A method for generating prompt words for large-scale models in power communication network operation and maintenance.

Abstract

This application relates to the fields of computer data processing and artificial intelligence technology, specifically disclosing a method for generating large-scale model prompt words for the operation and maintenance of power communication networks. First, it standardizes, cleans, and aligns the original multi-source heterogeneous sensor data according to their time series, mapping it onto a network topology map to anchor the fault source. Then, it uses an anisotropic influence propagation model incorporating business protection logic to perform topology walks, extracting a context-aware subgraph containing fault propagation paths and primary / backup relationships. Next, the structural features of this subgraph are transformed into a serialized text description, combined with alarm facts and thought chain guidance instructions, to assemble a structured prompt word input large-scale model with rigorous logical constraints. This method solves the problem of attribution illusion in large-scale models caused by the lack of spatial constraints on prompt words, and the problem of missed detection of potential risk nodes due to neglecting business protection logic, achieving accurate localization and forward-looking prediction of complex faults in power communication networks.

Description

Technical Field

[0001] This application relates to the fields of computer data processing and artificial intelligence technology, and more specifically, to a method for generating large model prompt words for the operation and maintenance of power communication networks. Background Technology

[0002] As the neural network for the safe and stable operation of the power grid, the power communication network carries critical production operations such as relay protection and dispatch automation. With the construction of new power systems, the communication network architecture is becoming increasingly complex, and the number of devices and alarm data is growing exponentially. Traditional operation and maintenance models that rely on manual experience can no longer meet the needs of real-time fault diagnosis. In recent years, leveraging the powerful semantic understanding and reasoning capabilities of large language models to assist operation and maintenance decision-making has become an industry trend. This aims to achieve automated fault root cause localization and handling suggestions by constructing high-quality prompt word-guided models.

[0003] However, existing prompt word generation technologies for operation and maintenance scenarios typically only perform simple keyword searches or flattened text concatenation on the original alarm logs. This approach reduces the multi-dimensional spatial structure of power communication network data to a linear text sequence, severing the original physical topology connections and service associations between devices. This leads to severe topology fragmentation problems during large-scale model inference. Models often rely solely on text similarity rather than actual network connections for attribution, easily resulting in illusory diagnoses that defy physical common sense. More critically, the few existing attempts that combine graph algorithms often simply assume that the impact of faults spreads isotropically throughout the network, meaning that risks are uniformly transmitted from the fault point to all physical neighbors. This generic model severely ignores the highly logically coupled primary / backup protection mechanisms (such as 1+1 protection or ring network protection) prevalent in power communication networks. In actual operation and maintenance, fault risks will preferentially propagate along the service protection logic to backup routing nodes, rather than spreading uniformly to all physical interfaces. Due to the lack of mathematical representation and cue word constraints for this anisotropic propagation characteristic, existing methods struggle to identify potential cascading risks on backup links, resulting in a lack of foresight and accuracy in the generated operation and maintenance decision reports, which cannot effectively guide complex troubleshooting work.

[0004] Therefore, an optimized scheme for generating large-scale prompt words for the operation and maintenance of power communication networks is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method for generating large-scale model prompts for power communication network operation and maintenance, comprising: S1: Standardize and clean the raw sensor data streams collected from the power communication network to generate standardized event objects; S2: Parse the device identifier in the standardized event object, and perform anchoring retrieval and tagging in the preset network topology map to obtain the set of faulty nodes; S3: Centered on the set of faulty nodes, perform topology walk and fault-related subgraph extraction in the whole network topology graph to obtain a context-aware subgraph containing the fault propagation path; S4: Perform text serialization and causal chain description on the business flow direction and link attributes in the context-aware subgraph to obtain serialized topology description text; S5: The serialized topology description text and standardized event objects are assembled and optimized using thought chain-guided prompts to obtain structured prompts containing topological logic constraints. The structured prompts are then input into the large language model for reasoning interaction to output an operation and maintenance decision report.

[0006] Compared with existing technologies, this application proposes a method for generating large-scale prompt words for power communication network operation and maintenance. It cleans and standardizes the collected raw sensor data of the power communication network, parses equipment identifiers, and anchors fault sources in the network topology map. Centering on this, an anisotropic propagation algorithm incorporating business protection logic weights is used for topology walks to extract context-aware subgraphs containing physical connections and logical coupling relationships, which are then translated into serialized descriptive text with causal temporal order. Subsequently, this topology description and alarm facts are filled into a pre-set template and assembled into a structured prompt word input large language model with thought chain guidance. This method, by explicitly injecting topological structure and business logic constraints during the prompt word generation stage, not only bridges the semantic gap between flat text and graph-structured networks but also ensures that the large model can follow the protection switching mechanism of power services during inference, thereby accurately locating the root cause of the fault and identifying potential cascading risks, achieving high availability and illusion-free generation of operation and maintenance decisions. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1 This is a flowchart of a method for generating large model prompt words for power communication network operation and maintenance according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow in the large model prompt word generation method for power communication network operation and maintenance according to an embodiment of this application; Figure 3This is a flowchart illustrating the method for generating large model prompts for power communication network operation and maintenance according to an embodiment of this application. The method involves performing topology walkthrough and fault-related subgraph extraction in the entire network topology map, centered on the fault node set, to obtain a context-aware subgraph containing fault propagation paths. Figure 4 The flowchart illustrates the iterative calculation of node influence scores for each node in the candidate neighborhood node set to obtain a weighted node list, based on the large model prompt word generation method for power communication network operation and maintenance according to the embodiments of this application. Figure 5 This is a flowchart illustrating the process of generating large model prompt words for power communication network operation and maintenance according to an embodiment of this application, which involves text serialization and causal chain description of service flow direction and link attributes in a context-aware subgraph to obtain serialized topology description text. Figure 6 According to the embodiments of this application, the method for generating large model prompt words for the operation and maintenance of power communication networks involves assembling and optimizing thought-chain-guided prompt words from serialized topology description text and standardized event objects to obtain structured prompt words containing topological logical constraints. The structured prompt words are input into a large language model for reasoning interaction to output a flowchart of the operation and maintenance decision report. Detailed Implementation

[0009] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0010] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0011] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0012] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] Existing large-scale model operation and maintenance technologies typically only flatten and concatenate the original alarm logs, severing the original physical topology connections and business relationships between devices. This leads to severe "topology fragmentation" and logical illusion problems during model inference due to the lack of spatial structural constraints. Furthermore, general models simply assume isotropic diffusion of fault impacts, ignoring the crucial business protection logic in power communication networks and failing to accurately identify potential risks on backup links. Therefore, this application proposes a method for generating large-scale model prompts for power communication network operation and maintenance. Specifically, it first anchors the cleaned standardized event objects in the network topology map, and then performs anisotropic topology walks based on business protection logic weights around these objects to extract context-aware subgraphs that accurately cover fault propagation paths, effectively eliminating irrelevant noise. Subsequently, it uses text serialization technology to translate the business flow and link attributes in the subgraphs into causal chain description text that the large model can understand, and fills it together with the alarm facts into a pre-set thought chain guidance template. Finally, it generates structured prompt words containing explicit topology logic constraints, forcing the large language model to follow the protection switching mechanism and physical connection rules of power business during inference, thereby achieving phantom-free accurate fault location and decision support in complex concurrent alarm scenarios.

[0014] Figure 1 This is a flowchart of a method for generating large model prompts for the operation and maintenance of power communication networks according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in the large-scale model prompt word generation method for power communication network operation and maintenance according to an embodiment of this application. Figure 1 and Figure 2As shown, the method for generating large model prompt words for power communication network operation and maintenance according to an embodiment of this application includes: S1, standardizing and cleaning the collected raw sensor data stream of the power communication network to generate standardized event objects; S2, parsing the device identifiers in the standardized event objects and performing anchoring retrieval and tagging in a preset full network topology map to obtain a fault node set; S3, using the fault node set as the center, performing topology walk and fault-related subgraph extraction in the full network topology map to obtain a context-aware subgraph containing fault propagation paths; S4, performing text serialization and causal chain description on the service flow direction and link attributes in the context-aware subgraph to obtain serialized topology description text; S5, performing thought chain-guided prompt word assembly and optimization on the serialized topology description text and standardized event objects to obtain structured prompt words containing topological logic constraints, and inputting the structured prompt words into a large language model for reasoning interaction to output an operation and maintenance decision report.

[0015] Specifically, in step S1, the collected raw sensor data stream is standardized and cleaned to obtain a normalized time-series tensor, and a pre-trained text embedding model is used to vectorize the operation and maintenance log text to obtain an operation and maintenance log vector representation. It should be noted that, given the massive amounts of data involved in power communication network operation and maintenance scenarios, such as equipment optical power and bit error rate, which are characterized by numerous sources, dispersed distribution, and high-frequency noise, while unstructured texts such as operation and maintenance logs face challenges of semantic obscurity and inconsistent formats, the complexity of this multi-source heterogeneous data severely hinders subsequent correlation analysis and large-scale model inference accuracy. Based on this, the technical solution of this application first standardizes and cleans the collected raw sensor data stream to obtain a normalized time-series tensor, and then uses a pre-trained text embedding model to vectorize the operation and maintenance log text to obtain an operation and maintenance log vector representation. This eliminates the dimensional differences and transient jitter between data from different devices, and maps discrete text information into a high-dimensional semantic space vector that can be efficiently retrieved by computers, achieving feature alignment between physical measurement data and textual logical data. The above processing can effectively improve the accuracy of multi-source data governance, ensuring that the prompt words input into the large model have high data quality and semantic relevance, thus laying a solid data foundation for realizing full-link automated operation and maintenance.

[0016] More specifically, in a specific example of this application, step S1 includes: performing jitter suppression and denoising processing on the acquired raw sensor data stream based on a sliding window to obtain a denoised data stream; extracting key operation and maintenance indicators from the denoised data stream based on a regularization engine to obtain a structured field set; and performing linear interpolation or nearest neighbor matching on each data point in the structured field set based on a global standard time base to obtain a standardized event object.

[0017] Accordingly, the acquired raw sensor data stream is subjected to jitter suppression and denoising processing based on a sliding window to obtain a denoised data stream. It should be noted that due to the complex physical operating environment of power communication networks, and the influence of strong electromagnetic interference or thermal noise from the electrical components of the equipment itself, the acquired raw sensor data is often accompanied by a large amount of non-fault-related transient jitter or high-frequency noise. If this dirty data, which does not reflect the actual changes in network status, is directly input into subsequent stages, it is highly likely to trigger false alarms or interfere with the large model's extraction of fault features. Based on this, the technical solution of this application first performs jitter suppression and denoising processing based on a sliding window on the acquired raw sensor data stream to obtain a denoised data stream, thereby filtering out random interference signals and retaining effective low-frequency components that can truly reflect the operating trend of the equipment. Through the above processing, the signal-to-noise ratio and quality of the underlying operation and maintenance data can be effectively improved, preventing misjudgments in operation and maintenance caused by minor data fluctuations, and providing a clean and reliable data foundation for generating high-confidence standardized event objects and subsequent accurate inference of large models.

[0018] More specifically, in a particular example of this application, the raw sensor data stream includes: the optical power value of the optical transmission device, the B1 / B2 / B3 bit error counts of SDH / OTN frames, the device CPU / memory utilization, the CRC error statistics of port transceiver packets, and SNMPTrap alarm messages. For continuously changing indicators such as the optical power value of the optical transmission device or the device CPU utilization, a fixed-length time window is first set as a sliding kernel based on the sampling frequency. This window covers several consecutive historical sampling time points. Subsequently, as the sampling time step advances, the window moves forward point by point on the time axis of the raw sensor data stream. At each step, all sampling data points within the current window range are captured. Based on this, an arithmetic mean filtering algorithm is used to smooth the sampled values ​​within the window. That is, the algebraic average of all data points within the window is used to replace the original observation value at the current moment, and the calculated smoothed values ​​are rearranged in time sequence to generate a continuous and smooth denoised data stream.

[0019] Accordingly, key operational metrics are extracted from the denoised data stream using a regular expression engine to obtain a structured field set. It should be noted that although the denoised data stream eliminates random jitter noise in the time domain, its data form is essentially still a semi-structured or unstructured text sequence. Furthermore, the proprietary log protocols and message formats of different device manufacturers vary significantly, causing key operational metrics to be deeply embedded within a large number of redundant descriptive characters and formatting symbols. Downstream algorithms struggle to directly perform numerical calculations or logical associations on such heterogeneous text. Therefore, the technical solution of this application further extracts key operational metrics from the denoised data stream using a regular expression engine to obtain a structured field set. This allows for deep semantic parsing and format normalization of multi-source heterogeneous data through standardized syntax rules, forcibly transforming non-standardized log text into structured variables that can be directly indexed, sorted, and computed by computers. Through this processing, the protocol differences between underlying heterogeneous devices can be effectively masked, completely eliminating the ambiguity and confusion of the data source, ensuring that subsequent steps such as topology anchoring and large-scale model inference are based on semantically clear, formatted, and machine-readable data.

[0020] More specifically, in a concrete example of this application, a pre-built regular expression template library adapted to the log formats of heterogeneous vendors such as optical transmission equipment and routers in power communication networks is first loaded. Then, the efficient pattern matching capability of the regular expression engine is utilized to perform a full scan of the denoised data stream to identify feature strings conforming to predefined rules. For text-based data such as SNMPTrap alarm messages, specific regular expression operators are used to accurately match and capture key topology location and status information, including equipment rack number, board slot index, port logical ID, and alarm severity level. For performance monitoring data, floating-point or integer key values ​​such as optical power values ​​of optical transmission equipment, B1 / B2 / B3 error counts of SDH frames, CRC error statistics of port transceiver packets, and equipment CPU and memory utilization are extracted. Finally, these discrete information fragments extracted from the original messages are mapped and encapsulated according to a unified metadata specification using key-value pairs, thereby assembling them into a structured field set containing standard attribute fields.

[0021] Accordingly, based on a global standard time base, linear interpolation or nearest neighbor matching is performed on each data point in the structured field set to obtain standardized event objects. More specifically, in a specific example of this application, a globally unified standard time axis is first established based on a high-precision network time protocol, and a fixed time slice granularity is set as the reference grid for data alignment according to the minimum time resolution required by the business. Subsequently, differentiated alignment strategies are implemented for different types of data characteristics in the structured field set. For analog quantity fields that change continuously with time, such as the optical power value of optical transmission equipment and the CPU memory utilization rate of equipment, a linear interpolation algorithm is used to calculate the estimated value at the standard time point based on the linear slope relationship between the two original sampling points before and after the standard time grid, thereby filling the data gaps caused by inconsistent sampling frequencies. For discrete enumeration type fields that undergo state changes, such as port packet CRC error statistics and SNMPTrap alarm codes, a nearest neighbor matching algorithm is used to directly select the original sampling value with the smallest time difference from the current standard time point as the state value at that moment, or a zero-order hold method is used to maintain the state of the previous moment until the next state change occurs. Finally, all fields that have undergone time-series alignment are bound and encapsulated with a unified standard timestamp to generate a standardized event object that has time-series consistency and contains complete operational metrics.

[0022] It should be noted that, due to the fact that heterogeneous network elements such as optical transmission equipment and routers in power communication networks often rely on local clocks to operate independently, and different types of monitoring indicators have different sampling frequencies, the structured fields extracted in the preceding steps exhibit severe asynchronous and discrete distribution characteristics in the time dimension. This misalignment and non-alignment of multi-source data in time sequence seriously hinders the construction of a unified time series tensor from the dispersed operation and maintenance data, thus failing to meet the requirement of strict time sequence alignment for accurate causal correlation analysis of large models. Based on this, the technical solution of this application further uses a global standard time reference to perform linear interpolation or nearest neighbor matching on each data point in the structured field set to obtain standardized event objects. This forces all heterogeneous data from different sources and with different frequencies to be mapped onto a unified discrete time grid and eliminates the impact of clock drift between devices. Through the above processing, the time sequence synchronization and alignment of multi-source operation and maintenance data can be effectively achieved, ensuring that the generated event objects have strict identity in time logic, thereby providing a reliable time reference for the subsequent construction of a topology subgraph that accurately reflects the time sequence relationship of fault propagation.

[0023] Specifically, in step S2, the device identifier in the standardized event object is parsed, and anchoring retrieval and tagging are performed in a preset full-network topology map to obtain the fault node set. It should be noted that although the standardized event objects generated in the previous steps have unified temporal and numerical attributes, they are essentially isolated logical data points, lacking location information and topological context in the physical network space. This separation of data and model makes it impossible for subsequent algorithms to determine the geometric starting point of fault propagation. If reasoning is directly based on this, the large model will fall into the illusion trap of matching based solely on text similarity due to the lack of spatial constraints. Based on this, the technical solution of this application further performs string parsing and format normalization on the data payload of the standardized event object to extract the device unique identifier. The device unique identifier is used as the search key to perform indexing and matching in the preset full-network topology map to obtain the anchor network element node. Dynamic fault semantic tags containing timestamps and alarm levels are injected into the anchor network element node to obtain the fault node set. This constructs a mapping bridge for operation and maintenance data from the logical domain to the physical topology domain, realizing the accurate location and state binding of fault entities in the full-network topology. Through the above processing, discrete alarm events can be effectively transformed into anchor nodes with graph attributes, providing accurate starting coordinates and boundary constraints for subsequent graph theory-based fault propagation walks and subgraph extraction.

[0024] More specifically, in a specific example of this application, step S2 includes: performing string parsing and format normalization on the data payload of the standardized event object to extract the device unique identifier; using the device unique identifier as a search key to perform index lookup and matching in a preset full network topology map to obtain the anchor network element node; and injecting dynamic fault semantic tags containing timestamps and alarm levels into the anchor network element node to obtain a fault node set.

[0025] More specifically, the process first reads the data payload of the standardized event object, locates the field storing network element identity information based on a predefined communication protocol field mapping table, extracts the unique identifier string representing the physical device using a string parsing algorithm, and performs normalization processing such as removing meaningless leading zeros or standardizing case to generate a globally unique device identifier. Subsequently, this unique device identifier is used as a high-dimensional index key, and a hash lookup or B+ tree search is performed in a pre-built graph structure database containing all network connections. This process follows the formula: in, This indicates the anchored network element node obtained from the retrieval. This represents the set of all network element nodes in the network topology diagram. This is an indicator function that returns 1 when the two input parameters are exactly equal. The attribute ID stored for the node. This is the unique identifier for the input device. After successfully locking the topology node, an encapsulation container is created, using the original graph node as the core payload, and injecting it with dynamic fault semantic tags containing the current event timestamp and alarm severity level. Finally, these node objects carrying spatiotemporal multidimensional context information are added to the collection container to form a fault node set.

[0026] Specifically, in step S3, with the fault node set as the center, a topology walk and fault-related subgraph extraction are performed in the entire network topology map to obtain a context-aware subgraph containing the fault propagation path. It should be noted that, given the highly structured nature of power communication networks, which widely employ service protection mechanisms such as synchronous digital hierarchy ring protection or 1+1 linear protection, strong logical coupling exists between nodes. Traditional isotropic diffusion models assume that fault risks are uniformly transmitted to all physical neighbors, often ignoring the crucial fact that risks are asymmetrically propagated to backup routing nodes during primary / backup switching. This leads to critical early warning information on backup links being easily misjudged as noise and removed during conventional topology pruning. Based on this, the technical solution of this application further uses the fault node set as the center, performs a topology walk and fault-related subgraph extraction in the entire network topology map to obtain a context-aware subgraph containing the fault propagation path, and introduces an anisotropic influence diffusion mechanism based on the logical coupling of service protection during the walk process. Specifically, a logical coupling matrix is ​​constructed to quantify the non-physically strong coupling between primary and backup nodes, and a Softmax weighted normalization strategy is used to calculate the anisotropic transition probability. This guides the targeted propagation and iterative calculation of fault impact along high-risk business protection logic paths. Through this processing, the distortion problem of fault correlation analysis caused by the primary / backup switchover mechanism can be effectively solved, accurately capturing potentially risky nodes that are physically distant but logically close, thereby improving the coverage and topological integrity of the generated context-aware subgraph for key logical nodes.

[0027] Figure 3 This is a flowchart illustrating a method for generating large-scale prompt words for power communication network operation and maintenance, based on an embodiment of this application. The method involves performing topology walkthroughs and extracting fault-related subgraphs within the entire network topology map, centered on a set of fault nodes, to obtain a context-aware subgraph containing fault propagation paths. For example... Figure 3 As shown, step S3 includes: S31, taking each node in the fault node set as the starting point of the walk, performing a bidirectional breadth-first search topology walk in the entire network topology graph to obtain a candidate neighbor node set; S32, performing iterative calculation of the node influence score for each node in the candidate neighbor node set to obtain a weighted node list; S33, pruning and reconstructing the weighted node list based on a set relevance truncation threshold to obtain a context-aware subgraph.

[0028] In step S31, each node in the fault node set is used as the starting point for the walk, and a bidirectional breadth-first search topology walk is performed in the entire network topology map to obtain a candidate neighbor node set. It should be noted that due to the high connectivity and cascading effect of power communication networks, equipment failures are often not isolated events, but rather trigger chain reactions in the physical topology and business logic of upstream and downstream systems. Focusing only on the alarm node itself makes it difficult to understand the transmission path and scope of the fault. However, performing an indiscriminate traversal scan of all nodes in the entire network faces the dilemma of redundant computing resources and irrelevant noise interference. Based on this, the technical solution of this application further uses each node in the fault node set as the starting point for the walk, performing a bidirectional breadth-first search topology walk in a pre-set entire network topology map to obtain a candidate neighbor node set. This allows the fault source to radiate outwards, simultaneously covering potential root cause nodes upstream and affected business nodes downstream, thus initially delineating the topological boundary of the fault's impact in the entire network topology. Through the above processing, a candidate neighborhood node set containing all potentially relevant network elements can be effectively extracted from the massive network data, ensuring that subsequent causal analysis is based on the complete fault propagation context, while avoiding the introduction of redundant background noise unrelated to the fault.

[0029] More specifically, in a concrete example of this application, the fault node set generated in the preceding steps is first read. Each anchor network element node marked as a fault trigger source is initialized as a seed node for topology traversal and placed at the head of the search queue. Subsequently, relying on the breadth-first search algorithm in graph theory, an iterative traversal program is initiated in the entire network topology database. This program is configured in bidirectional mode, that is, simultaneously tracing upstream aggregation nodes and downstream access nodes along the physical links connected by optical cables. To prevent the unlimited expansion of the search range from causing a computational dimensionality explosion, a maximum hop count threshold (e.g., 3 hops) is set as the termination condition for the traversal depth. In each iteration, all directly adjacent nodes of the nodes in the current queue are traversed, added to the candidate set, and marked with their access status, until the preset hop count depth is reached or all relevant connected components are traversed. Finally, a candidate neighbor node set containing all visited nodes and their connection relationships is output.

[0030] In step S32, the node influence score of each node in the candidate neighboring node set is iteratively calculated to obtain a weighted node list. It should be noted that, given that power communication networks widely employ technologies such as synchronous digital hierarchy ring protection or 1+1 linear protection to ensure the high availability of critical services, there is extremely strong business logic coupling between physically distant nodes. Traditional isotropic diffusion models completely ignore this primary / backup protection relationship, assuming that fault risk is uniformly transmitted to all physical neighbors. This makes it impossible to quantify the asymmetric characteristics of risk directional transmission to backup nodes during primary / backup switching, easily misjudging critical early warning information on backup links as background noise and eliminating it. Based on this, the technical solution of this application further iteratively calculates the node influence score of each node in the candidate neighboring node set to obtain a weighted node list. In the calculation process, an anisotropic influence diffusion mechanism based on business protection logic coupling is introduced. By constructing a logic coupling matrix, the directional coupling coefficient between primary and backup nodes is explicitly quantified, and the anisotropic transition probability tensor is calculated using a Softmax weighted normalization strategy to guide the fault influence to propagate preferentially along high-risk business protection logic paths. Through the above processing, the original model can effectively correct the defect of ignoring strong business logic correlation, deeply integrate the static topology structure with the dynamic business protection logic, accurately capture those silent nodes that currently have no alarms but have extremely high potential risks due to master-slave correlation, thereby improving the ability of the generated node list to predict potential cascading failures and the integrity of the topology.

[0031] Figure 4 This is a flowchart illustrating the iterative calculation of node influence scores for each node in the candidate neighborhood node set, based on the large-scale prompt word generation method for power communication network operation and maintenance according to embodiments of this application, to obtain a weighted node list. For example... Figure 4 As shown, step S32 includes: S321, constructing a logical coupling matrix of the candidate neighborhood node set based on the protection configuration data; S322, estimating the anisotropic transition probability of the candidate neighborhood node set based on the logical coupling matrix to obtain the anisotropic transition tensor; S323, performing a two-factor weighted influence iteration on each node in the candidate neighborhood node set based on the anisotropic transition tensor to obtain a weighted node list.

[0032] In step S321, a logical coupling matrix of the candidate neighboring node set is constructed based on the protection configuration data. It should be noted that, given the widespread primary / backup protection mechanisms in power communication network operation and maintenance scenarios, which result in extremely strong business logic coupling between nodes, traditional isotropic diffusion models, which only calculate based on physical connectivity degrees, cannot quantify the limitations of non-physical strong coupling between primary and backup nodes, making it difficult to accurately assess the potential risks of backup routing nodes. Therefore, the technical solution of this application further constructs a logical coupling matrix of the candidate neighboring node set based on the protection configuration data, thereby transforming the invisible business protection relationships into computable parameters and executing the construction of the primary / backup switching coupling matrix. Through the above processing, a logical coupling matrix that can explicitly characterize the logical coupling strength between nodes can be generated, shortening the distance of logically related paths at the computational level. This constructs a data structure that accurately reflects the inherent protection logic of power services, providing a key basis for subsequent non-uniform risk diffusion calculations.

[0033] More specifically, in a concrete example of this application, the service protection logic information is first extracted from the complex network configuration by parsing the input power communication network protection configuration data. Next, node pairs with a primary / backup protection relationship are identified in the candidate neighbor node set, clearly distinguishing between the primary and backup nodes. Subsequently, based on the importance level of the service in its protection configuration, a directional coupling coefficient value much higher than that of a normal physical connection is assigned to this pair of nodes. The calculation logic for this matrix construction can be expressed as the following mathematical expression: in, For the slave node in the logical coupling matrix To the node The coupling coefficient, To replace the bias factor, the risk of the standby node being activated when the primary node is abnormal is quantified. For nodes The level of the service carried, is the base of the natural logarithm. Furthermore, if the node... Only nodes If the nodes are physical neighbors but have no protection relationship, the coefficient is set to 1.0; otherwise, it is set to 0. This step amplifies the risk transmission capability of high-level services between primary and backup nodes through an exponential function. For example, in scenarios involving UHV transmission line relay protection services, assuming the nodes... It carries extremely high-level scheduling and control services, and the nodes It is the core router of its off-site disaster recovery backup center. Although the two are hundreds of kilometers apart in terms of physical fiber optic topology and have extremely low correlation at the physical connection level, protection configuration data indicates that they have a 1+1 linear protection relationship. Based on the above formula, due to the service level... Extremely high, exponential operation This will generate a large number of directional coupling coefficients, making the logical coupling matrix... arrive The weight far exceeds To its physical neighboring nodes The weight is 1.0. This numerical differentiation forces subsequent risk analysis algorithms to overcome the limitations of physical distance and prioritize backup nodes. This allows for the precise simulation of the risk jump phenomenon that occurs during a business transition.

[0034] In step S322, anisotropic transition probabilities are estimated for the candidate neighborhood node set based on the logical coupling matrix to obtain the anisotropic transition tensor. It should be noted that, given that fault propagation in power communication networks is not uniform but strongly guided by business logic, simply relying on the degree of the physical topology for normalization is insufficient to meet accuracy requirements in order to identify and prioritize propagation along high-risk logical paths during the influence diffusion process. Therefore, the technical solution of this application further estimates the anisotropic transition probabilities for the candidate neighborhood node set based on the logical coupling matrix to obtain the anisotropic transition tensor. This corrects the isotropic normalization method based on degree in the original model while giving the influence diffusion a directionality. Through the above processing, the impact of the fault is biasedly distributed to neighboring nodes with strong logical coupling. This directly simulates the biased characteristics of the fault impact in the power protection switching scenario, so that the impact is focused on the real risk-bearing path. This generates an anisotropic transfer tensor that can guide the non-uniform and directional propagation of the impact, and finally makes the diffusion behavior of the algorithm highly aligned with the physical laws and business logic of the real world.

[0035] More specifically, in a concrete example of this application, based on the logical coupling matrix generated in the previous step, a Softmax weighted normalization strategy is used to perform a nonlinear transformation and probability redistribution on the connection weights between nodes in the candidate neighborhood node set. In this process, instead of simply allocating the risk transfer probability equally based on the number of physical connections, an exponential function is introduced to amplify the coupling strength, and the influence transfer probability between nodes is recalculated. This process can be expressed as the following mathematical expression: in, For the node To the node The anisotropic transition probability, This is the clustering coefficient, used to regulate the degree to which risk concentrates towards protected nodes. The larger the coefficient, the more pronounced the trend of risk concentration towards highly coupled nodes. For nodes The set of physical neighbor nodes, For nodes The set of logically protected nodes, This is the logical coupling coefficient calculated in the previous steps. Through this calculation, the values ​​in the logical coupling matrix are mapped to a probability distribution with a sum of 1, thus completing the construction of the anisotropic transition tensor. Taking the primary and backup communication routing of an ultra-high voltage substation as an example, assume that there is a high-level service protection relationship between the primary node A and the backup node B (high logical coupling coefficient), while node A and the physically adjacent ordinary monitoring node C only have a physical connection (low logical coupling coefficient). When performing anisotropic transition probability estimation, the Softmax strategy greatly amplifies the transition probability from A to B (e.g., calculated to 0.9), while suppressing the transition probability from A to C (e.g., reducing it to 0.1). This means that in subsequent risk iteration calculations, the impact of a fault in node A will mainly flow to node B, which is spatially distant but logically closely related, rather than flooding to the physically adjacent but service-unrelated node C, thus accurately simulating the risk jump brought about by service switching.

[0036] In step S323, based on the anisotropic transition tensor, a two-factor weighted influence iteration is performed on each node in the candidate neighborhood node set to obtain a weighted node list. It should be noted that although the previous steps have established the propagation direction of fault influence through the anisotropic transition tensor, the simple propagation probability does not consider the historical stability and inherent risk attributes of the nodes themselves. If only the current topology is used for deduction, it is difficult to quantify the potential collapse risk of nodes that have frequently experienced failover or have unstable performance in the past when handling sudden business traffic. Therefore, the technical solution of this application further performs a two-factor weighted influence iteration on each node in the candidate neighborhood node set based on the anisotropic transition tensor to obtain a weighted node list. To obtain a final risk assessment that is closer to actual operation and maintenance, the two-factor weighted influence iteration calculation is performed, aiming to apply the improved transition probability to the iteration calculation and incorporate historical operation and maintenance experience data to achieve a comprehensive risk assessment. Through the above processing, information from three different dimensions—static topology, dynamic business logic, and historical stability—can be integrated into a unified risk calculation framework, thereby generating a risk-aware weighted node list. This list can not only accurately assess the quality and reliability of the impact of current alarms, but also proactively identify silent nodes that currently have no alarms but have extremely high potential risks due to historical or logical connections, thus providing strong data support for predictive maintenance.

[0037] More specifically, in a concrete example of this application, all nodes in the candidate neighborhood node set are first initialized. Nodes already anchored as fault sources are given initial high weights, while other nodes are given zero or low background weights. Subsequently, an iterative calculation process is initiated. In each iteration, the traditional average degree is no longer used as a decay factor. Instead, when iteratively updating the node influence score, the anisotropic transition tensor generated in the previous step is used to replace the original degree normalization factor, and a service interruption penalty term is introduced. This penalty term is assigned a value based on the node's historical master / standby switchover frequency. If a node's historical state is unstable, its current risk score is correspondingly increased. In this process, the specific calculation logic of the two-factor weighting follows the following mathematical expression: in, For nodes The final influence score, These are the inherent weighting coefficients. For nodes The initial fault weights, The historical penalty coefficient, For nodes The business interruption penalty term characterizes the instability of its historical state. This formula is derived from the first part... Captured from neighboring nodes External risks, transmitted via business logic paths, are then further addressed through the second part. It integrates the node's own fault status and historical health. The iterative process continues until the score changes of all nodes converge to a preset error range, and finally outputs a list of nodes containing a comprehensive risk score.

[0038] Taking a specific UHV (Ultra-High Voltage) supporting communication network scenario as an example, assume there is a primary routing node A and a backup routing node B, which have a 1+1 protection relationship. At the current moment, node A issues an alarm due to a board failure, while node B, although operating normally and without alarms, has experienced multiple unexplained resets due to a software bug within the past three months, causing its... The value is relatively high. When performing this step, the algorithm first transfers the tensor based on anisotropy. The impact of A's failure is likely to be redirected to B; at the same time, due to the historical stability penalty of node B... If a numerical value exists, the second factor in the formula will further increase the final score of node B. This mechanism ensures that in the final weight list, node B, although currently without problems, is marked as high-risk. This prompts the large model to not only address the failure of node A when generating operational recommendations, but also to focus on checking the carrying capacity of node B or to recommend a preventative restart of node B before failover. This effectively avoids a complete business disruption caused by a faulty backup device being put into operation.

[0039] In step S33, based on a set relevance truncation threshold, the weighted node list is pruned and reconstructed to obtain a context-aware subgraph. It should be noted that while the weighted node list generated in the previous steps through topology walks and influence iteration quantifies the correlation between each network element node and the fault source, it still essentially contains a large number of background noise nodes with extremely low influence scores, existing only as weak physical connections without any actual business logic impact. If these redundant nodes are directly retained, they will not only needlessly consume the valuable context window resources of the large language model but also introduce irrelevant interference information, causing the model to become distracted during inference and even produce incorrect causal attributions. Therefore, the technical solution of this application further prunes and reconstructs the weighted node list based on a set relevance truncation threshold to obtain a context-aware subgraph, thereby performing rigorous data cleaning and topology simplification operations, forcibly removing all non-critical nodes with relevance scores below the safety threshold, and reorganizing the connection relationships of the remaining core high-risk nodes in the original network. Through the above processing, the complex candidate neighborhood can be effectively converged into a highly refined context-aware subgraph that contains only the root cause of the failure and its exact propagation path. This ensures that the information input to the large model has a very high signal-to-noise ratio and topological logical coherence, thus laying a solid spatial data foundation for the subsequent generation of accurate operation and maintenance decision reports.

[0040] More specifically, in a concrete example of this application, firstly, based on the operational tolerance of the power communication network or the statistical distribution characteristics of historical fault data, a numerical correlation truncation threshold is set. This threshold serves as the decision boundary distinguishing effective fault propagation paths from background topology noise. Next, a traversal and filtering procedure is initiated on the weighted node list, reading the final influence score of each network element node in the list one by one and comparing it numerically with the preset truncation threshold. Nodes with scores below the threshold are judged as redundant objects with no substantial causal relationship to the current fault and are removed, while nodes with scores above or equal to the threshold are marked as core retained nodes. Subsequently, based on the original connection relationships stored in the network topology database, the physical optical cable connections and logical service mappings between all core retained nodes are retrieved and restored, reconstructing the edge relationships between these discrete high-scoring nodes. When core retained nodes are not connected in the original topology, the shortest path intermediate node in the original topology is allowed to be introduced as a bridging node to ensure that the generated subgraph maintains connectivity in the topology structure. Finally, these selected nodes and their reconstructed connections are encapsulated into an independent graph data object, namely a context-aware subgraph. This subgraph fully preserves all key network elements that spread from the source of the fault along the path of high-risk business logic, thus achieving a precise spatial characterization of the scope of the fault's impact.

[0041] Specifically, in step S4, the business flow direction and link attributes in the context-aware subgraph are serialized and described using causal chains to obtain serialized topological description text. It should be noted that, given that large language models are essentially sequence-based probabilistic prediction models, they cannot directly perceive or parse graph structure data in non-Euclidean space. If the extracted subgraph is directly converted into flat log text, key edge attributes, business flow directions, and spatial topological relationships between nodes will inevitably be lost, resulting in the model being unable to construct effective thought chains due to a lack of structured constraints. Based on this, the technical solution of this application further serializes and describes the business flow direction and link attributes in the context-aware subgraph using text to obtain serialized topological description text. By introducing domain-specific language, the complex two-dimensional or three-dimensional graph structure is dimensionality-reduced and mapped into a one-dimensional linear text sequence containing explicit causal logic, and the business logic relationships between upstream aggregation nodes and downstream access nodes are clearly marked, thereby constructing a semantic bridge from unstructured graph data to the text understanding capabilities of large models. Through the above processing, the semantic gap between heterogeneous graph data and large model text input can be effectively bridged, providing clear and accurate topological context support for subsequent thought chain reasoning. This forces the large model to follow physical connection and business protection mechanisms during reasoning, reducing the risk of illusion caused by topological fragmentation.

[0042] Figure 5This is a flowchart illustrating the process of generating large-scale prompt words for power communication network operation and maintenance according to an embodiment of this application. It describes the text serialization and causal chain description of service flow direction and link attributes in a context-aware subgraph to obtain serialized topology description text. For example... Figure 5 As shown, step S4 includes: S41, performing business flow analysis and link attribute extraction on the edges and nodes in the context-aware subgraph to obtain an attribute-enhanced graph object; S42, performing DSL-based structured translation on the nodes and edges in the attribute-enhanced graph object to obtain DSL code fragments; S43, performing causal chain logic sorting and text concatenation on the DSL code fragments to obtain serialized topology description text.

[0043] In step S41, the edges and nodes in the context-aware subgraph are analyzed for business flow and extracted for link attributes to obtain attribute-enhanced graph objects. It should be noted that although the context-aware subgraph generated in the preceding steps retains the physical connections related to the fault in its geometric structure, it is essentially a static topology skeleton. It lacks dynamic operational status information and business logic flow characteristics crucial for fault attribution analysis, making it difficult for subsequent algorithms to distinguish between simple physical connections and the logical paths actually carrying services, and also making it difficult to judge the severity of the fault based on real-time link quality indicators. Therefore, the technical solution of this application further analyzes the business flow and extracts link attributes from the edges and nodes in the context-aware subgraph to obtain attribute-enhanced graph objects, thereby achieving deep alignment and semantic fusion between the isolated static topology structure and real-time operation and maintenance monitoring data and business configuration information. Through the above processing, the topology subgraph can be effectively endowed with rich multi-dimensional semantic connotations, ensuring that the prompt words of the large model generated later not only include the spatial connection relationship between network elements, but also accurately express the transmission direction of services and the real-time health status of links, thereby providing complete data support for the construction of accurate causal reasoning chains for the large model.

[0044] More specifically, in a concrete example of this application, the process first involves a full traversal of all edge and node elements in the context-aware subgraph. Based on a pre-configured power communication network service routing database, each physical link is compared and identified to determine its specific service type and logical transmission direction, thus marking the originally undirected physical connections as directed service paths with clearly defined source and destination ends. Next, using the fault occurrence timestamp as a benchmark, key performance indicators for each network element node and port within a specific time window are retrieved from a real-time time-series database. These indicators specifically include the optical power value of the optical transmission equipment, the bit error count of SDH frames, and the resource utilization rate of the equipment. These discrete numerical indicators are then transformed into qualitative state description labels through threshold mapping. Finally, the obtained service flow labels, logical role definitions, and dynamic performance attributes are encapsulated as key-value pairs and attached to the corresponding node and edge objects of the subgraph, thereby generating an attribute-enhanced graph object that includes both spatial structure and temporal state.

[0045] In step S42, the nodes and edges in the attribute-enhanced graph object undergo DSL-based structured translation to obtain DSL code fragments. It should be noted that although the attribute-enhanced graph object fully preserves the topological connections and service status information of the power communication network at the data structure level, it is essentially still a high-dimensional discrete data structure. Large language models are primarily trained and inferred based on serialized text data, making it difficult to directly perform efficient semantic parsing and spatial logic deduction on complex graph data objects. Therefore, the technical solution of this application further performs DSL-based structured translation on the nodes and edges in the attribute-enhanced graph object to obtain DSL code fragments. This maps the nonlinear spatial topology structure into domain-specific language code that the large language model has a strong understanding capability, utilizing the rigorous syntactic structure of the code to carry the logical connections and attribute features between nodes. Through the above processing, the cognitive load of the large model on the graph data can be effectively reduced, enabling it to parse the network topology like reading program code, thereby providing highly readable and logically structured input for subsequent causal inference.

[0046] More specifically, in a concrete example of this application, a pre-built domain description language conversion template adapted to the topology representation of power communication networks is first loaded. This template is constructed based on Mermaid or GraphvizDOT syntax rules. Then, a full traversal of the attribute-enhanced graph objects is performed. For each network element node, its device identifier, device type, and operating status attributes are read, and a unique node declaration statement is generated according to the node definition syntax. For each edge containing the service flow direction, its source node, destination node, and link transmission quality attributes are read, and a connection statement with direction indicators and attribute labels is generated according to the relation definition syntax. Finally, all the generated node declaration statements and connection statements are combined according to the topology hierarchy to generate a complete DSL code snippet capable of textually reconstructing the topology structure of a fault area.

[0047] In step S43, the DSL code fragments are logically sorted according to causal chains and concatenated with text to obtain serialized topology description text. It should be noted that since the DSL code fragments generated in the preceding steps are often discrete and disordered in physical storage, and large language models heavily rely on the sequential logic of text input when establishing causal inference chains, simply concatenating the disordered topology code fragments directly can make it difficult for the model to capture the directional characteristics of fault propagation between network layers, leading to causal inversion or logical breaks in the inference process. Therefore, the technical solution of this application further performs logical sorting according to causal chains and concatenates text on the DSL code fragments to obtain serialized topology description text. This forces the physical order of service flows and the logical order of fault propagation in the power communication network to be strictly mapped to the reading order of the large model's text input, constructing a linear text context that conforms to causal laws. Through the above processing, the weight allocation of the large model's attention mechanism on the topology path can be effectively optimized, guiding the model to prioritize upstream root cause nodes and sequentially deduce to downstream damaged nodes, thereby improving the logical accuracy and interpretability of fault root cause localization.

[0048] More specifically, in a concrete example of this application, the process first performs a topology hierarchy and logical order parsing operation. Based on the service routing configuration data and fault propagation direction of the power communication network, it determines the upstream and downstream relative positions of each network element node in the DSL code segment within the service chain, and establishes a logical sorting index with the service source end as the starting point and the destination end as the ending point. Next, it performs a code segment rearrangement operation. Based on the established logical sorting index, the code segment describing the core aggregation layer device or upstream primary routing node is placed at the beginning of the text sequence, and the code segment describing the access layer device or downstream backup routing node is placed at the end, thus forming a code sequence with a clear causal orientation. Subsequently, it performs a semantic guidance information injection operation, inserting natural language annotation text or guiding statements at the beginning, end, and key transition points of the sorted code segments. These statements are used to explicitly explain the starting boundary of the topology and the service flow direction. Finally, it performs a text stream merging operation, concatenating the sorted and guided code segments into a continuous string object in a determined order, removing redundant whitespace characters and formatting noise, and generating the final serialized topology description text.

[0049] Specifically, in step S5, the serialized topology description text and standardized event objects are assembled and optimized using thought-chain-guided prompts to obtain structured prompts containing topological logical constraints. These structured prompts are then input into a large language model for reasoning interaction to output an operation and maintenance decision report. It should be noted that while general-purpose large language models possess powerful generalized semantic understanding capabilities, when faced with highly specialized and structurally complex power communication network operation and maintenance tasks, they often lack a deep understanding of the underlying physical topology and service protection logic, making them prone to reasoning illusions. Furthermore, the randomness of their output results makes it difficult to guarantee the stability and accuracy of decision-making results when processing multi-source heterogeneous alarm data. Based on this, the technical solution of this application further assembles and optimizes the serialized topology description text and standardized event objects using thought chain-guided prompts to obtain structured prompts containing topological logical constraints. These structured prompts are input into a large language model for reasoning interaction to output an operation and maintenance decision report. This approach, based on a clear definition of the operation and maintenance expert's role and task background, uses thought chain technology to decompose complex fault diagnosis tasks into a series of intermediate reasoning steps reflecting the logical structure of power business. The serialized topology structure is explicitly injected as a hard constraint into the model's context window, forcing the model to follow the logical path of first analyzing the topology propagation path and then combining alarms to determine the root cause. Through the above processing, the logical transparency and reasoning robustness of the large model under complex concurrent fault scenarios can be effectively improved, completely eliminating attribution errors caused by topology fragmentation, and ensuring that the final output operation and maintenance decision report has high logical self-consistency and physical interpretability. This achieves a fully intelligent closed loop from passive alarm reception to proactive and accurate decision-making.

[0050] Figure 6 To generate large-model prompt words for power communication network operation and maintenance according to embodiments of this application, the method assembles and optimizes sequential topology description text and standardized event objects using thought chain-guided prompt words to obtain structured prompt words containing topological logical constraints. These structured prompt words are then input into a large language model for reasoning interaction to output a flowchart of the operation and maintenance decision report. For example... Figure 6 As shown, step S5 includes: S51, reading a preset template containing system prompts and task background descriptions from the prompt template library, and initializing it as an initial prompt skeleton with blank spaces to be filled; S52, filling the topology environment slot and alarm fact slot of the initial prompt skeleton with serialized topology description text and converted natural language list form of standardized event objects, respectively, to generate context-filled prompts; S53, injecting a thought chain instruction containing step-by-step guidance logic at the end of the context-filled prompts to obtain structured prompts.

[0051] In step S51, a pre-set template containing system prompts and task background descriptions is read from the prompt word template library and initialized as an initial prompt word skeleton with blank spaces to be filled. It should be noted that, due to the lack of clear role settings and task background constraints, the reasoning output of general-purpose large language models often exhibits high randomness and divergence. Furthermore, power communication network operation and maintenance scenarios require extremely high diagnostic accuracy and professionalism. Directly inputting raw alarm data can easily lead to unprofessional general answers or logical illusions due to the lack of context. Therefore, the technical solution of this application further reads a pre-set template containing system prompts and task background descriptions from the prompt word template library and initializes it as an initial prompt word skeleton with blank spaces to be filled. This pre-establishes the role of power communication network operation and maintenance experts and the task boundaries of fault diagnosis for the large language model, and constructs a standardized command input paradigm. Through the above processing, the professional knowledge parameters of the large model in specific vertical fields can be effectively activated, minimizing the arbitrariness of model reasoning and providing a standardized logical container for the accurate injection of subsequent dynamic topology data and alarm facts.

[0052] More specifically, in a concrete example of this application, a persistent prompt word template library is first established and maintained. This library stores standardized instruction templates adapted to different operation and maintenance scenarios, covering various specific task types such as optical transmission interruption analysis, SDH ring network switching diagnosis, and equipment performance degradation assessment. When the prompt word generation process is triggered, based on the current specific operation and maintenance task requirements, the most suitable pre-set template file is retrieved and loaded from this template library through an index matching mechanism. This pre-set template embeds validated, high-weight system prompts, such as explicitly defining that the model must play the role of a senior engineer with extensive experience in power communication networks, and clearly defining the task context of using multimodal data for root cause troubleshooting. Next, the read template text content is parsed in a structured manner to accurately identify the predefined dynamic data anchors or placeholders. These anchors correspond to the serialized topology descriptions and standardized alarm events that will be injected later. By converting these anchors into program-addressable blank spaces to be filled, an initial prompt word skeleton object containing only static logical constraints and role settings, but not yet filled with specific fault data, is instantiated in memory.

[0053] In step S52, the serialized topology description text and the standardized event objects in the form of a converted natural language list are respectively filled into the topology environment slot and alarm fact slot of the initial prompt word skeleton to generate context-filled prompt words. It should be noted that although the initial prompt word skeleton establishes the expert role and task boundary, it is essentially still an abstract logical framework lacking specific fault context. Without injecting specific network space structure and real-time alarm data, the large language model will not be able to conduct substantial causal reasoning for specific operation and maintenance events. Furthermore, directly mixing structured data and topology code in non-natural language format into the prompt words can easily cause obstacles to the model's information reading or semantic confusion. Based on this, the technical solution of this application further fills the serialized topology description text and the standardized event objects in the form of a converted natural language list into the topology environment slot and alarm fact slot of the initial prompt word skeleton to generate context-filled prompt words. This instantiates the abstract reasoning template into diagnostic task instructions with specific spatiotemporal characteristics and ensures that the model can receive physical topology structure and fault fact data in a reading order that conforms to human cognition. Through the above processing, the deep integration of static logical framework and dynamic operation and maintenance data can be effectively achieved, and a complete semantic context that includes both macro-level network connection relationships and micro-level equipment operation indicators can be constructed for the large language model, thereby ensuring that subsequent thought chain reasoning is based on accurate and complete information.

[0054] More specifically, in a concrete example of this application, the standardized event objects containing key-value pairs such as optical power values, bit error counts, and alarm codes are first semantically converted. Natural language generation algorithms then translate these structured technical indicators into a declarative natural language list format that conforms to human reading habits, clearly describing the specific time of the fault, the network elements involved, and the characteristics of the abnormal state. Next, string matching and replacement techniques are used to accurately locate the reserved topology environment slot identifiers and alarm fact slot identifiers in the initial prompt word skeleton. Subsequently, the serialized topology description text describing the service flow and link attributes generated in the previous steps is injected into the topology environment slots, while the converted natural language list format alarm information is injected into the alarm fact slots. This process, through a strict slot mapping mechanism, organically embeds domain-specific language codes describing the spatial structure and natural language text describing time-series facts into a pre-defined logical container, ultimately generating a context-filled prompt word containing a complete panoramic view of the fault.

[0055] In step S53, a thought chain instruction containing step-by-step guiding logic is injected into the end of the context-filled prompt to obtain a structured prompt. It should be noted that although the context-filled prompt generated in the preceding steps already possesses a static topological structure and dynamic alarm facts, large language models, when handling complex logical reasoning tasks involving multi-hop propagation paths, often tend to make jumpy answers based on probabilistic intuition, lacking rigorous intermediate reasoning processes, leading to the inexplicability and logical discontinuity of the final conclusion. Therefore, the technical solution of this application further injects a thought chain instruction containing step-by-step guiding logic into the end of the context-filled prompt to obtain a structured prompt, thereby explicitly constraining the thinking path of the large model and forcing it to follow a logical order conforming to power operation and maintenance troubleshooting specifications. Through the above processing, the deep reasoning ability of the large model can be effectively stimulated, transforming one-step black-box prediction into visualized step-by-step deduction, improving the logical consistency and credibility of fault root cause localization.

[0056] More specifically, in a concrete example of this application, the instruction library is first invoked to retrieve step-by-step guidance corpus suitable for the current fault type from a pre-built thought chain instruction library. Next, the logical arrangement of instructions is performed to construct reasoning guidance text containing explicit sequence identifiers. This text is designed as a series of progressive logical instructions, such as first requiring the model to analyze the upstream convergence relationship of the fault node based on the topology, then requiring the model to determine the causal direction of fault propagation based on the alarm timestamp, then requiring the elimination of secondary alarm noise caused by business switching, and finally requiring a unique root cause determination conclusion based on the above analysis. Subsequently, the text is concatenated and validated, appending the arranged thought chain instructions to the end of the context-filled prompts, and performing an integrity check on the combined text to ensure that the instructions are placed after the context information to maximize attention guidance, thereby generating the final structured prompts used for reasoning.

[0057] In summary, the method for generating large-scale model prompts for power communication network operation and maintenance according to the embodiments of this application is explained. First, standardized and cleaned multi-source alarm data is anchored to the entire network topology map. An anisotropic influence diffusion mechanism based on service protection configuration is introduced to quantify the non-uniform propagation of fault risks to backup routing nodes, thereby accurately extracting context-aware subgraphs containing implicit logical relationships. Subsequently, the topology and link attributes of this subgraph are serialized into a text sequence using a domain description language, and then assembled into structured prompts containing explicit topological constraints using a thought chain-guided strategy. This approach reconstructs discrete alarm data into causal chains with spatial logic, forcing the large language model to follow the objective laws of physical connections and business logic during reasoning, achieving accurate capture and high-quality diagnosis of potential cascading faults.

[0058] As described above, the large-model prompt word generation method for power communication network operation and maintenance according to the embodiments of this application can be implemented in various power information and communication infrastructures and intelligent operation and maintenance management platforms, such as power communication network management center servers, digital twin simulation platforms, or new-generation intelligent operation and maintenance integrated machines. In one possible implementation, this method can be integrated into a unified network management system or panoramic command center as a fault early warning engine or intelligent auxiliary decision-making component. For example, the method can be an independent fault diagnosis application running on an operation and maintenance workstation, or an advanced alarm analysis plugin and handling suggestion generation module for an existing network management system, or a middleware service deployed on the server and distributing handling strategies through an interface; of course, the topology subgraph extraction based on anisotropic propagation, graph structure text serialization, and thought chain-guided structured prompt word assembly module in this method can also run on AI servers equipped with high-performance GPU computing power, edge computing nodes, or dedicated AI acceleration cards, serving as the underlying inference foundation of the power communication intelligent management and control system.

[0059] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A large model prompt word generation method for power communication network operation and maintenance, characterized in that, Comprise: S1: standardize cleaning of collected raw power communication network sensor data stream to generate standardized event objects; S2: analyze device identifiers in standardized event objects, and anchor search and label in pre-set full network topology map to obtain fault node set; S3: centering on fault node set, topological walk and fault related subgraph extraction are carried out in full network topology map to obtain context aware subgraph containing fault propagation path; S4: text serialization and causal chain description are carried out on service flow direction and link attribute in context aware subgraph to obtain serialized topology description text; S5: thought chain oriented prompt word assembly and optimization are carried out on serialized topology description text and standardized event objects to obtain structured prompt word containing topology logical constraint, and the structured prompt word is input into large language model for reasoning interaction to output operation and maintenance decision report. 2.The method of claim 1, wherein the method is characterized by, The raw sensor data stream includes: optical power value of optical transmission equipment, B1 / B2 / B3 error code count of SDH / OTN frame, device CPU / memory utilization, port transceiving packet CRC error statistics and SNMPTrap alarm message. 3.The method of claim 1, wherein the method is characterized by, Step S1, comprising: The collected raw sensor data stream is subjected to jitter suppression and denoising processing based on sliding window to obtain denoised data stream; The denoised data stream is subjected to key operation and maintenance index extraction based on regular engine to obtain structured field set; Based on global standard time reference, linear interpolation or nearest neighbor matching is carried out on each data point in the structured field set to obtain standardized event objects. 4.The method of claim 1, wherein the method is characterized by, Step S2, comprising: The data load of the standardized event object is subjected to string parsing and format normalization processing to extract the device unique identification code; The device unique identification code is used as a search key value to index search and match in the pre-set full network topology map to obtain the anchor network element node; The anchor network element node is injected with dynamic fault semantic label containing timestamp and alarm level to obtain the fault node set. 5.The method of claim 1, wherein the method is characterized by, Step S3, comprising: Each node in the fault node set is taken as a walk starting point, and bidirectional breadth-first search topological walk is carried out in the full network topology map to obtain a candidate neighborhood node set; The node influence score of each node in the candidate neighborhood node set is iteratively calculated to obtain a weighted node list; Based on the set correlation cutoff threshold, the weighted node list is pruned and reconstructed to obtain a context aware subgraph. 6.The method of claim 1, wherein the method is characterized by, Step S4, comprising: The edge and node in the context aware subgraph are subjected to service flow direction analysis and link attribute extraction to obtain an attribute enhanced graph object; The nodes and edges in the attribute enhanced graph object are subjected to DSL based structured translation to obtain a DSL code fragment; The DSL code fragment is subjected to causal chain logic sorting and text splicing to obtain a serialized topology description text. 7.The method of claim 1, wherein the method is characterized by, Step S5, comprising: A pre-set template containing system prompt and task background description is read from the prompt word template library, and is initialized as an initial prompt word skeleton with blank positions to be filled; The serialized topological description text and the converted standardized event object in the form of natural language list are filled into the topological environment slot and the alarm fact slot of the initial prompt word skeleton respectively to generate a context-filled prompt word; A thinking chain instruction containing step-by-step guiding logic is injected at the end of the context-filled prompt word to obtain a structured prompt word.

8. The method of claim 5, wherein the method is characterized by, An iterative calculation of node influence score is performed on each node in the candidate neighborhood node set to obtain a weighted node list, including: Based on the protection configuration data, a logical coupling matrix of the candidate neighborhood node set is constructed; Based on the logical coupling matrix, an anisotropic transition probability estimation is performed on the candidate neighborhood node set to obtain an anisotropic transition tensor; Based on the anisotropic transition tensor, a double-factor weighted influence iteration is performed on each node in the candidate neighborhood node set to obtain a weighted node list.

Images