, source of the maintenance ticket, maintenance level, and maintenance start and completion time. The job content review agent is used to review the completeness, accuracy, and compliance of job content. The business impact verification agent is used to verify the consistency between the impact on business and the business ledger.

[0016] A second aspect of the present invention provides an intelligent approval system for power communication maintenance tickets, used to implement the above-mentioned intelligent approval method for power communication maintenance tickets, the system comprising: The acquisition module is used to acquire power communication maintenance ticket data pending approval. The semantic understanding and information extraction module is used to perform semantic understanding and information extraction on the power communication maintenance ticket data to be approved using the neural network module, and output structured semantic information. The logical reasoning module is used to input structured semantic information into the symbolic reasoning module. The symbolic reasoning module performs logical reasoning and compliance verification based on a differentiable symbol rule base and outputs the approval result. The conversion module is used to convert the power communication maintenance ticket approval rules into weighted symbolic logic expressions using differentiable logical operators, and to build a differentiable symbolic rule library. The optimization module is used to construct a two-layer optimization framework, which jointly optimizes the parameters of the neural network module and the weights of the rules in the differentiable symbolic rule base based on the upper-layer approval task loss and the lower-layer symbolic rule logic consistency loss.

[0017] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when loaded onto the processor, implements the above-described power communication maintenance ticket approval method.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described power communication maintenance ticket approval method.

[0019] Compared with the prior art, the beneficial effects of the present invention include at least the following: First, approval efficiency is significantly improved. By constructing a fully automated approval architecture, the processing time for a single maintenance ticket has been reduced from an average of about 12 minutes to 30 seconds. Second, approval accuracy is significantly improved. Through a neural symbol dual verification mechanism, the approval accuracy rate reaches 96%, which is about 11% higher than traditional methods. Third, the decision-making process is interpretable. A complete traceability chain is established from text features to symbolic concepts, from symbolic concepts to logical rules, and from logical rules to approval conclusions, meeting the requirements of power communication security audits. Fourth, approval quality is strictly reliable. Through a nondeterministic management mechanism, the system can identify and mark low-confidence approval scenarios and automatically trigger manual review, thereby effectively controlling operational risks. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a schematic diagram of a power communication maintenance ticket approval method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the architecture of a power communication maintenance ticket approval system provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0022] Combination Figure 1 As shown, Embodiment 1 of this application provides a method for approving power communication maintenance tickets, including the following steps: Step 1: Acquire multi-source data from the power communication operation and maintenance system, and construct a training set based on the multi-source data. Step 1 specifically includes: Step 1.1: Collect and clean various raw data from the power communication operation and maintenance system to form a unified and standardized data source.

[0023] The original data sources include a historical maintenance ticket database, a standard template database, a business resource ledger, and a regulations and rules knowledge base. The historical maintenance ticket database collects all archived maintenance tickets from the past two years, ensuring coverage of all types, including planned and temporary maintenance, and includes the final approval results. The standard template database includes typical communication maintenance ticket templates, including but not limited to three major categories: optical cables, power supplies, and equipment. The business resource ledger includes complete business resource data, including but not limited to business types such as dispatch data networks and integrated data networks. Each business entry includes key information such as the mode order number, business name, bearer system, and protection mode. The regulations and rules knowledge base includes regulations and documents such as the "Communication Maintenance Management Measures" and the "Communication Maintenance Classification and Grading Standards."

[0024] Fill in missing values ​​and correct outliers in historical maintenance tickets; establish a mapping dictionary between business ledgers and equipment and site names in maintenance tickets to resolve naming inconsistencies; divide procedure documents into paragraphs and extract key clauses to prepare for subsequent rule formalization.

[0025] Step 1.2: Label the cleaned data and construct the training set.

[0026] To achieve effective learning in the neural symbolic system, the core task of this step is to assign machine-understandable semantic and logical labels to the cleaned data, forming multi-layered, structured "supervision signals." A systematic three-level annotation process is executed for each maintenance ticket, ultimately constructing a high-quality training dataset. The three-level annotation process includes: First, result-level annotation is performed to provide the system with macro-level decision-making objectives. This level annotates the final approval results of maintenance tickets, including but not limited to approval or rejection for modification. The annotation method is mainly based on historical approval records, and questionable cases are reviewed by domain experts to supervise the overall accuracy of the model's task decisions.

[0027] Then, rule-based annotation is performed to provide direct logical supervision for the symbolic reasoning module. This layer formalizes the requirements in typical service tickets and the regulations in safety procedures into specific logical rules, and domain experts determine which specific rules the service ticket meets or violates, thereby guiding the system to learn the reasons behind decisions, rather than just the results.

[0028] Finally, semantic layer annotation is implemented to establish a stable and interpretable correspondence between the cleaned data perceived by the neural network module and the abstract symbols of the symbolic reasoning module. Semantic entities are extracted from the cleaned data and mapped to predefined symbols, achieving "neural symbol grounding". This layer extracts key semantic entities from the natural language description of the maintenance ticket, including but not limited to "500kV", "Bihu Station", "optical cable", and "SNCP", and maps them one by one to predefined symbolic concepts, such as VoltageLevel (500), Site (BiHu), EquipmentType (Fiber), and ProtectionMode (SNCP). This work uses named entity recognition tools for initial screening, followed by manual calibration and confirmation by experts, ultimately constructing an accurate text-to-symbol mapping table to ensure that the results perceived by the neural network module can be accurately understood and processed by the symbolic system.

[0029] Through the joint annotation of the above three levels, the original maintenance ticket is transformed into a structured training sample, as follows: (Passed, [Rule 2.4: Satisfied, Rule 3.2: Satisfied, ...], [EquipmentType(...), Site(...), ...]) After obtaining labeled samples, they are serialized according to a unified format to complete the construction of the training set. This training set guides the neural network module to learn how to extract correct symbolic concepts from text and guides the symbolic reasoning module to learn how to use rules to perform effective logical reasoning on these concepts.

[0030] In this embodiment, a high-quality training dataset for neural symbolic systems to learn from is constructed through systematic data collection, governance, and annotation, laying the foundation for subsequent model training and system optimization.

[0031] Step 2: After pre-training the neural network module using professional texts in the field of power communication, the neural network module is then supervised fine-tuned and reinforced using the training set constructed in Step 1.

[0032] The neural network module can be a large language model. This step aims to build a language understanding model specifically for the power communication field, achieving deep adaptation to domain knowledge and enhanced logical reasoning capabilities through a progressive training strategy. This stage adopts a two-stage training architecture to ensure that the model, in addition to mastering professional terminology and business logic, further possesses symbolic reasoning consistency. Step 2 specifically includes: Step 2.1: Based on the general large language model, incremental pre-training is carried out using large-scale professional texts in the field of power communication to establish a solid domain knowledge foundation.

[0033] It is understandable that the professional texts in the field of power communication are general corpora in the field of power communication, including but not limited to knowledge related to maintenance tickets.

[0034] Specifically, the masked language modeling task enables the model to learn the specialized vocabulary and expression patterns unique to power communications, with the optimization objective being:

[0035] in, The loss function represents the masked language modeling task and is used to measure the model's performance in text reconstruction tasks; The input text sequence is derived from the training set. One example of sampling is that the training set C consists of professional texts in the field of power communications; It is the set of locations of randomly selected and masked words in a text sequence. This indicates the context information visible except for the masked location; Is the model in a given context Under the condition of predicting the masked position Correct vocabulary The probability, This means calculating the mathematical expectation of the input sample x sampled from the training set C, ensuring the unbiasedness of the loss calculation.

[0036] By minimizing the loss function of the masked language modeling task, the model is forced to learn the ability to infer masked words from context, thereby gaining a deeper understanding of the technical terms and language patterns in the field of power communications.

[0037] In this embodiment, the masked language modeling task utilizes professional texts in the field of power communications to learn high-quality language representations. Through the mask-prediction self-supervised task, the large language model captures the contextual dependencies between words in the text, thereby providing quantifiable semantic anchors for the abstract concepts of entities and relations in the symbol system.

[0038] Step 2.2: Based on the domain knowledge injection, the training set constructed in Step 1 is used to further perform supervised fine-tuning and reinforcement learning on the large language model. By combining supervised fine-tuning and reinforcement learning, the logical reasoning ability and symbol rule consistency of the large language model are improved.

[0039] The basic decision-making and logical alignment of the model are trained using labeled data. Supervised fine-tuning and reinforcement learning are combined sequentially: first, supervised fine-tuning is performed to lay the foundation for the model; then, the fine-tuned model is used as the initial policy for reinforcement learning, and the inference process is further optimized through a reward function (combining task accuracy, logical rationality, and exploration diversity) to improve generalization ability and robustness. Step 2.2 specifically includes: First, a multi-task learning framework is constructed, introducing symbolic rule consistency loss on top of traditional classification tasks. The composite loss function of this framework is defined as:

[0040] in, The overall loss function representing multi-task learning, It is the standard cross-entropy loss. It is the weight of the symbolic rule consistency loss. It is the symbolic rule consistency loss, calculated as follows:

[0041] in, This indicates that the neural network module responds to the input. The semantic output representation, The symbolic reasoning module represents the same input. The representation of the logical judgment result in vector space. This represents the square of the Euclidean distance.

[0042] Cross-entropy loss is used to supervise and optimize the final decision of the task, i.e., whether the approval result is correct. Symbolic rule consistency loss is used to supervise and optimize whether the intermediate semantic representation is aligned with the logical judgment of the symbolic reasoning module. Symbolic rule consistency loss is an auxiliary alignment loss used in the fine-tuning stage of the large language model. Its purpose is to make the internal representation of the large language model closer to the logical judgment of the symbolic reasoning module. It belongs to the separate training of the large language model before joint training.

[0043] This composite loss term forces the semantic output of the neural network to be consistent with the logical judgment of the symbolic reasoning module, ensuring that the model learns reasoning patterns that conform to domain rules.

[0044] Furthermore, a reinforcement learning mechanism based on the integrity of the reasoning path is designed to circumvent the reasoning shortcut behavior of large language models. The reward function is defined as:

[0045] in, Represents the overall reward value, used to guide strategy optimization during the reinforcement learning process; It is a reward that measures the accuracy of task completion and is evaluated based on the performance of the large language model in approval tasks; It is a reward for evaluating the rationality of the logical reasoning path, calculated based on the degree of conformity between the reasoning process of the large language model and the symbol rule base; It is the policy entropy regularization term. ,in Indicates the state Select action The policy probability is used to encourage large language models to explore diverse reasoning paths and prevent them from prematurely converging to local optima inference shortcuts. , and These are the weighting coefficients for each reward.

[0046] During this training phase, by designing a loss function and constructing a reward mechanism, we ensure that the large language model, while mastering professional knowledge in the field of power communication, possesses logical reasoning capabilities that are highly consistent with the symbol rule base, thus laying a solid foundation for the subsequent collaborative processing of the neural network module and the symbol reasoning module.

[0047] Step 3: Use differentiable logical operators to transform the natural language rules in power communication maintenance approval into weighted symbolic logical expressions, and construct a differentiable symbolic rule base.

[0048] This step aims to transform the natural language rules used in power communication maintenance approval into a computable and learnable symbolic system, and establish a formal reasoning mechanism that supports gradient optimization. Step 3 specifically includes: Step 3.1: Transform the approval rules into machine-understandable symbolic logic expressions. Based on the requirements in typical communication maintenance tickets and regulations such as the "Power and Communication Maintenance Management Measures" and the "Communication Maintenance Classification Standards," establish a formalized rule knowledge base.

[0049] First-order logic is used to transform approval rules into implication expressions of the "IF-THEN" form. For example, the rule "If the maintenance time exceeds 8 hours, the specific duration of business disruption must be clearly stated in the maintenance work content" can be formalized as:

[0050] Where Duration(x) represents the maintenance duration of maintenance ticket instance x. This shows an example of a maintenance ticket. This describes the duration of the business disruption. For maintenance ticket example x, the specific duration y of the business disruption must be clearly described in the maintenance content.

[0051] Handling complex conditions: For rules containing multiple conditions, logical conjunction is used for combination. For example, the rule "If maintenance involves detours, interruptions, or restorations, then the maintenance classification should be major maintenance" can be expressed as:

[0052] Rule weight configuration: Introduce trainable importance weight parameters for each rule. This reflects the relative importance of rules in approval decisions. Weights are learned automatically through training data, supporting dynamic optimization of the rule system.

[0053] It is understandable that differentiable logical operators are used to transform the formalized approval rules into a set of continuous functions that support gradient calculation, thereby constructing a differentiable symbol rule base.

[0054] Step 3.2: Transform logical operations into differentiable mathematical functions to achieve differentiability of the symbolic reasoning process, thereby supporting joint training with neural network modules. Specifically, this includes: First, the product trigonometric norm is used to achieve continuous approximation of logical operations, thus enabling the differentiability of logical operators.

[0055] For conjunction operations, define a differentiable conjunction operator:

[0056] For disjunction operations, a probability sum approximation is used:

[0057] in, Represents the truth value of a logical atom.

[0058] Secondly, based on differentiable logic operators, the degree of satisfaction of each rule is calculated.

[0059] For implied rules When condition P is triggered, the corresponding requirement Q is executed. The degree of satisfaction of this rule is defined as follows:

[0060] in, Indicates the level of satisfaction. These represent the truth values ​​of the premise and the conclusion, respectively. Satisfaction level directly affects the adjustment of rule weights. During training, gradient backpropagation optimizes the rule weights, increasing the weights of important rules and decreasing the weights of less important rules. The gradient calculated from the satisfaction level is backpropagated to the neural network, adjusting its parameters to make the symbolic representation of the neural network's output more consistent with the requirements of the symbolic rules.

[0061] Finally, gradient propagation design is implemented. Differentiable logic operators ensure that gradients during symbolic reasoning can be backpropagated to the neural network module. In calculating rule satisfaction, a smooth gradient approximation is used to avoid the gradient vanishing problem in traditional symbolic reasoning. Through backpropagation via the differentiable reasoning mechanism, the system can transform the logical constraints of symbolic rules into gradient signals, directly guiding the neural network to optimize its internal representation and predictions. This ensures that the neural network's output not only fits the data but also conforms to domain logic rules, allowing the model to learn to "think according to rules," achieving a deep integration and co-evolution of perceptual ability and logical reasoning.

[0062] Step 4: Establish a two-layer optimization framework to collaboratively optimize the weights of the pre-trained neural network module in Step 2 and the rules in the differentiable symbol rule base in Step 3.

[0063] This step aims to achieve deep integration and collaborative optimization of the neural network module (i.e., the domain-adapted pre-trained large language model) and the symbolic reasoning module by establishing an effective joint training mechanism. Specifically, it includes three levels: a two-layer optimization framework, an alternating training strategy, and nondeterministic management. Step 4 specifically includes: Step 4.1: First, establish a hierarchical optimization objective system to ensure that the entire system maintains logical rigor while pursuing task performance. Define the upper-level optimization objective as minimizing the overall loss of the approval task:

[0064] in, This indicates the overall loss of the approval task. This indicates that the maintenance ticket data has been entered. To ensure the accuracy of the approval results, Let x be a mapping function in a neural symbolic system, representing the complete mapping from the input inspection ticket text x to the final approval result. The loss function for the maintenance ticket approval decision task. These are the parameters of the neural network.

[0065] It should be noted that the approval decision-making task is a binary classification task, and the loss function for the maintenance ticket approval decision-making task can be a binary cross-entropy loss function, which is existing technology and will not be elaborated here. The text x in the inspection ticket represents the probability of passing.

[0066] The lower-level optimization objective focuses on the logical consistency of symbol rules:

[0067] in, The logical consistency loss function representing the symbol rules. For a set of rules, For rules The weight, This represents the probability distribution of the output of a neural symbolic system, consisting of a neural network module and a symbolic reasoning module, with respect to the symbolic variable z, given an input text x. For rules Logical constraints, As a measure of distribution distance, This represents the mathematical expectation. The logical consistency loss function of the symbolic rules is the core constraint loss used in the joint training phase of the large language model and the symbolic system, directly forcing the final symbolic output distribution of the entire system to conform to the formal rules.

[0068] Specifically, a differentiable satisfaction function is used to calculate the distribution distance metric. Logical constraints such as "if A then B" are transformed into a continuous function. This function calculates the degree to which the constraint is satisfied according to the symbolic distribution rule, ranging from 0 to 1. The complement of the satisfaction degree is then used as the distribution distance metric. That is, d = 1 - the degree of rule satisfaction. Thus, the distance is 0 when the symbol distribution completely conforms to the logical rule and 1 when it completely violates it, and the whole process is differentiable and optimizable.

[0069] Step 4.2: Optimize the neural network and symbolic system using an alternating parameter update mechanism to achieve stable optimization at time steps. The update process is as follows:

[0070]

[0071] in, This represents the parameters of the neural network at time step t+1. This represents the parameters of the neural network at time step t+1. This represents the symbolic rule weight at time step t. This represents the symbolic rule weight at time step t+1. For neural network parameters, For symbolic rule weights, and This represents the corresponding learning rate.

[0072] Furthermore, to ensure training convergence, a progressive constraint strategy is introduced:

[0073] in, This represents the total loss function for neural networks and symbolic systems. The weighting coefficients increase over time, with the initial stage focusing on task learning and gradually increasing logical constraints.

[0074] Specifically, a linear growth strategy linked to the number of training steps or cycles is adopted: in the early stage of training, it is set to 0 or a very small value to focus on task learning, and gradually increases linearly as the training progresses, reaching a maximum value such as 1.0 in the middle and late stages of training, thereby progressively strengthening the logical constraints and ensuring that the model first masters the basic task and then gradually and strictly aligns with the rule logic.

[0075] Furthermore, to address the ambiguity and uncertainty inherent in the approval process, a nondeterministic management mechanism is introduced into the system, incorporating an entropy regularization term into the optimization objective:

[0076] in, This represents the final total loss function of the neural network module and the symbolic reasoning module. For information entropy function, The coefficient is dynamically decayed. A larger value is set in the early stage of training to encourage exploration, and then gradually decayed to guide convergence in the later stage. This represents the probability distribution of the output of a neural symbolic system with respect to the symbolic variable z, given the input text x.

[0077] A linear decay strategy linked to the number of training steps is typically used. Initially, a large positive value such as 1.0 is set to encourage the model to explore diverse outputs, and then it is gradually decayed to close to 0 to stabilize convergence. The decay timing can be based on the training cycle or triggered when the model entropy value drops to a threshold.

[0078] This collaborative training mechanism ensures that the large language model can understand complex semantics while incorporating the logical reasoning advantages of the symbol system. Through the system's optimization strategies and nondeterministic management, it achieves effective complementarity and performance improvement.

[0079] Step 5: Based on the trained neural symbol system, construct a multi-agent system with professional review capabilities to achieve multi-dimensional and automated review of maintenance tickets, ultimately outputting approval conclusions and interpretable reports. Step 5 specifically includes: Step 5.1: Based on the collaboratively optimized large language model and differentiable symbol rule base, establish a professional intelligent agent review system, adopting a collaborative architecture of "general control intelligent agent - specialized intelligent agent".

[0080] Each specialized intelligent agent integrates a collaboratively optimized neural network module and a corresponding subset of symbolic rules, forming an independent processing unit with neural perception and symbolic reasoning capabilities. They share the same model foundation, but according to their respective review tasks, including but not limited to basic information checks and work content reviews, they load different subsets of symbolic rules and prompt word templates, thereby achieving specialized division of labor and collaborative review.

[0081] The central control agent sets the workflow and is responsible for the unified distribution of tasks and the orderly invocation of specialized agents. Each specialized agent integrates an optimized large language model and a corresponding subset of symbolic rules to form a closed-loop processing unit with "neural perception-symbolic reasoning" capabilities. Differentiated prompt templates are designed to address the characteristics of different review tasks, generating understandable and executable instruction sequences to effectively support the collaborative work of the three specialized agents: the basic information check agent checks the title's standardization, source validity, hierarchical accuracy, and temporal logic; the work content review agent focuses on reviewing the clarity of the work's cause, the rationality of the repair steps, the accuracy of the impact scope, and the completeness of transitional measures; and the impact business verification agent focuses on verifying the consistency between the impact business description and the business ledger records.

[0082] Step 5.2: After each specialized intelligent agent performs the review task, it outputs the review results with confidence assessment and the complete inference path. Based on this, a corresponding result aggregation and confidence management mechanism is established.

[0083] In terms of result aggregation, the review results are integrated based on a strict logical conjunction relationship. If any specific review fails, the overall review fails, forming a necessary condition verification. At the same time, a progressive decision-making process is adopted, with the results of previous reviews serving as the basis for subsequent judgments, ensuring process dependence and logical coherence.

[0084] In terms of confidence management, the review boundary is identified based on the confidence scores output by each specialized intelligent agent. Low-confidence results are automatically triggered to trigger a manual review process to ensure the reliability of key decisions. In addition, the overall review quality is evaluated based on the confidence distribution, providing a basis for subsequent model optimization and rule iteration, thereby achieving continuous improvement.

[0085] Based on the above decision-making rules, the overall control agent integrates the review results of each specialized agent and outputs a clear conclusion of "pass" or "return for modification".

[0086] Step 5.3: To achieve the traceability and verifiability requirements of power safety audits, generate an approval report with a complete decision-making path for each maintenance ticket.

[0087] The report first establishes an end-to-end traceability chain from semantic understanding to symbolic reasoning, fully recording the initial parsing results of the large language model and the layer-by-layer reasoning process of symbolic rules, ensuring that each review conclusion has a clear logical basis. Based on this, the reporting system presents the review results of each specialized agent, listing the review conclusions while simultaneously displaying the confidence assessment and corresponding judgment criteria, forming a complete chain of evidence available for verification. For issues discovered during the review, based on violations of symbolic rules and combined with standard examples from a typical template library, targeted rectification suggestions are automatically generated.

[0088] The entire report is structured around the power safety audit standards, enabling full-chain traceability and auditability from the original work order to the final approval conclusion, effectively supporting the compliance of the approval process.

[0089] Through the above steps, this method transforms the technical advantages of neural symbol systems into practical approval capabilities, ensuring both the professionalism of the review dimensions and the accuracy and reliability of the final decision through a rigorous logical integration mechanism, thus establishing a complete intelligent solution for the approval of power communication maintenance tickets.

[0090] Table 1 is a comparison table of experimental data applied in an embodiment of this disclosure.

[0091] Table 1

[0092] As shown in Table 1, the method proposed in this invention achieves significant improvements in multiple performance indicators compared to traditional methods: by constructing an automated closed loop of "neural perception-symbolic reasoning," end-to-end intelligent approval of maintenance tickets is realized, reducing the processing time per ticket from approximately 12 minutes to 30 seconds, improving efficiency by more than 95%; relying on the dual verification mechanism of the neural symbol system, human error is effectively compensated for, and the approval accuracy rate is improved by 11 percentage points; based on a formal differentiable rule base, comprehensive coverage of power communication approval rules is achieved, controlling operational risks caused by rule omissions from the source; and the end-to-end traceable approval report generation mechanism makes the approval process transparent and credible, fully meeting the stringent requirements of power safety auditing.

[0093] In this embodiment, a large language model and a differentiable symbol rule base are deeply integrated through end-to-end collaborative training to construct a two-layer optimization framework of "neural perception-symbolic reasoning". Based on this, a collaborative architecture of "central control agent-specialized agent" is adopted to achieve fully automated approval of maintenance tickets from semantic understanding and logical verification to report generation. This application solves the problems of low efficiency, easy omission of rules, and lack of interpretability in existing intelligent methods in traditional manual approval, significantly improving the efficiency, accuracy, and auditability of approval, and is suitable for the intelligent processing of power grid communication maintenance services.

[0094] Embodiment 2 of the present invention provides a power communication maintenance ticket approval system, which operates the power communication maintenance ticket approval method as described in Embodiment 1. The system includes: The acquisition module is used to acquire power communication maintenance ticket data pending approval. The semantic understanding and information extraction module is used to perform semantic understanding and information extraction on the power communication maintenance ticket data to be approved using the neural network module, and output structured semantic information. The logical reasoning module is used to input structured semantic information into the symbolic reasoning module. The symbolic reasoning module performs logical reasoning and compliance verification based on a differentiable symbol rule base and outputs the approval result. The conversion module is used to convert the power communication maintenance ticket approval rules into weighted symbolic logic expressions using differentiable logical operators, and to build a differentiable symbolic rule library. The optimization module is used to construct a two-layer optimization framework, which jointly optimizes the parameters of the neural network module and the weights of the rules in the differentiable symbolic rule base based on the upper-layer approval task loss and the lower-layer symbolic rule logic consistency loss.

[0095] Combination Figure 2 As shown, in some embodiments, the power communication maintenance ticket approval system includes: Data Management and Labeling Subsystem: As the foundational support layer of the system, this subsystem implements all the functions of step 1 in the method embodiment. Through the pipeline operations of the data integration and governance unit and the three-level data labeling unit, this subsystem completes the entire process from multi-source heterogeneous data collection, cleaning, and standardization to the construction of a high-quality training dataset, providing structured supervision signals for the subsequent training of the neural symbol system.

[0096] The neural symbol training and optimization subsystem, serving as the intelligent core layer of the system, implements all the functions of steps 2 to 4 in the method embodiment. This subsystem completes the deep injection of domain knowledge and logical consistency reinforcement of the large language model through a domain-adaptive training engine; it implements the formal representation of approval rules and the construction of a differentiable reasoning mechanism through a differentiable symbol rule manager; and the collaborative optimization controller executes a two-layer optimization framework, employing an alternating gradient descent strategy to jointly optimize the parameters of the large language model and the symbol rule weights, and effectively manages the nondeterminism in the system training process through a dynamically decaying entropy regularization term.

[0097] The intelligent approval subsystem for maintenance tickets, serving as the system's business application layer, implements all the functions of step 5 in the method embodiment. This subsystem constructs a collaborative architecture of "central control intelligent agent - specialized intelligent agents": the central control intelligent agent acts as the intelligent agent collaborative scheduling center, responsible for task distribution and process scheduling; three types of specialized intelligent agents (basic information inspection intelligent agent, work content review intelligent agent, and impact on business verification intelligent agent) integrate optimized large language models and symbolic rule subsets to form a closed-loop processing unit of "neural perception - symbolic reasoning." The confidence fusion decision-maker aggregates the review results of each specialized intelligent agent based on logical conjunction relations and identifies boundary conditions through a confidence management mechanism. The interpretable report generator ultimately constructs an end-to-end traceability chain from semantic understanding to symbolic reasoning, outputting a complete approval report that meets the requirements of power safety auditing. The data management and annotation subsystem includes a data integration and governance unit and a three-level data annotation unit.

[0098] The data integration and governance unit is responsible for collecting multi-source data from the power communication operation and maintenance system, including historical maintenance ticket database, standard template database, business resource ledger, and regulations and rules knowledge base. It performs a complete data cleaning and standardization process, including missing value filling, outlier correction, naming mapping dictionary construction, and extraction of key information from regulations and rules, to form a unified and standardized data source.

[0099] The three-level data annotation unit includes result layer annotation subunits, rule layer annotation subunits, and semantic layer annotation subunits, realizing a systematic three-level annotation: the result layer annotates the final approval conclusion of the maintenance ticket, the rule layer annotates the state of satisfaction of various formal symbol rules, and the semantic layer realizes "neural symbol grounding" - extracting key semantic entities from the text and establishing a mapping relationship with symbolic concepts, and finally transforming the original maintenance ticket into a structured training sample to complete the construction of the training set.

[0100] The neural symbol training and optimization subsystem includes a domain-adaptive training engine, a differentiable symbol rule manager, and a collaborative optimization controller. The domain-adaptive training engine first injects domain knowledge deeply through a masked language modeling task to establish the domain knowledge foundation of the model; then, by introducing a symbolic rule consistency loss function and a reinforcement learning reward mechanism based on the integrity of the inference path, it achieves logical consistency reinforcement, ensuring that the model output remains highly consistent with the symbolic system.

[0101] The differentiable symbolic rule manager uses first-order logic to formalize approval rules into "IF-THEN" implication expressions, and uses differentiable logic operators to achieve continuous approximation of logical operations, constructing a differentiable inference mechanism that supports gradient calculation, ensuring that gradients in the symbolic inference process can be backpropagated to the neural network module.

[0102] The collaborative optimization controller establishes a two-layer optimization objective system that includes task loss and logical consistency loss. It adopts alternating gradient descent to jointly optimize the parameters of the large language model and the symbolic rule weights. Furthermore, it manages the nondeterminism in the training process by introducing a dynamically decaying entropy regularization term, thereby achieving effective complementarity and performance improvement between the neural network and the symbolic system.

[0103] The intelligent approval subsystem for maintenance tickets includes an intelligent agent collaborative scheduling center, a confidence fusion decision-maker, and an interpretable report generator. Among these, The Intelligent Agent Collaborative Scheduling Center, acting as the central control intelligent agent, is responsible for setting up approval workflows and sequentially scheduling three types of specialized intelligent agents: the Basic Information Checking Intelligent Agent checks the standardization of titles, the validity of sources, the accuracy of hierarchical classification, and the rationality of time logic; the Work Content Review Intelligent Agent focuses on reviewing the clarity of work reasons, the rationality of maintenance steps, the accuracy of the scope of impact, and the completeness of transition measures; and the Impact Business Verification Intelligent Agent focuses on verifying the consistency between the impact business description and the business ledger records.

[0104] The confidence fusion decision-maker receives the confidence-based review results and complete reasoning paths from each specialized agent, and integrates the results based on a strict logical conjunction: if any specialized review fails, the entire process fails; at the same time, a progressive decision-making process is adopted to ensure logical coherence. Through a confidence management mechanism, low-confidence review results are automatically identified and a manual review process is triggered to ensure the reliability of key decisions.

[0105] The interpretable report generator establishes an end-to-end traceability chain from semantic understanding to symbolic reasoning by fully recording the initial parsing results of the large language model and the layer-by-layer reasoning process of symbolic rules. The system presents the review conclusions, confidence assessments, and judgment criteria of each specialized intelligent agent, forming a complete chain of evidence. For issues discovered during the review, it automatically generates targeted rectification suggestions based on violations of symbolic rules and a typical communication maintenance template library, and finally outputs an approval report that complies with power safety audit standards.

[0106] Regarding the system in the above embodiments, the specific manner in which each unit performs operations has been described in detail in the embodiments related to the method, and will not be elaborated here.

[0107] Embodiment 3 of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the power communication maintenance ticket approval method described in Embodiment 1.

[0108] Embodiment 4 of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the power communication maintenance ticket approval method according to Embodiment 1.

[0109] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0110] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0111] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0112] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0113] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for approving power communication maintenance tickets, characterized in that, include: Obtain data on pending power and communication maintenance tickets; The neural network module is used to perform semantic understanding and information extraction on the power communication maintenance ticket data to be approved, and output structured semantic information. Structured semantic information is input into the symbolic reasoning module, which performs logical reasoning and compliance verification based on a differentiable symbol rule base, and outputs the approval result. Among them, differentiable logical operators are used to transform the power communication maintenance ticket approval rules into weighted symbolic logical expressions, and a differentiable symbolic rule library is constructed. A two-layer optimization framework is constructed, which jointly optimizes the parameters of the neural network module and the weights of the rules in the differentiable symbolic rule base based on the upper-layer approval task loss and the lower-layer symbolic rule logic consistency loss.

2. The method for approving power communication maintenance tickets according to claim 1, characterized in that: The neural network module was pre-trained using professional texts in the field of communications through a masked language modeling task; Acquire multi-source data from the power communication operation and maintenance system, and annotate the multi-source datasets to construct a training set; The neural network module is fine-tuned using the training set, and the fine-tuned neural network module is used as the initial policy for reinforcement learning. Based on the initial policy, the neural network module is optimized through a compound reward function.

3. The method for approving power communication maintenance tickets according to claim 2, characterized in that: The composite reward function includes a first reward function, a second reward function, and a third reward function. The first reward function is used to evaluate the performance of the neural network module in the approval task, the second reward function is used to evaluate the rationality of the logical reasoning path of the neural network module, and the third reward function is used to encourage the neural network module to explore diverse reasoning paths.

4. The method for approving power communication maintenance tickets according to claim 2, characterized in that: The loss function in the pre-training process includes cross-entropy loss and symbolic rule consistency loss. The symbolic rule consistency loss function is introduced to enhance the logical output. The symbolic rule consistency loss is calculated based on whether the structured semantic information output by the neural network module is consistent with the logical reasoning result of the symbolic reasoning module.

5. The method for approving power communication maintenance tickets according to claim 2, characterized in that: Labeling multi-source datasets includes: Based on typical communication maintenance ticket templates and safety procedures, formalized symbol rules are defined. For each historical maintenance ticket, mark the approval result; mark the satisfaction status of each historical maintenance ticket for each symbol rule; mark the key semantic entities extracted from the historical maintenance ticket text, and establish the mapping relationship between key semantic entities and symbol rules.

6. The method for approving power communication maintenance tickets according to claim 1, characterized in that: Calculating the lower-level symbolic rule logical consistency loss includes: Calculate the probability distribution of the output with respect to the symbolic variable given the input text; Calculate the degree to which the rules corresponding to the symbolic variables are satisfied based on the probability distribution; The complement of the satisfaction level is used as a measure of the distance between the probability distribution of the symbolic variable and the distribution of the corresponding rule; The expected value of the sum is calculated by weighting and summing the distribution distance metric and the corresponding rule weights, and then the logical consistency loss of the lower-level symbolic rule is obtained.

7. The method for approving power communication maintenance tickets according to claim 1, characterized in that: In the two-layer optimization framework, a dynamically decaying entropy regularization term is added to the objective function, which consists of the upper-layer approval task loss and the lower-layer symbolic rule logic consistency loss. The information entropy of the output probability distribution with respect to symbolic variables is calculated under the given input text. The expected value of the information entropy is calculated and multiplied by the dynamic decay coefficient to obtain the entropy regularization term. The final total loss function is obtained by subtracting the entropy regularization term from the objective function.

8. The method for approving power communication maintenance tickets according to claim 1, characterized in that: The method further includes: A multi-agent collaborative structure is established to output approval results through multi-agent collaboration. The multi-agent collaborative structure includes a central control agent and specialized agents. The central control agent distributes approval tasks to the corresponding specialized agents. After each specialized agent executes the approval task, it outputs the review results including confidence assessment and the inference path and sends them to the central control agent. The central control agent outputs the final approval result based on the approval results of each agent. Intelligent agents are constructed based on optimized neural network modules and symbolic reasoning modules. Each specialized intelligent agent corresponds to a differentiable symbol rule base for different tasks. According to different review tasks, the corresponding differentiable symbol rule base and prompt word template are loaded so that each specialized intelligent agent can perform different approval tasks.

9. The method for approving power communication maintenance tickets according to claim 7, characterized in that: Specialized intelligent agents include: The basic information checking agent is used to check the basic information of the maintenance ticket, including the title, source of the maintenance ticket, maintenance level, and maintenance start and completion time. The job content review agent is used to review the completeness, accuracy, and compliance of job content. The business impact verification agent is used to verify the consistency between the impact on business and the business ledger.

10. A smart approval system for power communication maintenance tickets, used to implement the smart approval method for power communication maintenance tickets as described in any one of claims 1-9, characterized in that, The system includes: The acquisition module is used to acquire power communication maintenance ticket data pending approval. The semantic understanding and information extraction module is used to perform semantic understanding and information extraction on the power communication maintenance ticket data to be approved using the neural network module, and output structured semantic information. The logical reasoning module is used to input structured semantic information into the symbolic reasoning module. The symbolic reasoning module performs logical reasoning and compliance verification based on a differentiable symbol rule base and outputs the approval result. The conversion module is used to convert the power communication maintenance ticket approval rules into weighted symbolic logic expressions using differentiable logical operators, and to build a differentiable symbolic rule library. The optimization module is used to construct a two-layer optimization framework, which jointly optimizes the parameters of the neural network module and the weights of the rules in the differentiable symbolic rule base based on the upper-layer approval task loss and the lower-layer symbolic rule logic consistency loss.

11. An electronic device, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the power communication maintenance ticket approval method according to any one of claims 1-9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the power communication maintenance ticket approval method according to any one of claims 1-9.

Images