A method and system for full life cycle management of an optical cable intelligent label

Abstract

A kind of optical cable intelligent label full life cycle management method and system, comprising: based on configuration file, structured data is extracted by XML parsing engine, unstructured text semantic analysis is carried out, symbol quantization text features are combined with fractional regression, and knowledge graph is constructed;According to knowledge graph, the size of text, the position of two-dimensional code and fault tolerance rate are dynamically adjusted according to optical cable type and pasting scene using generative adversarial network technology, and optical cable intelligent label adapted to different scenes is generated;Label image is collected, computer vision technology is used, and structured label identification data containing optical cable type, connection relationship and transmission signal type are converted;The consistency is checked, potential anomalies are analyzed, influence is evaluated and graded warning is generated by comparing identification data with knowledge graph;Microservice architecture system is constructed, data analysis, AI generation, intelligent identification engine and knowledge graph service are integrated, and full life cycle management is realized.The application improves the intelligent level of optical cable label full life cycle management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent management of power systems, in particular to a method and system for whole life cycle management of optical cable intelligent tags, which is suitable for the generation, identification and operation and maintenance management of optical cable tags in intelligent substations. BACKGROUND

[0002] In the construction and operation of smart grids and substations, optical cable tag management is a key link to ensure the reliable operation of communication networks. At present, the mainstream optical cable tag management methods mainly include two modes of manual data entry to generate tags and bar code / QR code scanning identification. For example, the traditional optical cable tag management system usually generates tags by manually entering design data, and then identifies and queries through a handheld scanning device. This method has low initial deployment cost and is simple to operate. Another common method is to use a database-based management system to store optical cable information in a central database, and to realize information query and update through simple code scanning.

[0003] The closest prior art is an optical cable tag management system integrated with basic image recognition functions. This system can identify simple tag content using computer vision technology and match it with the background database. The system uses a static data structure to store optical cable information, uses a preset tag template to generate optical cable tags, and uses basic image processing algorithms to identify and analyze the tags. In the operation and maintenance link, the system also provides a simple maintenance record function to support recording the maintenance history of the optical cable.

[0004] However, this prior art has obvious technical defects: first, the tag generation process relies on manual entry of design data, which is prone to errors and inefficient; second, the tag identification technology has poor adaptability to complex scenarios such as optical cable tag contamination and inclination, and the recognition success rate is less than 90% in actual harsh environments; third, it lacks intelligent data correlation and analysis capabilities, and cannot effectively cope with the dynamic changes of physical and logical loop scenarios in intelligent substations, making it difficult to support real-time monitoring and abnormal warning of optical cable connection relationships; finally, the system architecture is closed, making it difficult to interact and integrate with other systems, limiting the management efficiency of the whole life cycle of optical cables. SUMMARY

[0005] The present application aims to provide a method and system for whole life cycle management of optical cable intelligent tags, which aims to solve the technical problems of existing technologies, such as reliance on manual entry for optical cable tag generation, poor adaptability of identification technology, lack of intelligent data correlation and analysis capabilities, and closed system architecture.

[0006] To achieve the above-mentioned purpose, the present application provides a method for whole life cycle management of optical cable intelligent tags, comprising the following steps:

[0007] Based on the smart substation configuration file including the SCD file and the SPCD file of the smart substation, structured data including logical nodes, logical devices, communication connections and physical device descriptions are extracted through an XML parsing engine, semantic analysis is performed on unstructured text by using natural language processing technology, text features are quantitatively processed by combining a symbolic quantile regression method, and a knowledge graph containing the association relationship of devices-ports-optical cables-virtual terminals is constructed;

[0008] According to the knowledge graph, the generative adversarial network technology is used to dynamically adjust the text size, the position of the two-dimensional code and the fault tolerance rate according to the optical cable type and the pasting scene, and to generate optical cable intelligent labels adapted to different scenes;

[0009] Based on the optical cable intelligent label, a label image of the field optical cable is collected, and a multi-modal fusion computer vision technology is used to convert it into structured label recognition data containing the optical cable type, the connection relationship and the transmission signal type;

[0010] Relying on the structured label recognition data and the knowledge graph, a real-time data comparison algorithm is used to check the consistency of the label information and the expected connection relationship, analyze potential abnormalities, use a multi-objective evolutionary algorithm to evaluate the safety impact, reliability impact and functional impact of the abnormalities and generate graded warning information and targeted processing suggestions;

[0011] A modular system based on a microservice architecture is constructed, including a data parsing engine, an AI generation engine, an intelligent recognition engine and a knowledge graph service, and through standardized API interfaces, data exchange and functional integration with other systems of the smart substation are supported, realizing intelligent management of the whole life cycle of optical cable labels from generation, recognition to operation and maintenance.

[0012] Preferably, based on the smart substation configuration file including the SCD file and the SPCD file of the smart substation, structured data including logical nodes, logical devices, communication connections and physical device descriptions are extracted through an XML parsing engine, semantic analysis is performed on unstructured text by using natural language processing technology, text features are quantitatively processed by combining a symbolic quantile regression method, and a knowledge graph containing the association relationship of devices-ports-optical cables-virtual terminals is constructed, including:

[0013] Based on the smart substation configuration file, a graph neural network technology is used to mine the implicit connection relationship between devices, identify the indirect association of cross-bay optical cables, and establish a multi-level graph structure association network containing physical devices, ports, optical cables and virtual terminals in the entity layer, connection relationships, transmission relationships and ownership relationships in the relationship layer, and device models, port protocols and optical cable core numbers in the attribute layer;

[0014] The reservoir computing technology is used to predict potential cable connection modes based on historical connection modes, the Lyapunov index is used to analyze the reliability of a dynamically adjusted prediction strategy, missing cable purpose labels and spare core information are automatically completed, and completed label information is obtained;

[0015] According to the completed label information and historical operation and maintenance data including historical operation and maintenance records, alarm logs, cable test data and label use feedback, a multi-objective evolutionary algorithm with decision space diversity is used to simultaneously consider data integrity, consistency and timeliness for multi-dimensional data fusion, connection relationship conflicts, obsolete information and abnormal patterns are detected and corrected based on the uniformity of Hamming distance, and a verified and optimized knowledge graph is obtained.

[0016] Preferably, the binding symbol quantile regression method quantitatively processes text features, including:

[0017] For the unstructured text containing cabinet names, port descriptions and virtual terminal signals, the power field professional vocabulary and rule base are used for word segmentation and part-of-speech tagging, symbolic processing is performed to map professional terms and descriptions into symbolic representations, and symbolic text is generated;

[0018] Based on the symbolic text, a quantile-based regression model is established to map text semantic features to different quantiles, analyze the correlation strength between symbols, calculate the conditional quantiles of different features to determine the importance of symbols in the cable label and determine the importance weight in the cable label;

[0019] Based on the importance weight, the cable types including main cable, tail cable and single / multi-mode optical fiber are extracted, the connection relationship between devices including signal flow direction and logical relationship is extracted, the transmission signal types including GOOSE signal, SV sample value and MMS message are extracted, the priority information including key protection signal and non-key monitoring signal is extracted, and the quantitatively processed text features are obtained.

[0020] Preferably, the generation of the cable intelligent label adapted to different scenarios includes:

[0021] The knowledge graph is used to analyze the physical space constraints, environmental factors and viewing condition characteristics of different application scenarios including narrow space of cabinets, outdoor environment and high-frequency review area, historical label use data is learned by reservoir computing technology to predict label use environment parameters, and an initial label template candidate set including compact label, waterproof and weather-resistant label and wear-resistant label is generated.

[0022] According to the initial label template candidate set, label content data containing cable ID and connection information, and target scene parameters, a generative adversarial network technology is used, a label design scheme is generated by a generator according to the cable type and the pasting scene, the information integrity, readability and recognition rate of the label design are evaluated by a discriminator based on historical label samples, content constraints are added to ensure that key information is clear and visible, two-dimensional code constraints adjust the size, position and fault tolerance rate according to the use environment, space constraints optimize the overall layout according to the pasting position, and visual constraints ensure the contrast of text and background, to generate an optimized label layout scheme;

[0023] Based on the optimized label layout scheme and historical label use data containing label durability, scanning success rate, a reinforcement learning model is constructed, the influence of parameter adjustment on label performance is analyzed by Lyapunov index, the two-dimensional code encoding density, error correction level and ultraviolet resistance coding ability are adaptively optimized, the text font size, thickness and contrast are adaptively optimized, the material parameters including label thickness, adhesive type and protective layer treatment are adaptively optimized, and the cable intelligent label is obtained.

[0024] Preferably, the generative adversarial network technology is used to optimize the label design, including:

[0025] A large number of historical label samples containing successful cases, failed cases and expert annotations and corresponding use effect evaluations are used as training data, a generative adversarial network model including a generator responsible for generating a label design scheme and a discriminator evaluating whether the generated label design meets the expected standard is trained, content constraints ensuring that key information is clear and visible, two-dimensional code constraints automatically adjusting the size, position and fault tolerance rate of the two-dimensional code according to the use environment, space constraints optimizing the overall layout according to the available space of the pasting position, and visual constraints ensuring the contrast of text and background to improve readability are added, and a GAN model under the constraint condition is formed;

[0026] According to the GAN model under the constraint condition, different types of label design schemes are generated by specially adjusting the ODF label simplifying information display and highlighting port number and cable direction, the device-to-device connection label emphasizing clear indication of device information and signal type, and the tail cable label optimizing information layout under small size to ensure readability, considering the resolution limit of the printing device, the ink absorption characteristics of different materials, the possible deformation in the pasting process, and the visual changes in the aging process;

[0027] Based on the multiple label design schemes, multiple rounds of optimization iteration are performed to balance the functional requirements, aesthetics and practicality, considering the functional requirements to meet the label recognition and information transmission requirements, the aesthetics to improve the visual effect, and the practicality to ensure the actual use convenience, to obtain the optimized label layout scheme.

[0028] Preferably, the conversion is a structured label recognition data containing cable type, connection relationship and transmission signal type, comprising:

[0029] Based on the optical cable intelligent label, the label image collected by the mobile terminal is combined with the environmental supplementary data obtained in the label pollution, line of sight obstruction and high-density optical cable area, the illumination intensity, collection angle and collection distance are recorded synchronously, the adaptive histogram equalization, non-local mean filtering and super-resolution reconstruction image enhancement algorithm are applied, the geometric correction algorithm and illumination compensation algorithm are applied, and the standardized label image with uniform brightness, contrast, viewing angle and resolution is generated;

[0030] According to the standardized label image, the two-dimensional code area is located and extracted through gradient analysis, morphological verification and positioning point detection, the low-resolution or blurred two-dimensional code is processed by using the residual learning network combined with the perception loss function super-resolution reconstruction technology, the partially worn or broken two-dimensional code is processed by using the edge repair technology of structural integrity analysis, context perception filling and topological structure preservation, the attention mechanism of spatial attention focusing effective area, channel attention dynamically adjusting feature weight and multi-scale fusion capturing local and global features is combined, the two-dimensional code recognition is realized through adaptive binaryzation, perspective correction and error correction enhancement decoding, and the two-dimensional code recognition result is obtained;

[0031] Based on the two-dimensional code recognition result, the multi-objective evolutionary algorithm of decision space diversity is introduced, which encodes the key parameters and strategies in the recognition process into decision vectors, generates an initial solution set with diversity based on the uniformity measurement of Hamming distance, and obtains multiple complementary recognition strategies through double-layer selection of Pareto advantage and decision space diversity, at the same time, the optical character recognition is carried out on the text information on the label image, the multi-modal fusion technology of information consistency verification, complementary information integration and redundant information utilization is used to fuse the two-dimensional code information and the text recognition result, the natural language processing technology is used to convert into a standardized data structure, and the structured label recognition data containing cable type, connection relationship and transmission signal type is obtained.

[0032] Preferably, based on the optical cable intelligent label, the label image of the on-site optical cable is collected, and the multi-modal fusion computer vision technology is used to convert into the structured label recognition data containing cable type, connection relationship and transmission signal type, further comprising:

[0033] The key parameters and strategies in the recognition process are encoded into decision vectors, the Hamming distance between each decision vector in the solution set is calculated, and the diversity index is taken as an additional optimization target by designing a diversity evaluation function based on density;

[0034] According to the diversity evaluation function, an initial solution set with diversity is generated, new candidate solutions are generated by mutation and crossover to explore the decision space, a double-layer selection is performed based on the Pareto advantage and the diversity of the decision space, a dynamic adjustment mechanism is used to adjust the diversity weight according to the convergence in the optimization process, adjust the learning rate according to the sensitivity of different parameters, and dynamically balance the exploration and utilization, and a variety of complementary identification strategies are obtained;

[0035] Based on the plurality of complementary identification strategies, a strategy that pays more attention to light compensation is selected in a strong light interference environment, a more aggressive image repair strategy is selected in a serious pollution situation, a simplified strategy with higher calculation efficiency is selected when the calculation resources are limited, and the most suitable identification method for the current situation is dynamically selected or combined according to the real-time detected environmental conditions, so as to realize the diversity enhancement of the decision space.

[0036] Preferably, the structured tag identification data and the knowledge graph are used to check the consistency of the tag information and the expected connection relationship by a real-time data comparison algorithm, analyze potential abnormalities, evaluate the safety impact, reliability impact and function impact of the abnormalities by using a multi-objective evolutionary algorithm, and generate hierarchical warning information and targeted processing suggestions, including:

[0037] Based on the structured tag identification data and the knowledge graph, the topological relationship and logical function are matched by multi-dimensional feature matching, a fuzzy matching strategy of character similarity calculation, semantic similarity evaluation and structure similarity analysis is adopted, multi-source data collaborative verification is performed by integrating design document data, historical scanning records and related optical cable data, historical abnormal cases are learned by using an effective prediction time method of reservoir computing technology, and abnormal modes including mismatch, tag error, unauthorized change and potential risk are configured, and Lyapunov index is calculated to evaluate the abnormal influence time scale and range, the case that the physical connection does not match the design drawing is detected, and the abnormal detection result is obtained;

[0038] Based on the abnormal detection result and the tag operation and maintenance history data including fault history, maintenance record and performance trend, the severity and potential impact of the abnormality are evaluated from multiple angles including safety threat degree, reliability impact, function impact evaluation function loss range and economic impact evaluation repair cost and resources by using a multi-objective evolutionary algorithm, a balance point is found in multiple evaluation dimensions, and the situational sensitivity of load condition, seasonal factor and concurrent event is considered, hierarchical warning information including emergency warning, important warning, general warning and attention prompt is generated according to the severity of the evaluation, similar abnormal case reasoning is searched based on the historical case library, and rule reasoning is applied based on the field expert rules, and targeted processing suggestions including emergency processing scheme, temporary relief scheme, root solution and prevention strategy are provided;

[0039] On the basis of the execution result of the targeted processing suggestion and the actual connection state confirmed on site, a change source identification and multi-level change confirmation process including plan changes, on-site discoveries and fault processing is established, an incremental update strategy is used to add, modify or delete entity nodes and connection relationships affected by changes using graph neural network technology, the direct impact, indirect impact and redundancy evaluation of network topology changes are re-evaluated, the connection relationship data in the knowledge graph is dynamically updated and the change version history is maintained, and dynamic association and abnormality early warning are realized.

[0040] Preferably, the modular system based on the micro-service architecture includes a data parsing engine, an AI generation engine, an intelligent identification engine and a knowledge graph service, supports data exchange and functional integration with other systems of the smart substation through standardized API interfaces, and realizes intelligent management of the optical cable label throughout its life cycle, including:

[0041] The hardware system includes a mobile terminal equipped with a GPU supporting AI inference, a high-definition camera with anti-shake and light compensation functions, an RFID identification module for auxiliary identification and a label printing device supporting multiple materials, the terminal processing capability is optimized through edge computing technology to support AI inference and image processing, ensuring data processing and label identification availability in weak network or offline environments, and a usable hardware system is obtained.

[0042] Based on the usable hardware system, the core modules including a data parsing engine for processing the smart substation configuration file combined with NLP and GNN technology, an AI generation engine for label optimized generation combined with GAN and reinforcement learning technology, an intelligent identification engine for multi-modal analysis combined with CV and multi-modal technology, and a knowledge graph service for storing association relationships and historical data are developed, containerization technology is used to realize loose coupling and flexible deployment between modules, and a micro-service architecture supporting independent development, deployment and expansion is formed.

[0043] Based on the micro-service architecture, a hierarchical data storage strategy is designed for different types of data to separate cold and hot data, a multi-level security mechanism is built including data encryption to protect sensitive information, access control to limit operation permissions and operation audit record system, a standardized API interface is developed to support data exchange and functional integration with other systems of the substation such as the SCADA system and the asset management system, a flexible plug-in mechanism is designed to allow third-party development and extension of functions, and full life cycle management is realized.

[0044] The application also provides an optical cable intelligent label full life cycle management system, which comprises:

[0045] A data analysis module is configured to extract structured data including logical nodes, logical devices, communication connections and physical device descriptions from smart substation configuration files including SCD files and SPCD files through an XML analysis engine, perform semantic analysis on unstructured text using natural language processing technology, quantitatively process text features by combining symbolic quantile regression methods, and construct a knowledge graph containing device-port-cable-virtual terminal association relationships.

[0046] A label generation module is configured to dynamically adjust text size, QR code position and fault tolerance according to cable types and pasting scenarios based on the knowledge graph, and generate smart cable labels adapted to different scenarios using generative adversarial network technology.

[0047] A label recognition module is configured to collect label images of field cables based on the smart cable labels, and convert them into structured label recognition data containing cable types, connection relationships and transmission signal types using multi-modal fusion computer vision technology.

[0048] An anomaly warning module is configured to check the consistency of label information and expected connection relationships by real-time data comparison algorithms, analyze potential anomalies, evaluate the safety impact, reliability impact and functional impact of anomalies using multi-objective evolutionary algorithms, and generate graded warning information and targeted processing recommendations.

[0049] A system integration module is configured to build a modular system based on a microservice architecture, including a data analysis engine, an AI generation engine, an intelligent recognition engine and a knowledge graph service, and support data exchange and functional integration with other systems of a smart substation through standardized API interfaces, realizing intelligent management of the entire life cycle of cable labels from generation, recognition to operation and maintenance.

[0050] The present application has the following advantages:

[0051] 1. The connection relationship of the cable is automatically parsed from the smart substation configuration file, a comprehensive knowledge graph is constructed, manual input errors are greatly reduced, and data accuracy is improved.

[0052] 2. Through generative adversarial networks and reinforcement learning techniques, adaptive optimization of label design is achieved, making the label more suitable for different scene usage requirements and improving the practicality and durability of the label.

[0053] 3. The multi-objective evolutionary algorithm with decision space diversity is innovatively introduced, combined with the uniformity measurement based on Hamming distance, which significantly improves the label recognition accuracy under non-ideal conditions such as contamination and inclination, and increases the recognition success rate from 90% of traditional methods to more than 99.5%.

[0054] 4. The effective prediction time method based on reservoir computing technology is combined with a multi-objective evolutionary algorithm to realize accurate detection and evaluation of optical cable connection abnormalities, support hierarchical early warning, and generate targeted processing suggestions;

[0055] 5. A microservice architecture and standardized API interface are adopted to build an open and extensible system framework, support seamless integration with other systems, and realize full life cycle management from label generation and identification to operation and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0057] Figure 1 is a method flow chart of the optical cable intelligent label full life cycle management of the present application;

[0058] Figure 2 is a system structure diagram of the optical cable intelligent label full life cycle management of the present application. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0060] As shown in Figure 1 The optical cable intelligent label full life cycle management method provided by the present application includes the following steps:

[0061] S1: Based on the intelligent substation configuration file including the intelligent substation SCD file and the SPCD file, the structured data including logical nodes, logical devices, communication connections and physical device descriptions are extracted through an XML parsing engine, the unstructured text is subjected to semantic analysis by using natural language processing technology, the text features are quantitatively processed by combining with the symbolic quantile regression method, and a knowledge graph containing the association relationship of device-port-optical cable-virtual terminal is constructed;

[0062] In the optical cable intelligent tag full life cycle management method of the application, first, step S1 is performed, that is, the optical cable tag intelligent data model layer is constructed. This step is the basis of the whole, and the knowledge graph required for optical cable tag management is established by deeply analyzing and processing the configuration files of the intelligent substation. The SCD (System Configuration Description) file and the SPCD (Substation Physical Connection Diagram) file in the intelligent substation are two types of core configuration files, which respectively describe the logical connection relationship and the physical connection relationship of the substation. First, the XML parsing engine specially designed is used to parse these files, and the key structured data is extracted, including the logical node (LN) information (such as protection function, measurement function, etc.), the logical device (LD) information (clear function grouping), the communication connection information (including GOOSE, SMV and other message types and communication paths) and the physical device description (such as screen cabinet position, terminal row number, etc.). In the parsing process, a distributed processing architecture is adopted to process the complex configuration files of large substations in segments in parallel, and a strict error detection and fault tolerance mechanism is established to ensure the accuracy and integrity of the data.

[0063] For the unstructured text in the substation configuration file, the natural language processing (NLP) technology is introduced for deep analysis. These unstructured texts include device description, port naming, signal description, etc., which are usually in the form of natural language written by artificial, and contain rich implicit semantics. A self-defined professional vocabulary and rule library in the power field are used to perform word segmentation and part-of-speech tagging on the text, and special attention is paid to the power-specific terms and expressions, such as "110kV bus PT signal", "circuit breaker tripping control loop" and other professional terms. In order to convert these text information into quantifiable features, the symbolic quantile regression (SQR) method is innovatively used for processing. The SQR method maps the text semantic features to different quantiles, calculates the conditional quantiles of different features by analyzing the correlation strength between symbols, and determines the importance and weight of each element in the optical cable tag. This method can accurately extract the optical cable type (such as main optical cable, tail cable, single / multi-mode optical fiber, etc.), the connection relationship (signal flow direction and logical relationship between devices), the transmission signal type (such as GOOSE signal, SV sampling value, MMS message, etc.) and the priority information (such as key protection signal, non-key monitoring signal, etc.) from the unstructured text. Through these extracted features, the knowledge graph containing the complete association relationship of device-port-optical cable-virtual terminal is constructed, providing a solid data foundation for subsequent tag generation and recognition.

[0064] S2: According to the knowledge graph, the generative adversarial network technology is used to dynamically adjust the text size, two-dimensional code position and fault tolerance rate according to the cable type and pasting scene, and generate intelligent cable labels that adapt to different scenes;

[0065] Step S2 implements the intelligent generation layer of the cable label. Based on the knowledge graph constructed in the previous step, advanced artificial intelligence technology is used to generate intelligent cable labels that adapt to different scenes. First, the characteristics of different application scenarios are analyzed through reservoir computing technology (RC), including physical space constraints (such as narrow space in screen cabinets, dense area of wire slots), environmental factors (such as outdoor exposure environment, high temperature area), viewing conditions (such as low light environment, long distance identification requirement) and use frequency, etc. RC technology learns the correlation between label parameters and use effect in different scenarios by processing historical label use data, thereby predicting the optimal label design parameters. A training data set containing scene characteristics and label effect evaluation is constructed, the key hyperparameters of the RC model are optimized, and an initial label template candidate set optimized for different scenarios is generated, such as compact label (suitable for narrow space in screen cabinet), waterproof and weather-resistant label (suitable for outdoor environment) and wear-resistant label (suitable for high-frequency reading area) etc.

[0066] After obtaining the initial label template candidate set, generative adversarial network (GAN) technology is introduced to optimize the label design in depth. The GAN model consists of a generator and a discriminator. The generator is responsible for generating label design schemes according to the cable type and pasting scene, and the discriminator evaluates the information integrity, readability and recognition rate of the design based on historical label samples. Various constraints are added in GAN training: content constraint ensures that key information is clear and visible; two-dimensional code constraint adjusts size, position and fault tolerance rate according to the use environment; space constraint optimizes the overall layout according to the pasting position; visual constraint ensures the contrast between text and background. Special adjustments are also made for different types of cable labels, such as ODF label (simplified information display, highlighting port number and cable direction), inter-device connection label (emphasizing device information on both ends, clearly indicating signal type) and tail cable label (optimizing information layout in small size) etc. The label design generated by GAN takes into account various practical factors such as the resolution limit of the printing equipment, the characteristics of the label material, the possible deformation during the pasting process and the visual changes during the aging process, etc.

[0067] Among them, the generative adversarial network (GAN) technology applied to the optimization of cable label design is an innovative artificial intelligence solution that automatically generates label design schemes that meet specific requirements through deep learning algorithms. This technology combines artistic design with engineering needs, solving the problems of high dependence on manual experience, low standardization and poor adaptability in traditional label design, and realizing the intelligent and personalized customization of label design.

[0068] From the technical principle, the GAN model of the application contains a specially designed generator and discriminator network. The generator adopts the U-Net architecture, contains an encoder-decoder structure and a skip connection, and can effectively preserve information details. The encoder extracts the feature representation of the input conditions (such as cable type, scene parameters) through multiple layers of convolution, and the decoder reconstructs the complete label design through transposed convolution. The discriminator adopts a multi-task learning structure, not only to evaluate the authenticity of the design, but also to specially evaluate key indicators such as information integrity, readability and expected recognition rate. To adapt to the special needs of label design, the GAN model introduces a conditional control mechanism, which injects design requirements (such as label type, use environment) as a condition vector into the generation process, ensuring that the generated results meet the needs of specific scenarios. The model training uses an improved Wasserstein GAN loss function, which adds multiple specific task loss terms such as content preservation loss, clarity loss and structure consistency loss, and forms the final optimization target through weighted combination.

[0069] The detailed implementation steps include six key stages: first, demand analysis and parameter extraction, analyze label use requirements, extract key design parameters such as functional requirements, environmental conditions, space limitations and identification methods, and convert these requirements into a structured condition vector. The second stage is to generate the initial design, the generator network generates 10-20 candidate design schemes based on the condition vector and random noise, each scheme contains complete layout, color scheme, font setting and two-dimensional code configuration. The third stage performs multi-dimensional quality evaluation, the discriminator scores each scheme for information integrity, readability, expected recognition rate and aesthetics, and calculates the overall quality score through weighted combination. The fourth stage verifies the constraint conditions, checks whether the design meets the information priority, two-dimensional code quality, minimum font size and space utilization, and filters out unqualified schemes. The fifth stage carries out environmental adaptability simulation, simulates the visual effect and performance of the label under different conditions in a virtual environment, including different light, angle, material aging and pollution, and evaluates the two-dimensional code recognition rate, calculates the environmental adaptability index. The last stage is iterative optimization and scheme selection, the system selects 2-3 best schemes for 3-5 rounds of deep optimization, specifically solves weaknesses such as increasing key information contrast, adjusting two-dimensional code parameters or optimizing layout. Finally, all evaluation indicators are integrated to select the optimal scheme and generate high-resolution design files and detailed implementation guidelines.

[0070] Through this deep optimization process, GAN technology realizes the transformation from traditional manual design to intelligent, data-driven design, not only shortens the design time from several hours to several minutes, but also increases the recognition success rate of the label by 15-20%, while ensuring the adaptability of the design in various complex environments, providing strong technical support for the optical cable management of intelligent substation.

[0071] To further optimize the label parameters, a reinforcement learning model is constructed to analyze the impact of parameter adjustment on label performance using Lyapunov exponents. The reinforcement learning model includes a state space (describing the current parameter configuration of the label), an action space (possible parameter adjustment operations), a reward function (evaluating the pros and cons of parameter adjustment according to the use effect of the label), and a policy network (learning the optimal parameter adjustment strategy). Through Lyapunov exponent evaluation of parameter stability and long-term effect prediction, the parameters of two-dimensional code (coding density, error correction level, anti-ultraviolet coding ability), text (font size, thickness, contrast), and material (label thickness, adhesive type, protective layer treatment) are adaptively optimized. The reinforcement learning model will continuously adjust the parameter selection strategy based on successful and failed cases in historical data, such as analyzing failed labels in high-temperature environments to optimize material selection, adjusting text and two-dimensional code parameters according to the output effect of different printing equipment, etc. Through this data-driven parameter optimization, high-quality intelligent cable labels that adapt to different scenarios can be generated, significantly improving the practicality and durability of the labels.

[0072] Among them, the construction and training process of the reinforcement learning model can be illustrated through a substation field application case. First, we collected comprehensive data on 100 installed optical cable labels in the substation, including their complete parameter configurations (such as two-dimensional code size, error correction level, material type, etc.) and actual performance data under different environmental conditions (such as recognition success rate, aging speed, etc.). Based on these data, an initial state space is constructed, using a 16-dimensional vector to represent the label parameter configuration, with each dimension corresponding to a key parameter. The action space is designed as discrete parameter adjustment operations, such as "increase two-dimensional code size by 10%", "increase error correction level by one level", etc., totaling 42 basic adjustment operations.

[0073] The reward function design is the core of the model, and a comprehensive score is constructed combining multiple performance indicators: R = 0.4 x recognition success rate + 0.3 x durability index + 0.2 x information integrity + 0.1 x cost-effectiveness. Among them, the recognition success rate is obtained through actual testing under different conditions; the durability index is estimated based on accelerated aging tests and field use data; the information integrity evaluates the ability of the label to maintain information readability during use; and the cost-effectiveness considers the cost changes brought by parameter adjustment.

[0074] The policy network adopts a deep Q-network (DQN) architecture, containing 4 fully connected layers with 128, 256, 128, and 42 neurons respectively (the last layer corresponds to the number of actions). The training process uses the experience replay technique, maintaining a replay buffer containing 10,000 experience records. The model is first pre-trained in a simulated environment, using an environment simulator constructed from historical data to allow the model to learn the relationship between parameters and performance by trying different parameter adjustments. After pre-training, the model is deployed in the actual environment for online learning, and whenever new labels are generated and actual usage feedback is obtained, the experience library is updated and the model is fine-tuned.

[0075] Liapunov exponent analysis played a key role in the training process. For example, it was found that in outdoor high-humidity environments, when the size of the QR code module was less than 0.8mm and the label thickness was less than 0.2mm, the Liapunov exponent increased sharply, indicating that in this parameter region, the system was extremely unstable and the recognition rate would quickly decrease over time. Based on this analysis, the model learned to automatically increase the size of the QR code module and the thickness of the label in outdoor high-humidity areas.

[0076] After about 3 months of training, the model gradually converged to an efficient strategy. A typical application case is when the model detects that the label will be deployed near a cooling device (with large temperature fluctuations and high humidity), it automatically recommends increasing the QR code error correction level from the standard L level to the H level, adjusting the label material from ordinary PVC to composite PET material, and adding an ultraviolet-resistant coating.

[0077] S3: Based on the optical cable intelligent label, collect the label image of the on-site optical cable, and use multi-modal fusion computer vision technology to convert it into structured label recognition data containing optical cable type, connection relationship, and transmission signal type;

[0078] Step S3 develops an intelligent recognition layer for optical cable label. This step realizes high-precision recognition of on-site optical cable labels through multi-modal fusion computer vision technology. In the complex environment of a substation, labels may be in various non-ideal states, such as uneven lighting, angle tilt, partial obstruction or contamination, etc., which brings challenges to label recognition. First, high-definition cameras mounted on mobile terminals (such as industrial tablets or smartphones) are used to capture label images. These cameras have high resolution, auto-focus, anti-shake functions, and light compensation features. At the same time, environmental supplementary data is also collected, including through RFID assisted recognition (especially in label contamination, line of sight obstruction or high-density optical cable areas) and environmental parameters (such as light intensity, collection angle and distance). After obtaining the original data, a series of preprocessing steps are performed, including image enhancement (adaptive histogram equalization, non-local mean filtering, super-resolution reconstruction), geometric correction (perspective transformation, distortion correction, size standardization), light compensation (light model estimation, reflection component separation, shadow removal) and region positioning (label detection, region of interest extraction, multi-label separation) etc. These processes convert the original images collected under various conditions into standardized high-quality label images, providing ideal input data for subsequent two-dimensional code recognition.

[0079] For the two-dimensional code on the label, based on the standardized image, the two-dimensional code region is located and extracted, and then integrated CV enhancement algorithms are applied for processing. These algorithms include super-resolution reconstruction technology (based on residual learning network, combined with perception loss function, for low-resolution or blurred two-dimensional codes), edge repair technology (through structure integrity analysis, context-aware filling and topology structure preservation, for partially worn or broken two-dimensional codes) and attention mechanism (spatial attention focusing on effective areas, channel attention dynamically adjusting feature weights, multi-scale fusion capturing local and global features). After these processes, through adaptive binarization, perspective correction and error correction enhancement decoding, high-precision two-dimensional code recognition is achieved. To further improve the recognition ability under extreme conditions, a multi-objective evolutionary algorithm with decision space diversity is innovatively introduced. This algorithm encodes the key parameters and strategies in the recognition process into a decision vector, evaluates the diversity of the decision space through uniformity measurement based on Hamming distance, and generates multiple complementary recognition strategies. This enables the dynamic selection or combination of the most suitable recognition method for the current situation according to the real-time detected environmental conditions, significantly improving the label recognition accuracy under non-ideal conditions such as contamination and tilt. Finally, the text information on the label is recognized by optical character recognition, and the two-dimensional code information and the text recognition results are integrated through multi-modal fusion technology, converting into structured label recognition data containing cable type, connection relationship and transmission signal type.

[0080] Among them, the multi-objective evolutionary algorithm with decision space diversity is an innovative computational optimization method designed specifically for the extreme condition identification problem of two-dimensional codes in optical cable labels. Based on the principle of multi-objective evolutionary computation, the algorithm maintains a highly diverse solution set in the decision space, achieving robust optimization of two-dimensional code recognition in various complex environments. Unlike traditional single recognition strategies, this method builds an integrated system containing multiple complementary recognition strategies, capable of dealing with various unpredictable recognition challenges.

[0081] From a technical perspective, the algorithm encodes key decision parameters in the two-dimensional code recognition process into a multi-dimensional decision vector, including preprocessing parameters (such as filter type, kernel size, enhancement intensity, etc.), binarization parameters (such as threshold method, local window size, etc.), perspective correction parameters (such as transformation matrix estimation method, reference point selection strategy, etc.), and decoding parameters (such as sampling rate, fault tolerance processing strategy, etc.). On this basis, the algorithm defines two main optimization objectives: maximum recognition accuracy and maximum decision space diversity. Among them, the decision space diversity is evaluated by the uniformity measure based on Hamming distance, ensuring that the generated strategy set is uniformly distributed in the decision space, covering different types of recognition scenarios.

[0082] The implementation steps of the algorithm include four key stages. First, initialize the population and randomly generate a set of decision vectors, each representing a recognition strategy. The population size is usually set to 20-30 individuals to ensure initial diversity. The second stage is fitness evaluation, which evaluates the performance of each decision vector on multiple objectives. The recognition accuracy objective is calculated by the average recognition success rate on the standard test set (containing two-dimensional code samples with different degrees of contamination, inclination, and illumination changes); the decision space diversity objective is evaluated by calculating the average Hamming distance between the current individual and other individuals in the population, the greater the distance, the higher the diversity. The third stage is evolution operation, including elite preservation, crossover and mutation. The algorithm uses non-dominated sorting and crowding distance calculation to select elite individuals; uses simulated binary crossover (SBX) to generate new offspring strategies in the decision space; and uses polynomial mutation to introduce random changes to prevent premature convergence. The evolution process usually runs for 100-200 generations until the Pareto frontier stabilizes. The last stage is strategy selection and deployment, which selects a representative subset of recognition strategies (usually 5-8 strategies) from the final non-dominated solution set and deploys them to the actual recognition system. In the real-time recognition process, the environmental conditions of the current two-dimensional code (such as contamination degree, inclination angle, and illumination uniformity) are first evaluated, and then the most suitable strategy or strategy combination is selected for recognition through simple decision rules or machine learning classifiers.

[0083] S4: Based on the structured label recognition data and the knowledge graph, check the consistency of the label information and the expected connection relationship through real-time data comparison algorithm, analyze potential abnormalities, evaluate the safety impact, reliability impact and functional impact of abnormalities using multi-objective evolutionary algorithm, and generate graded warning information and targeted processing suggestions;

[0084] Step S4 builds an intelligent association layer for optical cable labels. This step realizes the dynamic association of optical cable labels and actual physical connections and abnormal warning. First, a real-time data comparison mechanism is established to match the label information identified in step S3 with the knowledge graph constructed in step S1. This comparison involves not only primary key information such as optical cable ID, but also comprehensive comparison of multiple dimensions, including basic attributes (optical cable type, specification, length, etc.), topological relationships (starting device, terminating device, port number, etc.), and logical functions (transmission signal type, communication protocol, etc.). Considering the possible naming variants or slight differences in the actual environment, a fuzzy matching algorithm is used, including character similarity calculation, semantic similarity evaluation, and structural similarity analysis. At the same time, information from multiple sources is integrated for cross-validation, including design document data, historical scanning records, and related optical cable data. To detect potential abnormalities, the effective prediction time method of reservoir computing technology is used for analysis. By learning from historical abnormal cases through RC models, various abnormal patterns can be identified, such as configuration mismatch, label error, unauthorized changes, etc., and the Lyapunov index is used to evaluate the severity and impact range of abnormalities.

[0085] Among them, the Reservoir Computing (RC) model learns historical abnormal cases, which is an innovative method combining nonlinear dynamic system theory and machine learning, designed specifically for time series anomaly detection in optical cable label management. This technology takes advantage of the computational efficiency of RC and its excellent processing capability for time series data, by analyzing historical abnormal patterns, building a prediction model, and achieving early identification and warning of potential optical cable connection abnormalities.

[0086] The core principle of the RC model is based on dynamic system theory, which projects input signals into a high-dimensional nonlinear space for processing. Its basic structure consists of three key components: input layer, reservoir, and output layer. The input layer receives the optical cable label data stream, including recognition results, scanning timestamps, operator information, etc.; the reservoir is a recurrent network containing a large number of randomly connected neurons, with "echo state" characteristics, capable of preserving historical information of input signals; the output layer extracts useful features from the reservoir state through simple linear regression to predict abnormal probabilities. Unlike traditional deep learning methods, RC models only train the output layer weights, keeping the reservoir weights fixed, significantly reducing computational complexity and training difficulty, making it very suitable for edge computing environments.

[0087] The implementation steps of the model in the optical cable label management include four main stages. The first stage is data preprocessing and feature engineering. The system collects historical optical cable label scanning records, extracts time features (such as scanning frequency, time interval change), spatial features (such as scanning position distribution, movement pattern) and content features (such as label information consistency, change frequency), and performs normalization processing. The second stage is the construction of an abnormal case library. Through historical record analysis and expert annotation, a case library containing various typical abnormalities is established, covering configuration mismatch (such as optical cable type not matching system records), label error (such as incorrect port information), unauthorized change (such as connection changes during unplanned maintenance), etc. Each type of abnormality is assigned a severity level. The third stage is RC model design and training. Based on the requirements of anomaly detection, a suitable reservoir topology is designed, usually using an echo state network (ESN) containing 500-1000 neurons, setting appropriate sparsity (about 5%) and spectral radius (usually 0.8-0.95) to ensure the network has sufficient memory capacity and dynamic complexity. Input data is injected into the RC model in time series, activating reservoir neurons, then only training the output layer weights to make the model output match the known abnormal case labels. To improve model robustness, cross-validation and regularization techniques are used to prevent overfitting. The fourth stage is anomaly prediction and Lyapunov index analysis. Real-time optical cable label data is processed by the trained RC model to generate an abnormal probability score. When the score exceeds the preset threshold, a preliminary warning is triggered. Further calculation of the Lyapunov index quantifies the instability of the system state. The index is calculated based on the divergence rate of the RC model state vector, and a higher index value indicates that the system may experience drastic changes in a short period of time, indicating the severity and potential impact of the abnormality. Based on the Lyapunov index and the type of abnormality, a hierarchical warning is automatically generated, and possible root cause analysis and recommended repair measures are provided.

[0088] When an abnormal situation is detected, a multi-objective evolutionary algorithm is used to evaluate the severity and potential impact of the abnormality from multiple perspectives, including safety impact (threat level to safety), reliability impact (impact on reliability), functional impact (range of potential functional loss), and economic impact (cost and resources required for repair). Through the multi-objective evolutionary algorithm, a balance point is sought in multiple evaluation dimensions, while considering situational factors such as load conditions, seasonal factors, and concurrent events in the current operating environment. Based on the evaluation results, hierarchical warning information is generated, including emergency warnings (serious abnormalities that require immediate handling), important warnings (abnormalities that need to be handled as soon as possible but do not require immediate response), general warnings (minor abnormalities that can be handled in routine maintenance), and attention prompts (potential risks or conditions to be observed). At the same time, based on historical handling experience and best practices, targeted handling suggestions are generated, including emergency handling schemes, temporary mitigation schemes, root cause solutions, and prevention strategies. Finally, dynamic updating of connection relationships is supported, based on abnormal handling results and actual connection states confirmed on site, using graph neural network technology to incrementally update the knowledge graph, re-evaluate the impact of network topology changes, keep the data model synchronized with the actual physical connections, and form a closed-loop management.

[0089] S5: Construct a modular system based on microservice architecture, including data parsing engine, AI generation engine, intelligent recognition engine, and knowledge graph service, support data exchange and function integration with other systems of smart substation through standardized API interface, realize intelligent management of optical cable label from generation, recognition to operation and maintenance.

[0090] Step S5 builds the optical cable label full life cycle management system architecture, which integrates the above-mentioned technical layers to build a complete system framework. A comprehensive solution including hardware and software is designed. At the hardware level, the system includes mobile terminals equipped with GPUs (supporting AI inference), high-definition cameras (with anti-shake and light supplement functions), RFID auxiliary identification modules, and label printing devices. Through edge computing technology, the system optimizes the terminal processing capability and ensures the availability of data processing and label recognition in weak network or offline environment. At the software level, the system is based on microservice architecture and develops multiple core modules: data parsing engine (combining NLP and GNN technology to process SCD / SPCD files), AI generation engine (combining GAN and reinforcement learning to optimize labels), intelligent recognition engine (combining CV and multi-modal technology for label recognition), and knowledge graph service (storing association relationships and historical data). Containerization technology is used to achieve loose coupling and flexible deployment between modules, forming a microservice architecture that supports independent development, deployment, and extension.

[0091] In terms of data management, a hierarchical data storage strategy is designed, and suitable storage solutions are adopted for different types of data such as knowledge graph, historical data, and user information, realizing the separation of hot and cold data. At the same time, a multi-level security protection mechanism is constructed, including data encryption (protecting sensitive information), access control (limiting operation permissions), and operation audit (recording system usage), etc., to ensure the security and integrity of the cable data. To support integration with other systems, the system develops standardized API interfaces to support data exchange and functional integration with other systems in the substation (such as SCADA, asset management system, etc.). In addition, a flexible plug-in mechanism is designed to allow third-party development and extension of functions, forming an open optical cable label management ecosystem. Through this modular and extensible system architecture, the invention realizes the intelligent management of the whole life cycle of optical cable labels from generation, identification to operation and maintenance, improving the efficiency and reliability of optical cable management in smart substations.

[0092] The step S1 includes the following sub-steps:

[0093] S1.1: Based on the intelligent substation configuration file, use graph neural network technology to mine the implicit connection relationship between devices, identify the indirect association of cross-bay optical cables, and establish a multi-level graph structure association network containing physical devices, ports, optical cables, and virtual terminals in the entity layer, connection relationships, transmission relationships, and ownership relationships in the relationship layer, and device models, port protocols, and optical cable core numbers in the attribute layer;

[0094] S1.2: Based on the multi-level graph structure association network, use reservoir computing technology to predict potential optical cable connection patterns based on historical connection patterns, use Lyapunov index to analyze the reliability of the prediction strategy, automatically complete the missing optical cable purpose labeling and spare core information, and obtain the completed label information;

[0095] S1.3: According to the completed label information and historical operation and maintenance data including historical operation and maintenance records, alarm logs, optical cable test data, and label use feedback, use a multi-objective evolutionary algorithm with decision space diversity to consider data integrity, consistency, and timeliness for multi-dimensional data fusion, detect and correct connection relationship conflicts, outdated information, and abnormal patterns based on the uniformity of Hamming distance measurement, and obtain the verified and optimized knowledge graph.

[0096] In particular, step S1.1 details how to construct a multi-level graph structure association network based on the smart substation configuration file. In a smart substation, there are complex connection relationships between devices, including direct visible physical connections and implicit logical associations. Traditional methods often only capture explicit direct connections, ignoring potential indirect associations, which limits the depth and breadth of optical cable management. This step innovatively introduces graph neural network (GNN) technology to deeply mine the implicit connection relationships between devices. GNN can consider both node features and topological structure, learn deep interaction between nodes through message passing mechanism, and is particularly suitable for handling complex connection relationships in substations. Through GNN, indirect associations of cross-bay optical cables can be identified, for example, when two different bay protection devices are connected through multiple optical cables, GNN can infer their logical association, even if this association is not directly represented in the original configuration file. At the same time, a multi-level graph structure is constructed based on the extracted data: the entity layer contains physical devices (such as transformers, circuit breakers, protection devices, etc.), ports (such as fiber interfaces, Ethernet ports, etc.), optical cables (trunk cables, tail cables, etc.), and virtual terminals (such as GOOSE publishing points, subscription points, etc.); the relationship layer defines the connection relationships between entities, such as "connected to", "transmission", "belongs to", etc.; the attribute layer describes the characteristics of each entity, such as the model of the device, the communication protocol of the port, the number of fiber cores of the optical cable, etc. This multi-level graph structure design can comprehensively capture the complex associations between devices in the substation, including physical connection relationships and logical functional relationships, providing a comprehensive and in-depth data view for subsequent label generation and anomaly detection.

[0097] Among them, the graph neural network (Graph Neural Networks, GNN) technology is applied to the mining of implicit connection relationships between smart substation devices, which is a cutting-edge deep learning method specially designed to handle complex association patterns in non-Euclidean data structures. In the context of substation optical cable management, GNN can break through the limitations of traditional methods and discover functional associations that are not directly labeled in the configuration file but actually exist, providing a deeper semantic understanding for the optical cable labeling system.

[0098] The core principle of GNN is based on the message passing mechanism of graph structure data. Unlike traditional neural networks that handle images or sequences, GNN is specifically designed to handle graphs composed of nodes (substation devices) and edges (connection relationships). Through an iterative message passing process, each node continuously aggregates information from neighboring nodes and updates its own representation. As the number of layers increases, nodes can obtain more and more distant neighbor information, thus capturing indirect associations of multi-hop connections. This feature makes GNN particularly suitable for discovering implicit functional associations in substations that are formed by multiple optical cable connections, such as the logical relationships between cross-bay protection devices.

[0099] The implementation steps of applying GNN in intelligent substation optical cable management include four main stages. First is graph construction and initialization. The system extracts device information, communication configuration and connection relationship from IEC 61850 configuration file (SCD file), and constructs the initial graph structure. Each physical device, communication port and optical cable is represented as a node in the graph, and the direct physical connection is represented as an edge. The initial feature vector of the node contains device type, function code, physical location and other attribute information; the feature of the edge contains connection type, medium type, etc. For communication services such as GOOSE and MMS, virtual nodes are created to represent the publishing point and subscription point, and the connection relationship with the corresponding physical device is established. The second stage is GNN model design and training. According to the characteristics of the substation, a hybrid architecture containing multiple layers of graph convolution network (GCN) and graph attention network (GAT) is adopted. GCN layer is used to capture the overall structure features, while GAT layer highlights important connections through attention mechanism, especially focusing on the optical cable paths carrying key protection signals. The model training adopts semi-supervised learning method, using known device function association as training label, and introducing topology structure preservation loss function to ensure the key topology characteristics of substation network are preserved during the learning process. The third stage is implicit connection discovery and verification. The trained GNN model identifies potential implicit functional associations by analyzing the similarity and connection patterns in the node embedding space. The system uses an algorithm based on path importance to evaluate the key degree of different optical cable paths in information transmission, and identifies those device pairs that are physically separated but functionally closely related. To verify the discovered implicit connections, the system combines electrical protection logic analysis and historical operation data to confirm the effectiveness and importance of these connections. The last stage is multi-level graph structure integration. The system integrates the discovered implicit connections into the original graph structure to construct a more complete multi-level association network. At the entity level, virtual association nodes are added to represent functional associations; at the relationship level, new edge types such as "functional dependence" and "backup path" are introduced to describe the properties of implicit connections; at the attribute level, importance scores, signal types and other attributes are added to these new associations. The integrated graph structure not only represents the physical connection, but also contains the functional dependence and information flow path, providing a comprehensive semantic background for the optical cable labeling system.

[0100] Step S1.2 uses reservoir computing technology to conduct in-depth analysis and completion of the multi-level graph structure association network. In actual application, there may be data missing or incomplete in the substation configuration file, such as the optical cable purpose not being explicitly labeled, the backup core information missing, etc., which will affect the accuracy of subsequent label generation and management. To solve this problem, this step introduces reservoir computing (RC) technology, which predicts potential optical cable connection patterns based on historical connection patterns. RC is a special type of recurrent neural network, whose core feature is a fixed, randomly initialized reservoir, and only the output layer weight is trained, which makes RC have significant advantages in training efficiency and processing time series data. First, the historical connection data is used to train the RC model, and the reservoir size and spectral radius and other hyperparameters are optimized. For the prediction of optical cable connection patterns, the historical connection data is serialized and input into the RC model, and its "memory" capability is used to capture the time dependence of the connection pattern. Using the VPT (Valid Predictive Time) characteristic of RC, the reliability of the prediction result is evaluated. VPT is a key indicator of RC, representing the longest period of time that can maintain effective prediction, and is crucial for evaluating the long-term effectiveness of the prediction result. In addition, Lyapunov index analysis is also used to dynamically adjust the prediction strategy. Lyapunov index can quantify the sensitivity to changes in initial conditions, helping to identify areas of uncertainty in the prediction, and accordingly adjust the prediction strategy to improve the accuracy of long-term prediction. In this way, missing label information can be automatically completed, such as inferring the purpose of the optical cable that is not explicitly labeled from the historical connection pattern, or inferring the possible purpose of the backup core based on the device function and topological location. This automatic completion capability significantly reduces the need for manual intervention, improves the completeness and accuracy of the knowledge graph, and provides a more reliable data foundation for generating high-quality optical cable labels.

[0101] Step S1.3 realizes multi-dimensional data fusion and consistency verification by multi-objective evolutionary algorithm of decision space diversity. In practical applications, data from different sources may be inconsistent, conflicting or outdated, and direct use of these data may lead to incorrect label information and management decisions. This step will integrate the knowledge graph constructed in the previous step with historical operation and maintenance data, and ensure data quality through strict consistency verification. The integrated data sources include historical operation and maintenance records (actual maintenance situation), alarm logs (potential problem indication), optical cable test data (such as OTDR test results), and label use feedback (such as label scanning success rate, durability, etc.). Data fusion adopts a hierarchical strategy, including data cleaning (standardizing each source data), entity alignment (identifying different expressions of the same device or optical cable in different data sources), time sequence correlation (establishing an event timeline and correlating the status changes of the same optical cable or device at different time points), and spatial correlation (combining the physical layout information of the substation to establish a spatial distribution model of the data). To ensure data consistency, a multi-objective evolutionary algorithm of decision space diversity is introduced innovatively. Unlike traditional multi-objective optimization algorithms that mainly focus on the diversity of the objective space, this algorithm focuses on the diversity of the decision space, and evaluates the distribution of candidate solutions in the decision space through the uniformity of Hamming distance. In specific implementation, multiple evaluation indexes (such as graph structure integrity, data timeliness, conflict degree, etc.) are set, and multiple candidate solutions are generated, each representing a possible data modification strategy. The Pareto optimal solution is selected by non-dominated sorting, balancing multiple optimization objectives such as data integrity, consistency, and timeliness. Through this step, connection relationship conflicts (such as mismatch between logical connections described in SCD files and physical connections recorded in SPCD files), outdated information (changed connection relationships that have not been reflected in configuration files), abnormal patterns (relationships significantly deviating from normal connection patterns), and missing data (necessary but missing connection descriptions based on context information) can be detected and corrected. Finally, this step outputs a high-quality optical cable label intelligent data model that has been verified and optimized, which not only contains accurate connection relationships, but also integrates historical operation and maintenance experience and best practices, providing a comprehensive and reliable data foundation for subsequent label generation and management.

[0102] The combination symbol quantile regression method quantifies text features, including:

[0103] For the unstructured text containing cabinet name, port description, and virtual terminal signal, use the power field professional vocabulary and rule base for word segmentation and part-of-speech tagging, perform symbolic processing to map professional terms and descriptions into symbolic representations, and generate symbolic text;

[0104] Based on the symbolic text, a quantile-based regression model is established to map text semantic features to different quantiles, analyze the correlation strength between symbols, and calculate the conditional quantiles of different features to determine the importance of symbols in the cable label and the importance weight in the cable label.

[0105] Based on the importance weight, the cable types including backbone cable, tail cable, single / multi-mode optical fiber are extracted, the connection relationship between devices including signal flow direction and logical relationship is extracted, the transmission signal types including GOOSE signal, SV sampling value, MMS message are extracted, the priority information including key protection signal, non-key monitoring signal is extracted, and the quantized text features are obtained.

[0106] Specifically, in the processing process of the intelligent substation configuration file, the processing of unstructured text is a key challenge. These texts contain cabinet names (such as "110kV arrester cabinet", "primary equipment measurement and control screen"), port descriptions (such as "optical fiber interface TX1", "optical splitter port P3") and virtual terminal signals (such as "circuit breaker trip command", "measurement value output") and the like. In order to effectively process these unstructured texts and extract valuable features, the present application innovatively combines a symbol quantile regression method to realize accurate quantization processing of text features. The specific implementation process of this method is divided into three main stages: symbolization processing, quantile regression modeling and feature extraction.

[0107] In the symbolization processing stage, the unstructured text is first accurately segmented and tagged with parts of speech using a professional lexicon and rule base specially customized for the power field. This professional lexicon contains tens of thousands of power-specific terms, covering device names (such as "circuit breaker", "disconnector", "transformer"), technical parameters (such as "rated voltage", "rated current", "protection setting value") and professional abbreviations (such as "PT" for voltage transformer, "CT" for current transformer) and the like. At the same time, the rule base contains naming patterns and expression rules specific to the power field, such as the naming method of "device number + function + parameter" (for example, "1# circuit breaker trip control circuit"). Using a professional segmenter based on these resources, complex power professional texts are accurately decomposed into the smallest semantic units. Subsequently, the segmented results are symbolized to map professional terms and descriptions into symbolic representations. For example, "220kV line protection device" is mapped to symbol "HVL_PROT", and "optical fiber splitter output port" is mapped to symbol "FOSP_OUT". This symbolization process not only standardizes the text representation, but also reduces the dimensionality and complexity of the text, facilitating subsequent mathematical modeling and analysis. In this way, the original unstructured text is converted into a symbolic sequence containing specific semantic information, forming a symbolic text.

[0108] In the quantile regression modeling stage, based on the symbolic text generated in the previous step, a quantile-based regression model is established. Unlike traditional mean regression, quantile regression can analyze different quantiles of data distribution, providing more comprehensive distribution information, especially suitable for handling data with outliers and asymmetric distribution. First, the symbolic text is converted into a feature vector, with each symbol corresponding to a dimension. Then, through the expert-labeled training data set (containing symbols and their importance scores in different cable labels), the quantile regression model is trained. This model can map text semantic features to different quantiles, for example, it can analyze the weight performance of each symbol at 25%, 50%, 75%, and 90% quantiles. Through this analysis, the performance characteristics of symbols in different situations can be obtained, avoiding the information loss that may be caused by simple mean. At the same time, the correlation strength between symbols is analyzed, and the conditional quantile model is used to consider the mutual influence between symbols. For example, when the two symbols "circuit breaker" and "trip" appear at the same time, their combined importance may be much higher than the sum of their individual importance. By calculating the conditional quantile of different features, the importance of each symbol in the cable label is determined, and finally the importance weight of each symbol in the cable label is determined. These weights reflect the relative importance of each symbol in the label information, providing a scientific basis for subsequent feature extraction.

[0109] In the feature extraction stage, based on the importance weights determined in the previous step, deep feature extraction is performed on the original text. First, extract the cable type information, including trunk cable (backbone cable connecting major devices), tail cable (short cable from trunk cable to terminal device), and single / multi-mode fiber (according to the transmission mode of the fiber). By identifying specific identifier symbols (such as "TRUNK CABLE" and the analysis symbol combination pattern, the type information of the optical cable is accurately extracted. Secondly, the connection relationship information is extracted, including the signal flow direction (sending end and receiving end) and logical relationship (such as control relationship, membership relationship) between devices. This process relies on the identification and analysis of signal flow direction indicators (such as "->", "=>") and relationship descriptors (such as "control", "belongs to"). Thirdly, the transmission signal type is extracted, including GOOSE signal (for fast event transmission between devices), SV sample value (for measurement data transmission) and MMS message (for regular data exchange). By identifying specific protocol identifiers (such as "GOOSE", "SMV", "MMS") and analyzing context information, the type of the transmitted signal is determined. Finally, the priority information is extracted, distinguishing between critical protection signals (such as trip signals, lockout signals) and non-critical monitoring signals (such as state monitoring, parameter display). By identifying keywords (such as "protection", "trip", "monitoring") and combining their importance weights, the priority of the signal is determined. Through these steps, the original unstructured text is transformed into structured and quantitative text features, providing an accurate data basis for the generation and management of optical cable labels.

[0110] This text quantification method based on symbol quantile regression has significant advantages. First, it can effectively handle professional terms and expressions specific to the power field, overcoming the limitations of general natural language processing methods in specialized fields. Second, the quantile regression model can capture the overall data distribution, not only focusing on the mean but also analyzing the characteristics of different quantiles, providing more comprehensive information. Third, by conditionally analyzing the correlation strength between symbols, important symbol combinations can be identified and strengthened, improving the accuracy of feature extraction. Fourth, the feature extraction process based on importance weights ensures the relevance and importance of the extraction results, providing a scientific basis for the design of optical cable labels. Finally, the quantified text features have good interpretability and traceability, which helps continuous optimization and knowledge accumulation. In practical applications, this method has been verified in multiple smart substation projects, with an accuracy rate of text feature extraction of over 95%, significantly improving the quality and efficiency of optical cable label generation.

[0111] The step S2 includes the following sub-steps:

[0112] S2.1: Analyze the physical space constraints, environmental factors and viewing condition characteristics of different application scenarios including narrow space of screen cabinet, outdoor environment and high-frequency reference area using the knowledge graph, learn the historical label use data through reservoir computing technology to predict the label use environment parameters, and generate an initial label template candidate set including compact label, waterproof and weather-resistant label and wear-resistant label;

[0113] S2.2: Based on the initial label template candidate set, label content data containing cable ID and connection information, and target scene parameters, a generative adversarial network technology is adopted. A generator generates a label design scheme according to the cable type and the pasting scene. A discriminator evaluates the information integrity, readability and recognition rate of the label design based on historical label samples. Content constraints are added to ensure that key information is clear and visible. Two-dimensional code constraints adjust the size, position and fault tolerance according to the use environment. Spatial constraints optimize the overall layout according to the pasting position. Visual constraints ensure the contrast between text and background. An optimized label layout scheme is generated.

[0114] S2.3: Based on the optimized label layout scheme and historical label use data containing label durability, scanning success rate, a reinforcement learning model is constructed. Lyapunov index is used to analyze the influence of parameter adjustment on label performance. The two-dimensional code encoding density, error correction level and ultraviolet resistance coding ability are adaptively optimized. The text font size, thickness and contrast are adaptively optimized. The material parameters including label thickness, adhesive type and protective layer treatment are adaptively optimized. The cable intelligent label is obtained.

[0115] Specifically, step S2.1 describes the scene-adaptive label template design process in detail. In the generation of cable labels, different application scenarios have different requirements for labels. For example, outdoor environments require labels with strong weather resistance, while narrow spaces require compact labels. To achieve precise adaptation of label design, first, based on the cable label intelligent data model output in step S1, the characteristics and constraint conditions of different application scenarios are analyzed comprehensively. These characteristics and constraints include physical space constraints (such as narrow space of screen cabinet, dense wire slot space, under-ceiling / floor wiring space, etc.), environmental factors (such as outdoor exposure environment, high temperature area, humid environment, vibration area, etc.), viewing conditions (such as low light environment, long distance identification requirement, high frequency review scene, etc.) and use frequency (such as daily inspection high frequency area, emergency handling area, regular maintenance area, etc.). For these different scenarios, labels with different parameters need to be designed to meet the actual use requirements.

[0116] To scientifically predict the parameter requirements of various scenarios, the innovative application of reservoir computing technology (RC) technology. RC is a special type of recurrent neural network with a "reservoir" structure that can efficiently process time-series data, with low computational resource requirements and a simple training process, making it particularly suitable for embedded and real-time applications. A training dataset containing various scenario features and label effectiveness evaluations was constructed, including temperature range (-40°C to +85°C), humidity (relative humidity 0% to 100%), lighting conditions (0 to 100,000 lux), spatial size (from a few millimeters to a few centimeters), etc. Label effectiveness evaluations include recognition success rate (percentage of successful recognition under different conditions), durability (from a few months to a few years), readability (degree of difficulty for human eye recognition), etc. The key hyperparameters of the RC model were optimized, including reservoir size (determines model complexity, usually hundreds to thousands of nodes), spectral radius (affects dynamic characteristics, usually set between 0.8 and 1.2), and regularization coefficient (controls the degree of overfitting, usually adjusted between 0.01 and 0.1). By adjusting these parameters, the prediction performance of the RC model can be maximized, generating the best label parameter configuration for different scenarios.

[0117] Based on the optimized RC model, the best label parameters for new scenarios are predicted, generating an initial label template candidate set. These candidate templates are specifically optimized for different scenarios, for example, for narrow spaces in screen cabinets, compact labels are generated with the smallest effective size (usually 20x40mm), the information layout is optimized to ensure that core information (such as optical cable ID, connection endpoint) is prominently displayed, while secondary information (such as detailed description) can be appropriately simplified or reduced. For outdoor environments, waterproof and weather-resistant labels are designed, with increased material thickness (usually above 0.5mm), the use of UV-resistant ink, the addition of a waterproof layer and an antioxidant coating, and the improvement of text contrast to adapt to strong light environments. For high-frequency access areas, wear-resistant labels are designed, with bold font and high-contrast color, the use of wear-resistant materials (such as polyester, polycarbonate, etc.), and the addition of a protective layer to extend the service life. For low-light environments, enhanced visibility labels are designed, using fluorescent materials or reflective elements, increasing font size, improving contrast, and optimizing the error correction level of the QR code to improve scanning success rate. Through this data-driven label template design method, the most suitable initial label design can be generated for different application scenarios, providing a good starting point for further optimization.

[0118] Step S2.2 introduces generative adversarial network technology to optimize label design. After the initial label template is generated, it needs to be optimized in depth according to specific cable information to ensure the practicality and readability of the label. To this end, based on the initial label template candidate set generated in step S2.1, combined with label content data (such as cable ID, connection information, signal type, etc.) and target scene parameters, generative adversarial network (GAN) technology is used for deep optimization. GAN consists of two parts: generator and discriminator. Through the adversarial learning of these two networks, high-quality label design schemes can be generated.

[0119] The generator network is responsible for generating label design schemes according to input conditions (such as scene parameters, cable information, etc.). Its architecture uses a deep convolutional neural network (DCNN) that includes multiple convolutional layers, batch normalization layers, and activation functions. To improve the generation effect, a conditional control mechanism is introduced, allowing the generation process to be controlled through additional inputs. For example, the generation strategy can be adjusted through a scene condition vector (encoding physical space, environmental factors, etc.), or the information layout can be controlled through a content condition vector (encoding cable type, importance, etc.). The discriminator network evaluates whether the generated label design meets the expected standards, including information integrity, readability, recognition rate, etc. The discriminator also uses a DCNN architecture, but adds an attention mechanism that can focus on key areas in the label (such as QR codes, core text information, etc.). During the training process, a large number of historical label samples and their usage evaluation are used as training data, including successful cases (high recognition rate, durable label design), failed cases (recognition difficulty, easily damaged label design), and expert-labeled label design samples (high-quality label designed and evaluated by professionals).

[0120] The innovation of GAN training lies in the addition of multiple constraints to ensure that the generated label design is both aesthetically pleasing and practical. Content constraints ensure that key information is clearly visible, with different content elements prioritized based on an information importance score (based on the weight analysis of step S1.3) to ensure that critical information (such as cable ID, connection endpoints) is highlighted, while secondary information (such as detailed descriptions) can be simplified as appropriate. The two-dimensional code constraint automatically adjusts the size, position, and error tolerance of the two-dimensional code based on the use environment, increasing the size and error tolerance level (usually Q or H level) of the two-dimensional code in high-interference environments (such as outdoors, dusty areas), and using a more compact two-dimensional code to save space in controlled environments. The space constraint optimizes the overall layout based on the available space at the sticking position, taking into account the space limitations of different installation locations (such as cable surface, port vicinity, distribution frame, etc.), and dynamically adjusting the label shape and size. The visual constraint ensures the contrast between text and background to improve readability, automatically selects the best text-background color combination to ensure good readability under various lighting conditions, and considers colorblind-friendly design. During the optimization process, the GAN will make special adjustments for different types of cable labels, such as ODF labels (simplified information display, highlighting port numbers and cable routing), inter-device connection labels (emphasizing device information on both ends, clearly indicating signal types), and tail cable labels (optimizing information layout in small sizes). Through multiple iterations and fine-tuning, the final optimized label layout scheme is generated, meeting functional requirements while balancing aesthetics and practicality.

[0121] Step S2.3 uses a reinforcement learning model to adaptively adjust label parameters. In the final stage of label design, fine-tuning of various parameters is required to adapt to the specific needs of different use environments. To this end, this sub-step innovatively introduces reinforcement learning (RL) technology to adaptively optimize key label parameters based on historical use data. The core idea of reinforcement learning is to learn the optimal strategy through interaction with the environment, which is very suitable for solving parameter optimization problems that need to balance multiple goals and constraints.

[0122] The core components of the reinforcement learning model include state space, action space, reward function, and policy network. The state space describes the current parameter configuration of the label, including two-dimensional code parameters (size, position, fault tolerance level, etc.), text parameters (font, size, color, etc.), material parameters (material type, thickness, protective layer, etc.), and scene characteristics (temperature, humidity, illumination, etc.). The action space defines possible parameter adjustment operations, such as increasing / decreasing the size of the two-dimensional code, adjusting the font size, replacing the material type, etc. The reward function evaluates the pros and cons of parameter adjustment based on the label usage effect, considering factors such as recognition success rate, durability, cost, and aesthetics. The policy network learns the optimal parameter adjustment strategy and selects the best adjustment action based on the current state. Deep Q Network (DQN) is used as the implementation method of the policy network, combined with experience replay and target network technology to improve learning stability. At the same time, the Double Deep Q Network (DDQN) architecture is introduced to reduce the over-optimistic bias of value estimation and improve learning efficiency.

[0123] Among them, Double Deep Q-Network (DDQN) is an advanced reinforcement learning architecture designed to address the value overestimation problem in traditional DQN algorithms, playing a key role in the optimization of optical cable label parameters. DDQN decouples the action selection and value evaluation processes, significantly improving the stability and efficiency of learning, making label parameter optimization more accurate and reliable.

[0124] The core principle of DDQN is based on the analysis of value estimation bias in Q-learning. In the standard DQN algorithm, the same network is responsible for both selecting actions (such as which label parameter adjustment scheme is best) and evaluating the value of that action (how much benefit that adjustment can bring). This design leads to the selection of actions that are overestimated in the presence of estimation errors, forming a positive feedback loop and ultimately leading to a serious value overestimation problem. DDQN solves this problem by introducing two functionally complementary networks: the current policy network is responsible for selecting the optimal action based on the current state, while the target network is responsible for evaluating the value of that action. This decoupling design effectively breaks the positive feedback loop and significantly reduces estimation bias.

[0125] The steps of implementing DDQN in the optical cable label optimization system include four main stages. First is the network architecture design, the system builds two deep neural networks with the same structure but independent parameters: the current network and the target network. Considering the complexity of the label parameter optimization task, the network usually adopts a multi-layer perceptron structure, the input layer receives the state vector (including the current label configuration and environmental factors), the hidden layer uses 256-512 neurons with ReLU activation function to process features, and the output layer corresponds to the Q value of each possible parameter adjustment action. To handle continuous parameters (such as precise adjustment of the size of the QR code), a hierarchical discretization strategy is introduced to discretize the continuous action space into multiple levels of fine adjustment steps. The second stage is interaction and experience collection. Different label parameter configurations are tried in the actual or simulated environment, and the four-tuple experience (current state, selected action, obtained reward, next state) is recorded after each adjustment. To speed up learning, the priority experience replay technique is used, which assigns different priorities according to the time difference error of experience, and learns more frequently from important experiences. In addition, an uncertainty-based exploration strategy is implemented, which increases the exploration probability in areas of the parameter space that have not been fully explored, ensuring comprehensive coverage of possible optimization directions. The third stage is double Q learning update. In each training batch, the current network is used to select the optimal action for the next state, but the target network is used to evaluate the value of the action. The last stage is policy evaluation and deployment. The current policy is regularly evaluated in the test environment to measure key indicators such as label recognition rate, durability, etc. under different conditions. To improve the robustness of the policy, a multi-scenario test process is specially designed, including extreme temperatures (-40°C to 85°C), high humidity (relative humidity 95%), strong light (100,000 lux), etc. The well-trained DDQN model can recommend the optimal label parameter configuration for different application scenarios, such as automatically selecting heat-resistant materials and increasing the size of the QR code module in high-temperature environments, optimizing contrast and adding anti-glare coatings in strong light environments. The Lyapunov index analysis is particularly introduced to evaluate the impact of parameter adjustment on label performance. The Lyapunov index is a measure of the sensitivity of a dynamic system to initial conditions, which can quantify the impact of parameter changes on long-term behavior. By calculating the Lyapunov index of different parameter adjustment operations, it can identify key parameters that may cause significant changes in performance, and prioritize fine-tuning of these parameters. For example, in a high-temperature environment, the size and material type of the QR code may have a higher Lyapunov index, indicating that these parameters have a significant impact on the performance of the label in a high-temperature environment, and will be prioritized for optimization.

[0126] Based on the reinforcement learning model, the key parameters are self-adaptively optimized. The QR code parameter optimization includes the encoding density (adjusted according to the information amount and space constraints, usually between 10-30mm), the error correction level (from L level to H level, selected according to the environmental interference degree) and the anti-ultraviolet encoding ability (special encoding scheme is used in outdoor environment). The text parameter optimization includes the font size (from 5pt to 12pt, adjusted according to the available space and viewing distance), the thickness (increased when the environmental interference is large) and the contrast (adjusted according to the background color and lighting conditions). The material parameter optimization includes the label thickness (from 0.1mm to 0.8mm, adjusted according to the environmental conditions and durability requirements), the adhesive type (selected according to the surface material and environmental temperature) and the protective layer treatment (such as UV protection, waterproof coating, anti-scratch coating, etc.). The reinforcement learning model will continuously adjust the parameter selection strategy according to the success and failure cases in the historical data, for example, analyzing the labels that failed in high temperature environment in the past to adjust the material selection in similar environment, fine-tuning the text and QR code parameters based on the output effect of different printing devices, optimizing the protection treatment scheme based on the aging speed of labels in different regions, etc. Through this data-driven parameter optimization, the best parameter configuration can be generated for different scenarios, significantly improving the practicality and durability of the label.

[0127] Among them, the generative adversarial network technology is used to optimize the label design, including:

[0128] A large number of historical label samples containing successful cases, failure cases and expert annotations, as well as their identification rate, good durability or difficult identification, rapid damage and use effect evaluation are used as training data. The generative adversarial network model including the generator responsible for generating label design scheme and the discriminator evaluating whether the generated label design meets the expected standard is trained. The content constraint to ensure that the key information is clear and visible, the QR code constraint to automatically adjust the size and position of the QR code according to the use environment and the fault tolerance, the space constraint to optimize the overall layout according to the available space of the pasting position and the visual constraint to ensure the contrast between the text and the background to improve the readability are added to form the GAN model under the constraint condition;

[0129] Based on the GAN model under the constraint condition, different types of labels are specially adjusted, such as ODF label that simplifies information display and highlights port number and cable direction, device-to-device connection label that emphasizes clear indication of device information and signal type, and tail cable label that optimizes information layout under small size to ensure identification. The resolution limit of the printing device, the common label material characteristics of different materials to the ink absorbability and the possible deformation in the pasting process are considered, as well as the visual changes in the aging process, to generate multiple label design schemes;

[0130] Based on the plurality of label design schemes, through multiple rounds of optimization iteration, the function requirement satisfaction label recognition and information transmission requirement, the aesthetic improvement visual effect and the practicality ensure the actual use convenience are considered at the same time, the function requirement, the aesthetic and the practicality are balanced, and the optimized label layout scheme is obtained.

[0131] Specifically, in the optical cable label design process, the traditional method often relies on experience rules and manual adjustment, which is difficult to adapt to diversified application scenarios and requirements. To solve this problem, the present application innovatively introduces the generative adversarial network (GAN) technology to intelligently optimize the label design. GAN is a powerful deep learning architecture that can generate high-quality, realistic samples through the adversarial training of generators and discriminators. Applying GAN technology in the field of label design can significantly improve design efficiency and quality, and realize intelligent generation and optimization of labels.

[0132] Firstly, a rich and diverse training dataset is constructed to provide a learning basis for the GAN model. This dataset contains three types of key samples: successful cases, failed cases and expert annotated samples. Successful cases refer to label designs that perform well in actual use, with high recognition rate (recognition success rate in various environmental conditions exceeds 95%) and good durability (service life in standard environment exceeds 5 years). These cases usually have clear information layout, appropriate contrast and optimized two-dimensional code design. Thousands of successful cases from different substations and different application scenarios are collected, and their design parameters (such as font size, color configuration, two-dimensional code size, etc.) and performance indicators (such as recognition rate, service life, etc.) are recorded in detail. Failed cases are label designs that have problems in actual application, such as recognition difficulty (recognition success rate is less than 60%) or rapid damage (service life is less than 50% of the expected value). These cases usually have design defects, such as insufficient contrast, inappropriate two-dimensional code size or inappropriate material selection. These failure cases are collected, and their failure reasons and design defects are analyzed to provide negative learning samples for the model. Expert annotated samples are high-quality label samples designed and evaluated by industry experts. These samples not only contain design parameters, but also have detailed comments and improvement suggestions from experts. This kind of sample is particularly valuable, as it can provide the model with the experience and design philosophy of human experts.

[0133] Based on these rich training data, an advanced GAN model is constructed, including two core components: generator and discriminator. The generator adopts a deep convolutional neural network architecture, composed of multiple convolutional layers, batch normalization layers, and nonlinear activation functions. Its input includes conditional vectors (encoding label content, target scene, etc.) and random noise vectors (providing generation diversity), and the output is a complete label design scheme, including layout, color scheme, font settings, etc. The key innovation of the generator lies in its conditional control mechanism, which precisely controls the generation process by introducing multiple conditional vectors. These conditional vectors include scene condition vectors (encoding physical space, environmental factors, etc., such as "small space", "outdoor environment", etc.), content condition vectors (encoding cable types, importance, etc., such as "main cable", "key protection signal", etc.), and style condition vectors (encoding design style preferences, such as "simple style", "high contrast style", etc.). The discriminator also adopts a deep convolutional neural network structure, but adds an attention mechanism that can focus on key areas in the label. The input of the discriminator is the label design sample (real sample or generator-generated sample) and the corresponding conditional vector, and the output is the authenticity score of the sample and the prediction of various performance aspects (such as recognition rate, durability, etc.). The discriminator not only needs to distinguish between real samples and generated samples, but also needs to evaluate whether the generated samples meet various design standards and performance requirements.

[0134] In the training process of the GAN model, multiple constraints are introduced innovatively to ensure that the generated label design is both beautiful and practical. Content constraints ensure that key information is clear and visible, which is the most basic requirement in label design. Based on the information importance analysis in step S1.3, the label content is classified, such as first-level information (such as cable ID, connection endpoint, which must be the most prominent), second-level information (such as signal type, cable type, which needs to be clear and visible), and third-level information (such as detailed description, note information, which can be simplified appropriately). According to this classification, different attention weights are set in the generator to ensure that important information gets more "design resources" (such as larger font, more prominent position). The two-dimensional code constraint automatically adjusts the size, position and error tolerance of the two-dimensional code according to the use environment. In high interference environments (such as outdoor, dusty areas), the size of the two-dimensional code is increased (usually more than 25mm x 25mm) and the error tolerance level is increased (Q or H level, corresponding to 25% or 30% error correction capability); while in controlled environments, more compact two-dimensional codes (such as 15mm x 15mm) can be used to save space while reducing error tolerance (such as L level, corresponding to 7% error correction capability). The space constraint optimizes the overall layout according to the available space of the pasting position. Consider the space limitations of different installation positions, such as cable surface (usually elongated shape, width limited), port vicinity (usually small square area), wiring rack (usually standardized label slot), etc. According to these limitations, the shape and size of the label are dynamically adjusted to ensure that the label fits the target installation position. The visual constraint ensures the contrast between the text and the background to improve readability. The best text-background color combination is automatically selected by calculating the color contrast index (according to the W3C accessibility guidelines, the contrast between text and background should be at least 4.5:1) to ensure good readability under various lighting conditions. At the same time, colorblind-friendly design is also considered to avoid using red-green and other easily confused color combinations.

[0135] In the label design optimization process, special adjustments are made for different types of cable labels to meet their respective special needs. For ODF (Optical Distribution Frame) labels, since they usually need to provide port identification information in a small space, a strategy of simplifying information display is adopted. This type of label highlights the port number (usually using large bold font, occupying more than 30% of the label area) and the cable routing (using clear arrows or text description to indicate the start and end positions of the cable), while simplifying or reducing other information (such as detailed device description). At the same time, considering that labels in ODF environment are usually densely arranged, the side readability of the label is particularly optimized to ensure that key information can be clearly identified from different angles. For inter-device connection labels, since they need to clearly indicate the relationship between the two devices, they emphasize the two device information (equally highlighting the start and end device names) and the signal type (using color coding or icons to represent different types of signals such as GOOSE, SMV, etc.). This type of label usually adopts a symmetrical layout, with the two connected devices displayed on both sides of the label and the signal type and other connection details displayed in the middle. To improve information density, while ensuring readability, clever use of hierarchical layout and color coding is made to display more information in limited space. For tail cable labels, since they are usually small in size and need to be attached to thin cables, the information layout in small size is particularly optimized. This type of label adopts an efficient space utilization strategy, such as vertical arrangement of text (extending along the cable direction), use of abbreviations and codes (such as "SM" for single-mode fiber), optimization of font selection (selecting a font that remains clear in small size), etc. At the same time, the two-dimensional code readability of the tail cable label is enhanced, by optimizing the size, position and contrast of the two-dimensional code, to ensure that even in the case of limited label size, the two-dimensional code can still be reliably identified.

[0136] In generating the label design, various limiting factors in practical application are also fully considered. First, the resolution limitation of the printing device. Different printing devices have different resolution capabilities, from low-resolution thermal printers (such as 300 dpi) to high-resolution laser printers (such as 1200 dpi). The font size (ensuring that the minimum font is still recognizable at the target resolution, usually not less than 6pt), line thickness (ensuring that thin lines will not break after printing), and QR code density (ensuring that each module contains at least 3x3 printing points) will be adjusted according to the resolution capability of the target printing device. Second, the material characteristics of the label. Different materials have different ink absorption, durability, and applicable environments. For example, ordinary paper labels are low-cost but poor in durability, suitable for indoor protected environments; polyester (PET) labels are water-resistant and wear-resistant but more expensive, suitable for outdoor or harsh environments; polyvinyl chloride (PVC) labels have good flexibility, suitable for sticking on curved surfaces. According to the target use environment and durability requirements, the appropriate material is selected, and the design parameters are adjusted accordingly, such as increasing the line thickness on materials with strong ink absorption and adding background color on transparent materials. Third, the possible deformation during the sticking process, especially when the label needs to be stuck on a curved surface (such as the surface of an optical cable) or an irregular surface. By simulating different degrees of bending and stretching (usually considering a deformation range of ±10%), the robustness of the design is evaluated to ensure that even after deformation, the key information is still identifiable. Finally, visual changes during the aging process are also considered, such as color fading (simulated by reducing color saturation), contrast reduction (simulated by reducing light and dark contrast), and wear (simulated by adding noise and blur effects). Through these considerations, a variety of label design schemes that perform well in practical applications can be generated.

[0137] Finally, based on the generated multiple label design schemes, through multiple rounds of optimization iteration, the final optimized label layout scheme is obtained by balancing functional requirements, aesthetics, and practicality. In this process, three key aspects are considered simultaneously: functional requirements, aesthetics, and practicality. Functional requirements are the most basic consideration, and the label must meet the identification and information transmission requirements. Evaluate the information integrity (whether all necessary information is included), readability (whether the text is clear and visible), and identification rate (the success rate of QR code identification under various conditions) of each design scheme. Aesthetics, although not the most important consideration, good visual effects can improve user experience and work efficiency. Evaluate the overall balance, color coordination, and professional appearance of the design to ensure that the label not only has perfect functionality but also has a pleasing visual effect. Practicality focuses on the convenience of the label in actual use, including installation convenience (whether it is easy to accurately stick), search efficiency (whether it is easy to find the target label among multiple labels), and durability (whether it can maintain functionality in the expected environment for a long time).

[0138] The optimization process adopts an iterative method. After each iteration, the design parameters are adjusted based on the simulation test results and historical data comparison. For example, if a design scheme performs poorly in the recognition rate test, the size of the QR code or the contrast ratio will be increased; if a design scores low in the usability evaluation, the layout will be adjusted to improve the intuitiveness. Through 5-10 rounds of optimization iteration, the best design scheme that balances various needs can be found. The final output of the optimized label layout scheme not only meets various technical requirements, but also considers aesthetics and practicality, and can provide excellent performance and user experience in actual application. The actual application results show that through this GAN-based label design optimization method, the recognition success rate of the label is increased by 15-20%, the service life is prolonged by more than 30%, the user satisfaction is significantly improved, and the difficulty and maintenance cost of optical cable management in the transformer substation are greatly reduced.

[0139] The conversion into structured label recognition data containing optical cable types, connection relationships and transmission signal types in step S3 includes:

[0140] S3.1: Based on the optical cable intelligent label, the label image collected by the mobile terminal is used to acquire environmental supplementary data in the presence of label contamination, line-of-sight obstruction and high-density optical cable areas, and to record the intensity of light, the angle of collection and the distance of collection in synchronization. Image enhancement algorithms such as adaptive histogram equalization, non-local mean filtering and super-resolution reconstruction are applied, and geometric correction algorithms and light compensation algorithms are applied to generate standardized label images with uniform brightness, contrast, viewing angle and resolution;

[0141] S3.2: Based on the standardized label image, the QR code area is located and extracted through gradient analysis, morphological verification and positioning point detection. The super-resolution reconstruction technology of residual learning network combined with perception loss function is used to process low-resolution or blurred QR codes. The edge repair technology of structural integrity analysis, context perception filling and topological structure preservation is used to process partially worn or broken QR codes. The attention mechanism of spatial attention focusing on effective areas, channel attention dynamically adjusting feature weights and multi-scale fusion capturing local and global features is used. Through adaptive binarization, perspective correction and error correction enhancement decoding, QR code recognition is achieved, and the QR code recognition result is obtained;

[0142] S3.3: Based on the two-dimensional code recognition result, introduce a multi-objective evolutionary algorithm with decision space diversity that encodes the recognition process key parameters and strategies into decision vectors, generate an initial solution set with diversity based on the uniformity of Hamming distance, and obtain multiple complementary recognition strategies through double-layer selection of Pareto advantage and decision space diversity. At the same time, the optical character recognition of the text information on the label is carried out, the multi-modal fusion technology of information consistency verification, complementary information integration and redundant information utilization is used to fuse the two-dimensional code information and the recognition result of the text, and through the natural language processing technology, the standardized data structure is converted to obtain the structured label recognition data containing the cable type, connection relationship and transmission signal type.

[0143] Specifically, step S3.1 describes the multi-modal data acquisition and preprocessing process in detail. In the recognition of optical cable labels, it is necessary to first obtain high-quality raw data and carry out effective preprocessing to provide good input for subsequent recognition algorithms. This sub-step realizes efficient acquisition and enhancement of label data through various sensing devices and preprocessing technologies. First, the high-definition camera mounted on the mobile terminal (such as industrial tablet or smart phone) is used to capture label images. These cameras have a variety of advanced features, including high resolution (at least 12 million pixels, supporting 4K image acquisition), automatic focusing function (ensuring clear imaging at various distances), optical image stabilization function (reducing blur caused by hand shaking), and fill light (LED fill light, coping with low light environment). During the acquisition process, exposure parameters, white balance and color saturation are automatically adjusted to adapt to different lighting conditions, ensuring image quality.

[0144] In addition to image data, rich environmental supplementary data is also collected to improve the accuracy and robustness of recognition. These data include light intensity (measured by light sensor, unit: lux), acquisition angle (measured by gyroscope and accelerometer, recording horizontal and vertical offset angles), acquisition distance (estimated by laser ranging or depth camera, unit: centimeters), and environmental temperature and humidity (collected by temperature and humidity sensor, affecting image quality and two-dimensional code reflection characteristics). In addition, an RFID auxiliary recognition module is integrated, especially for recognition assistance in label contamination, line-of-sight obstruction or high-density optical cable areas. RFID technology provides a visual-independent label recognition method that can penetrate through slight obstacles and provide an alternative recognition approach in difficult visual recognition situations. The RFID module used works at 13.56 MHz frequency band, with a reading distance of up to 10 centimeters. The RFID chip embedded in the label is designed with anti-metal interference to ensure stable reading in metal environments.

[0145] After obtaining the raw data, a series of preprocessing steps are performed to convert the raw images collected under various conditions into standardized high-quality label images. First is image enhancement, including adaptive histogram equalization (enhancing local contrast, especially in low-light and overexposed environments), non-local mean filtering (removing noise while preserving details), and super-resolution reconstruction (improving image resolution through deep learning models, especially for small-sized labels collected at a distance). Then is geometric correction, including perspective transformation (correcting distortion caused by inclination angles), distortion correction (eliminating barrel or pincushion distortion caused by camera lenses), and size standardization (adjusting label images of different sizes to a uniform size for subsequent processing). Next is illumination compensation, including illumination model estimation (analyzing the illumination distribution in the image), reflection component separation (distinguishing between diffuse reflection and specular reflection on the label surface), and shadow removal (eliminating contrast unevenness caused by partial shadows). Finally is region positioning, including label detection (locating the label region in a complex background), interest region extraction (separating the QR code and text regions), and multi-label separation (handling multiple labels that may be contained in the same image). Through these preprocessing steps, high-quality standardized label images can be extracted from raw images collected under various non-ideal conditions, providing ideal input data for subsequent QR code recognition and text recognition.

[0146] Step S3.2 implements super-resolution enhancement and repair of the QR code region of the optical cable label. In actual applications, the QR code region may have poor image quality due to various reasons (such as long distance, poor light, partial wear, etc.), affecting recognition accuracy. This sub-step uses advanced computer vision technology to enhance and repair the QR code region, significantly improving recognition ability under various conditions. First, based on the standardized image after step S3.1 preprocessing, the QR code region is located and extracted. The positioning uses a cascade detector combining Haar features and AdaBoost algorithms to quickly and accurately identify the QR code region in the image. For complex backgrounds or partial occlusion, a deep learning target detection model (such as FasterR-CNN or YOLO) is introduced to improve the robustness of positioning. The extracted QR code region is precisely cropped and corner corrected to ensure correct alignment of the QR code matrix.

[0147] For low-resolution or blurred QR codes, a super-resolution reconstruction technique using a residual learning network combined with a perceptual loss function is applied. This technique is based on the Deep Residual Network (ResNet) architecture, which learns the residual mapping between low-resolution and high-resolution images to achieve high-quality super-resolution reconstruction. Unlike traditional methods that only use pixel-level MSE loss, a perceptual loss function is adopted, combining pixel-level loss, feature-level loss (difference in deep features extracted by a pre-trained VGG network), and adversarial loss (evaluating the authenticity of the reconstructed image by a discriminator network) to optimize visual quality and detail preservation. Special training datasets and network structures are customized for the special structure of QR codes (regular black and white module arrangement), including data augmentation strategies for QR codes (simulating various degradation conditions) and structure-aware residual block design (preserving the geometric structure of QR codes). Through this super-resolution reconstruction, the original resolution can be increased by 2-4 times, significantly improving the clarity and recognizability of QR codes.

[0148] Among them, the Deep Residual Network (ResNet) is a deep learning architecture that performs exceptionally well in the task of cable label QR code super-resolution reconstruction. It solves the gradient vanishing problem in deep neural network training through an innovative residual learning paradigm, enabling the construction of ultra-deep network structures to capture multi-scale features in images, making it particularly suitable for processing images such as QR codes that require accurate recovery of high-frequency details.

[0149] The core principle of ResNet is based on the idea of residual learning. Traditional convolutional neural networks directly learn the mapping function H(x) from input to output, while ResNet learns the residual mapping F(x) = H(x) - x, where the actual output of the network is F(x) + x. This design adds the input directly to the output through identity shortcut connections, forming a residual block structure. In the task of QR code super-resolution reconstruction, this mechanism is particularly important as it allows the network to focus on learning the difference information between low-resolution and high-resolution images, rather than completely reconstructing the entire image, significantly improving learning efficiency and accuracy, especially in restoring key identification features such as the edges and corners of QR codes.

[0150] The steps of implementing ResNet in the optical cable label QR code super-resolution system include four main stages. First is the network architecture design, which builds a super-resolution residual network (SR-ResNet) optimized for QR codes. The input layer receives low-resolution QR code images, which are processed through initial feature extraction convolution layers and then enter the main body of residual learning, composed of 16-20 residual blocks in series. Each residual block contains two 3×3 convolution layers, with a ReLU activation function in between, and the block's output is directly added to the input through a skip connection at the end. To adapt to the structural characteristics of QR codes, an attention mechanism module is introduced into the standard residual block to help the network focus on key areas such as positioning patterns and arrangement patterns. After feature extraction, the resolution is increased by 2×, 3×, or 4× through a pixel shuffle layer or transposed convolution, and finally, a reconstruction convolution layer generates the high-resolution output. The second stage is the construction and enhancement of specific datasets. For the optical cable label application scenario, a large number of high-quality QR codes are collected as high-resolution samples, and low-resolution samples are generated through a controlled degradation process. The degradation process simulates various quality degradation scenarios that actual optical cable labels may encounter under different conditions, including Gaussian blur (simulating unclear focus), motion blur (simulating jitter during scanning), JPEG compression (simulating digital transmission loss), and noise addition (simulating low-light conditions). At the same time, specific data enhancement strategies are designed, such as random occlusion (simulating partial label wear), brightness and contrast changes (simulating different lighting conditions), and geometric transformations (simulating different viewing angles). These data enhancement techniques significantly improve the model's generalization ability, enabling it to adapt to various scanning conditions in the field. The third stage is the design and training of a multi-objective loss function. To comprehensively optimize the visual quality and machine readability of QR codes, a triple loss function combination is developed. The pixel-level loss uses the L1 norm, which better preserves the clear edges of the QR code compared to the L2 norm; the feature-level loss extracts deep features through a pre-trained VGG-19 network and calculates the difference between the reconstructed image and the original high-resolution image in the feature space, focusing on the structural consistency of the QR code; the adversarial loss introduces a discriminator network to evaluate the authenticity of the reconstructed image, prompting the generative network to produce outputs closer to the true high-resolution image. In addition, a structure preservation loss term is designed to ensure the accurate reconstruction of key information such as positioning patterns and arrangement patterns. The training uses the Adam optimizer with an initial learning rate of 1e-4 and a cosine annealing strategy, a batch size of 16, and is performed on a server with 4 NVIDIA V100 GPUs, which takes an average of 48 hours to converge. The final stage is model optimization and deployment. To adapt to the computational resource limitations of edge devices, the trained model is compressed and optimized, including knowledge distillation (transferring the knowledge of a large teacher network to a smaller student network), pruning (removing unimportant connections), and quantization (converting 32-bit floating-point parameters to 8-bit integers).The optimized model size is reduced by 78%, and the inference speed is increased by 3.5 times, which can run in real time on mobile devices such as smartphones (each frame processing time is less than 150ms). In actual tests, it can improve the recognition rate of severely degraded two-dimensional codes from the original 37% to more than 94%, and even in extreme conditions (such as part of the optical cable label is worn, severely contaminated or insufficient light), it can maintain an identification success rate of more than 80%, significantly improving the reliability and user experience of the optical cable management system. This ResNet-based super-resolution reconstruction technology not only solves the limitations of traditional methods in processing low-quality two-dimensional codes, but also provides strong image recovery capabilities for optical cable label systems, ensuring stable operation in various harsh environments, and becomes a key technology pillar of modern optical cable management systems.

[0151] For partially worn or broken two-dimensional codes, edge repair techniques such as structural integrity analysis, context-aware filling, and topology preservation are used. First, the damage degree and location of the two-dimensional code are evaluated through structural integrity analysis, including missing module detection (identifying completely missing black and white modules), edge damage detection (identifying fuzzy or broken module edges), and pollution area identification (identifying areas covered by stains). Then, according to the structural characteristics and existing information of the two-dimensional code, context-aware filling is performed. This includes Markov Random Field (MRF) based module state inference (inferred missing module state based on surrounding module state), constraint repair based on two-dimensional code generation rules (using two-dimensional code check bits and format information for repair), and region completion based on deep generative models (using GAN or VAE models to generate context-compliant filling content). Finally, topology preservation techniques ensure that the repaired two-dimensional code maintains the correct geometric structure and connectivity, including edge enhancement (improving the clarity of module boundaries), shape regularization (ensuring that each module is approximately square), and global structure alignment (ensuring the regular arrangement of the overall two-dimensional code matrix).

[0152] To further improve the recognition capability under extreme conditions, a multi-objective evolutionary algorithm with decision space diversity is innovatively introduced. The algorithm encodes the key parameters and strategies in the recognition process into a decision vector, including preprocessing parameters (such as filtering strength, contrast enhancement degree), reconstruction parameters (such as super-resolution factor, regularization strength) and decoding parameters (such as binarization threshold, fault tolerance strategy). Then, the diversity of the decision space is evaluated by the uniformity measure based on Hamming distance, ensuring that the generated candidate solutions are uniformly distributed in the parameter space, covering different recognition strategies. At the same time, multiple objectives are optimized, including recognition accuracy, computational efficiency and robustness, and the Pareto optimal solution set is found by the Non-dominated Sorting Genetic Algorithm II (NSGA-II). In practical applications, the most suitable recognition method for the current situation will be dynamically selected or combined according to the real-time detected environmental conditions (such as light intensity, image quality, etc.). In this way, high recognition success rate can be maintained under various extreme conditions (such as severe contamination, strong light interference, long-distance shooting, etc.), significantly improving the adaptability and reliability.

[0153] Step S3.3 realizes the structured extraction of label information through multi-modal fusion technology. After completing the recognition of the two-dimensional code, the text information on the label needs to be recognized and the two-dimensional code and text information need to be fused to form complete structured label data. This sub-step realizes high-precision extraction and integration of label information through advanced OCR technology and multi-modal fusion method. First, the text area on the label is recognized by optical character recognition (OCR). An OCR model specially customized for the power industry is used. This model has been trained on a large number of transformer station optical cable label texts and can accurately recognize power industry terminology and special formats. The OCR processing flow includes text area detection (locating the text area through algorithms such as MSER or TextBoxes++), character segmentation (segmenting continuous text into individual characters, considering different fonts and spacing), feature extraction (using CNN to extract deep features of characters), and character recognition (determining the identity of each character through a deep learning classifier). To improve recognition accuracy, a context-aware language model is introduced, which uses language rules and common expressions in the power industry to correct possible recognition errors. For example, when "blocker" is recognized, the language model will infer that the correct word should be "blocker" based on the context.

[0154] After completing the individual recognition of the QR code and the text, the information is integrated into a unified structured data through multi-modal fusion technology. Multi-modal fusion adopts a hierarchical strategy, including feature-level fusion (combining low-level features of different modalities), decision-level fusion (integrating recognition results of different modalities), and semantic-level fusion (integrating semantic information from different sources). Feature-level fusion assigns different weights to features of different modalities through attention mechanism, dynamically adjusting the importance of each modality according to the current scene conditions. For example, in the case where the QR code is clear but the text is blurred, the weight of the QR code information is increased; while in the case where the text is clear but the QR code is partially damaged, the weight of the text information is increased. Decision-level fusion uses weighted voting or Bayesian decision framework to consider the confidence of each modality recognition result, and selects the most reliable result. For conflicting recognition results, conflicts are resolved based on prior knowledge (such as common error patterns) and current conditions (such as image quality of each modality). Semantic-level fusion focuses on the semantic consistency and completeness of information, ensuring that the fused data is logically reasonable and complete.

[0155] To improve the fusion effect, information complementarity and cross-validation mechanism are introduced. Information complementarity mechanism identifies unique information provided by different modalities, ensuring that the fusion result contains all available information. For example, the QR code may contain detailed unique identifiers and digital codes, while the text label may provide more intuitive device names and connection information. Cross-validation mechanism uses redundant information between different modalities to verify each other, improving data reliability. For example, the device ID in the QR code is compared with the device name mentioned in the text label to check for consistency. Through such multi-modal fusion, the advantages of various information sources can be fully utilized, maximizing the accuracy and completeness of label recognition. Finally, the fused recognition results are converted into structured label recognition data containing cable type (such as main cable, tail cable, single / multi-mode fiber, etc.), connection relationship (signal flow direction and logical relationship between devices), and transmission signal type (such as GOOSE signal, SV sample value, MMS message, etc.). These structured data directly support subsequent data comparison and anomaly detection, providing a solid data foundation for intelligent management of optical cable labels.

[0156] Among them, based on the optical cable intelligent label, the label image of the on-site optical cable is collected, and the multi-modal fusion computer vision technology is used to convert into structured label recognition data containing optical cable type, connection relationship and transmission signal type, which further includes uniformity measurement enhancement decision space diversity based on Hamming distance:

[0157] The key parameters and strategies in the recognition process are coded into decision vectors, the Hamming distance between each decision vector in the solution set is calculated, and a density-based diversity evaluation function is designed to take the diversity index as an additional optimization target;

[0158] Based on the diversity evaluation function, an initial solution set with diversity is generated, new candidate solutions are generated by mutation and crossover to explore the decision space, and the double-layer selection based on the Pareto advantage and the diversity of the decision space is carried out, the diversity weight is dynamically adjusted according to the convergence in the optimization process, the learning rate of different parameters is adjusted according to the sensitivity, and the adaptive adjustment mechanism of dynamic balance between exploration and utilization is obtained, and a variety of complementary identification strategies are obtained.

[0159] Based on the plurality of complementary identification strategies, in the strong light interference environment, the strategy paying more attention to light compensation is selected, in the serious pollution condition, the more aggressive image repair strategy is selected, in the limited computing resource, the simplified strategy with higher computing efficiency is selected, the most suitable identification method for the current situation is dynamically selected or combined according to the real-time detected environmental conditions, and the diversity of the decision space is enhanced.

[0160] Specifically, in the intelligent substation, the identification of optical cable labels faces various complex challenges, such as harsh environmental conditions, label aging and damage, poor imaging quality and the like. The traditional single identification method is often difficult to adapt to these variable scenes. In order to solve this problem, the present application innovatively introduces a uniformity measurement method based on Hamming distance, enhances the diversity of the decision space, and constructs a robust identification system that can adapt to various complex conditions. This method not only improves the success rate of identification, but also intelligently selects the best strategy according to different environmental conditions, greatly improving the overall performance of the system.

[0161] First, various key parameters and strategies in the identification process are encoded into structured decision vectors, laying the foundation for diversity evaluation and optimization. These decision vectors cover multiple key aspects of the identification process. Preprocessing parameters include filter method selection (such as Gaussian filtering, median filtering, bilateral filtering, etc.), filter intensity (usually adjusted between 0.5-5, the larger the value, the stronger the filtering effect), contrast enhancement method (such as histogram equalization, CLAHE adaptive histogram equalization, Gamma correction, etc.), contrast enhancement degree (usually expressed as a percentage, such as 30% enhancement) and edge enhancement parameters (such as sharpening coefficient, usually between 0.1-2). Reconstruction parameters include super-resolution factor (determines the magnification, usually 2x, 3x or 4x), super-resolution algorithm selection (such as SRCNN based on convolutional neural network, SRGAN based on generative adversarial network, etc.), regularization intensity (controls the smoothness in the reconstruction process, usually 0.001-0.1), iteration number (determines the complexity of the reconstruction process, usually 10-100 times) and feature preservation preference (such as edge preservation preference, texture preservation preference, etc.). Decoding parameters include binarization method (such as global thresholding, adaptive thresholding, Otsu method, etc.), binarization threshold (determines the black-white dividing point, usually an integer between 0-255), error correction code usage strategy, damaged area repair strategy, pattern recognition sensitivity (determines the strictness of the identification module) and check bit usage strategy (such as enabling all check bits, using only key check bits, etc.). These parameters are encoded into binary decision vectors, with each bit representing a parameter option or value range, for example, 8 bits can represent integer values between 0-255, and 2 bits can represent 4 different algorithm selections.

[0162] To evaluate the diversity of the decision space, an innovative diversity evaluation function based on Hamming distance is designed. Hamming distance is an indicator to measure the difference between two equal-length strings, defined as the number of different characters at corresponding positions. The larger the Hamming distance, the greater the difference between the two decision vectors, representing the more different strategies they adopt. The designed diversity evaluation function considers the Hamming distance distribution between each decision vector in the solution set, calculating the "crowding degree" of each solution in the decision space.

[0163] Based on the above diversity evaluation function, a complete multi-objective evolutionary algorithm process is implemented to generate and optimize a diverse set of recognition strategies. First, by controlling the parameter distribution during the generation process, an initial solution set with high initial diversity is constructed. This initialization does not use completely random generation, but uses techniques such as Latin Hypercube Sampling (LHS) to ensure that the initial solutions are uniformly distributed in the decision space. The initial solution set usually contains 100-200 different decision vectors, covering different parameter combinations and strategy choices. Next, the system generates new candidate solutions through mutation and crossover operations to explore more possibilities in the decision space. Mutation operations include bit flipping (randomly changing some bits in the decision vector, usually with a probability of 0.01-0.05), parameter adjustment (targeted adjustment according to parameter sensitivity), and strategy replacement (replace the strategy of a module as a whole, such as replacing the preprocessing method). Crossover operation generates offspring solutions with mixed characteristics by combining the advantages of different parent solutions. Methods such as gene fragment exchange (exchange parameter fragments between two decision vectors), uniform crossover (independently decide whether to exchange each bit according to a certain probability), and intelligent crossover (selectively exchange parameters according to the performance of parent solutions in different scenarios) are used.

[0164] Among them, Latin Hypercube Sampling (LHS) is an advanced statistical sampling technique used in optical cable label recognition systems to generate high-quality initial strategy sets. It can achieve uniform distribution of samples in high-dimensional parameter space, overcoming the defects of traditional random sampling methods that are prone to sample clustering or blank areas, providing a more comprehensive starting point for evolutionary algorithms, thereby accelerating convergence and improving the quality of the final solution.

[0165] The core principle of LHS is based on the mathematical concept of Latin square, which achieves efficient coverage of multi-dimensional space by ensuring uniform distribution of projections in each dimension. When generating m sample points in an n-dimensional parameter space, LHS divides each dimension equally into m intervals, then ensures that there is only one sample point in each interval of each dimension. This design ensures the uniform distribution of samples in the entire parameter space, while avoiding the dimension disaster problem, so that even in high-dimensional space, good space coverage can be achieved with relatively few sample numbers. For optical cable label recognition systems, this means that various parameter combinations (such as image processing parameters, feature extraction thresholds, classifier configurations, etc.) can be explored with the smallest number of initial strategies, greatly improving search efficiency.

[0166] The steps of implementing LHS in the optical cable label identification system include five main stages. The first stage is parameter space definition and boundary setting. The key parameters in the identification process are comprehensively analyzed, including preprocessing parameters (Gaussian filter kernel size, adaptive threshold window size, morphological operation structure element shape, etc.), feature extraction parameters (HOG descriptor unit size, SIFT feature point number threshold, CNN feature layer selection, etc.), and classification decision parameters (SVM kernel function type, random forest tree number, deep network activation function selection, etc.). Each parameter is set with reasonable upper and lower bounds according to professional knowledge, forming a parameter hyperspace with 25-30 dimensions, providing an accurate search range for LHS. The second stage is interval division and initialization. 160 initial solutions are generated (experience shows that this number can provide a good balance between diversity and computational efficiency under the complexity of the problem), and then the value range of each parameter is equally divided into 160 intervals. For continuous parameters (such as filter σ value), direct numerical division is performed; for discrete parameters (such as CNN layer selection), they are mapped to continuous space and then divided, and after sampling, they are mapped back to discrete values. To improve sampling quality, the system also implements minimum distance constraints to ensure that the Euclidean distance between any two sample points in the normalized space is not less than a pre-set threshold (usually 0.05), avoiding wasting computational resources due to too close sample points. The third stage is sampling matrix generation and optimization. First, create a 160x30 (sample number x dimension number) sampling matrix, where each column represents a parameter dimension. By randomly arranging the sequence of 1-160 and normalizing it to the [0,1] interval, the sampling point position in this dimension is generated. This process ensures that each interval of each parameter dimension has exactly one sampling point. To further optimize the sample distribution, the system implements maximum and minimum distance optimization: through column exchange operations, repeatedly optimize the minimum distance between sample points, and maximize the spatial dispersion of the entire sample set. This optimization process usually performs 500-1000 iterations until the minimum distance no longer increases significantly or reaches the computational time limit. The fourth stage is parameter conversion and strategy construction. Convert the optimized normalized sampling matrix back to the actual parameter value: for continuous parameters, map to the actual range through linear interpolation; for discrete parameters, use the nearest neighbor rounding method to determine the final value. In particular, for some parameter combinations that have dependent relationships (such as some feature extraction methods only applicable to certain pre-processed images), the system implements conditional constraint processing to ensure that the generated parameter combination is executable in the actual system. The converted parameter vector is directly used to construct a complete identification strategy, including image acquisition, preprocessing, feature extraction, classification decision, and post-processing of the complete process configuration. The last stage is diversity verification and adjustment.The diversity evaluation index of the initial solution set is calculated to verify the sampling effect: the coverage index measures the coverage ratio of the parameter space, usually requiring more than 95%; the dispersion index measures the uniformity of the spatial distribution of the sample points, and evaluates the statistical distribution of the nearest neighbor distance; the diversity index comprehensively considers the distribution of the strategy in the action space, and calculates the performance difference on the test data set. If the diversity index does not meet the expectation (usually requiring more than 85% of the action space coverage), local adjustment will be made: identify the low-coverage area in the parameter space, and increase the targeted sampling points; or remove similar strategies and replace them with new strategies in the low-coverage area. In the selection stage, an innovative double-layer selection strategy is implemented. The first layer is based on the Pareto advantage for selection, and the non-dominated solution (i.e. the solution that is not dominated by other solutions in all objectives) is retained. The fast non-dominated sorting algorithm (Fast Non-dominated Sorting) is used to divide the solution set into multiple levels, and the solutions with higher levels are preferentially selected. The solutions in the Pareto front represent the best trade-off in multiple objectives such as recognition accuracy, computational efficiency and robustness. The second layer is based on the diversity of the decision space for selection, and the solutions with sparse distribution in the decision space are preferentially retained, i.e. the solutions with large differences from other solutions. The crowding degree of each solution is calculated, and the solutions with low crowding degree (i.e. high diversity) are preferentially selected. This double-layer selection ensures that the solution set contains high-performance solutions and maintains sufficient diversity, avoiding premature convergence of the solution set to a single strategy. To further improve the adaptability of the algorithm, multiple adaptive adjustment mechanisms are introduced. The diversity weight is dynamically adjusted according to the convergence in the optimization process, and the diversity weight is increased in the early stage to promote extensive exploration, and the diversity weight is decreased in the later stage to promote meticulous optimization. The learning rate is adjusted according to the sensitivity of different parameters, and a smaller learning rate is used for key parameters (such as binary threshold, super-resolution factor, etc.) with greater impact to make fine adjustments, and a larger learning rate is used for parameters with smaller impact to speed up convergence. A dynamic balance is maintained between exploration and utilization, and the mutation rate and crossover rate are adaptively adjusted according to the performance and diversity state of the current solution set, and the exploration intensity is increased when the performance improvement is slow, and the utilization intensity is increased when potential areas are found. Through these strategies, a set of diverse and high-performance complementary recognition strategies can be obtained, providing rich choices for subsequent dynamic selection.

[0167] The fast non-dominated sorting algorithm (Fast Non-dominated Sorting) is a core algorithm in the field of multi-objective optimization, which is used to efficiently evaluate and stratify the strategy set in the optical cable label recognition system. By systematically comparing the dominance relationship between solutions, all candidate solutions are organized into different dominance levels, providing clear selection guidance for evolutionary algorithms, and balancing the trade-off between multiple conflicting objectives (such as recognition accuracy, computational complexity and environmental adaptability).

[0168] The core principle of this algorithm is based on the concept of Pareto dominance. For a multi-objective optimization problem, solution A is said to dominate solution B if A is better than B in at least one objective and not worse than B in all other objectives. Non-dominated solutions are those that are not dominated by any other solution in the set, and these solutions form the Pareto front, representing the optimal trade-offs between objectives. The fast and elitist non-dominated sorting algorithm efficiently divides the entire solution set into multiple front levels by calculating two key attributes for each solution: the domination count (how many solutions it dominates) and the dominated set (which solutions it is dominated by). The first front level contains all non-dominated solutions, the second front level contains non-dominated solutions after removing the first front level, and so on, forming a complete hierarchical structure.

[0169] The implementation of fast and elitist non-dominated sorting in the optical cable label identification system consists of three main stages. The first stage is the domination relationship calculation, which compares each pair of solutions in the candidate solution set and evaluates their performance in three key objectives: identification accuracy (the proportion of correct identifications on the standard test set), computational efficiency (the average time required to complete identification), and environmental adaptability (a measure of performance stability under different conditions). For each solution p, two sets are maintained: Sp (the set of solutions dominated by p) and np (the number of solutions that dominate p). The time complexity of the algorithm is O(MN 2 ), where M is the number of objectives (usually 3-5) and N is the number of solutions (usually 100-200). The second stage is front assignment, which divides the solution set into layers based on the calculated domination relationships. First, identify the first front F1, which contains all solutions with np = 0 (i.e., solutions that are not dominated by any solution). Then, for each solution p in F1, the system iterates through each solution q in its dominated set Sp, reducing q's domination count by 1; when a solution q's domination count becomes 0, it is added to the next front F2. Repeat this process until all solutions are assigned to the corresponding front. In the optical cable label system, usually 4-6 front levels are formed, with the first front containing 20-30 optimal identification strategies representing the performance boundary that the system can achieve. The third stage is rank assignment and selection, which assigns each solution a rank value based on its front level (the smaller the rank value, the higher the quality of the solution). In the selection operation, solutions with smaller rank values are preferred for the next generation. If it is necessary to select some solutions from the same front (for example, when the size of the current front exceeds the number of selectable solutions), further apply the crowding distance calculation to prefer solutions with smaller crowding distances (i.e., more isolated in the objective space) to maintain the diversity of the solution set. In practical applications, all solutions in the first front are retained, and from the second front onwards, the crowding distance selection is applied until the pre-set population size (usually 50-70% of the original population) is reached.

[0170] Based on the above-optimized multiple complementary recognition strategies, an environment-adaptive dynamic strategy selection is realized, which significantly improves the robustness and accuracy of optical cable label recognition. It can intelligently select or combine the most suitable recognition strategy according to different environmental conditions. In strong light interference environment, when it detects that there are high-light areas, strong light spots or obvious brightness unevenness in the image, it will automatically select a strategy that pays more attention to light compensation. Such strategies usually include high dynamic range processing (combining multiple images with different exposures to capture more details), light equalization (reducing the impact of over-bright and over-dark areas), and specular reflection removal (identifying and eliminating high light reflection on the label surface). For example, when it detects that more than 10% of the image is overexposed, it will enable a special highlight suppression algorithm to reduce the brightness of these areas while improving the visibility of shadow areas, making the black and white module contrast of the QR code clearer. In the case of severe contamination, when the system detects that there are obvious stains, scratches or missing areas on the label, it will select a more aggressive image repair strategy. Such strategies include high-intensity denoising (such as non-local mean filtering, which can effectively remove speckle noise), edge enhancement reconstruction (reconstructing fuzzy or broken module boundaries), and region completion based on deep learning (using a generative model to fill in missing areas). For example, when it detects that more than 20% of the QR code is contaminated, it will enable a special deep learning repair model that has been trained on a large number of contaminated QR codes and can effectively recover obscured or damaged information. When computing resources are limited, such as running on a handheld device or needing to process a large number of labels in real time, the system will select a simplified strategy with higher computational efficiency. Such strategies usually include lightweight preprocessing (such as using simple Gaussian filtering instead of complex non-local mean filtering), low-complexity super-resolution (such as using fast bicubic interpolation instead of deep learning super-resolution), and optimized decoding algorithms (such as simplified binarization methods and fast pattern recognition). These simplified strategies may have slightly lower performance in extreme conditions, but they can provide good performance in most ordinary situations while significantly reducing computing time and resource consumption.

[0171] The core innovation lies in the ability to detect environmental conditions in real-time and dynamically select or combine the most suitable recognition methods. The system assesses current environmental conditions in real-time through various sensors and image analysis techniques, including light intensity analysis (evaluating lighting conditions through image histograms and brightness distribution), dirt detection (evaluating label cleanliness through texture analysis and edge detection), and resource monitoring (monitoring CPU utilization, memory usage, and battery status of the current device). Based on these real-time data, a combination of rule-based decision trees and machine learning is employed to select the most suitable strategy for the current situation. In some complex cases, multiple strategies may be combined, for example, in the case of uneven lighting and slight dirt, moderate light compensation and mild image repair may be applied simultaneously. A strategy evaluation and adaptive adjustment mechanism is also implemented, recording the results and performance indicators of each recognition, constantly updating and optimizing the decision model for strategy selection. For example, if a certain strategy is found to have a lower success rate than expected under certain conditions, the weight of that strategy in similar conditions will be reduced; conversely, if a strategy performs well, its weight will be increased. In this way, the system can continuously learn and adapt to different environmental conditions and label states, achieving continuous optimization of recognition ability.

[0172] This method of uniformity measurement based on Hamming distance enhances the diversity of decision space, providing significant performance improvement for optical cable label recognition. Compared with traditional single strategy methods, the average recognition success rate is increased by more than 25%, especially in poor conditions, the recognition success rate in strong light environment is increased by 35%, and in severe dirt conditions, it is increased by 40%. At the same time, the adaptability and robustness are greatly enhanced, which can handle various complex and extreme situations such as severely faded labels, partially occluded labels, and labels under strong light. This method not only improves the accuracy of recognition, but also reduces the maintenance cost and the need for manual intervention, providing reliable technical support for the optical cable management of intelligent substations. Practical application shows that this method has achieved good results in multiple intelligent substations of different sizes and types, significantly improving the operation and maintenance efficiency and safety, and has received high praise from front-line workers.

[0173] The step S4 comprises the following sub-steps:

[0174] S4.1: Based on the structured label recognition data and the knowledge graph, the topological relationship and logical function are matched and compared through multi-dimensional feature matching, the fuzzy matching strategy of character similarity calculation, semantic similarity evaluation and structure similarity analysis is adopted, multi-source data collaborative verification is carried out by integrating design document data, historical scanning records and related optical cable data, the effective prediction time method of reservoir calculation technology is used to learn historical abnormal cases and configure abnormal modes including mismatch, label error, unauthorized change and potential risk and calculate Lyapunov index to evaluate abnormal influence time scale and range, detect the case that the physical connection does not match the design drawing, and obtain the abnormal detection result;

[0175] S4.2: Based on the abnormal detection result and the label operation and maintenance historical data containing fault history, maintenance record and performance trend, the severity and potential influence of the abnormality are evaluated from multiple angles including security threat degree, reliability influence, function loss range evaluation of function influence and repair cost and resource evaluation of economic influence, a balance point is found in multiple evaluation dimensions, and the situational sensitivity of load condition, seasonal factor and concurrent event is considered, the hierarchical warning information of emergency warning, important warning, general warning and attention prompt is generated according to the severity of the evaluation, the case reasoning of similar abnormal cases is searched in the historical case library and the rule reasoning of field expert rules is applied, and the targeted treatment suggestion including emergency treatment scheme, temporary relief scheme, root solution scheme and prevention strategy is provided;

[0176] S4.3: Based on the execution result of the targeted treatment suggestion and the actual connection state confirmed on site, the change source identification and multi-level change confirmation process including plan change, on-site discovery and fault handling are established, the graph neural network technology is used to add, modify or delete entity nodes and connection relationships affected by changes using incremental update strategy, the direct influence, indirect influence and redundancy evaluation of network topology change are re-evaluated, the connection relationship data in the knowledge graph is dynamically updated and the change version history is maintained, and dynamic association and abnormal warning are realized.

[0177] The step S5 includes the following sub-steps:

[0178] S5.1: Design a hardware system including a mobile terminal equipped with a GPU supporting AI reasoning, a high-definition camera with anti-shake and light supplementing functions, an RFID identification module for auxiliary identification, and a label printing device supporting multiple materials, optimize the terminal processing capability to support AI reasoning and image processing through edge computing technology, ensure the availability of data processing and label recognition in weak network or offline environment, and obtain a usable hardware system;

[0179] S5.2: Based on the available hardware system, develop core modules including a data parsing engine for processing the intelligent substation configuration file using NLP and GNN technologies, an AI generation engine for label optimization generation using GAN and reinforcement learning technologies, an intelligent recognition engine for multimodal parsing using CV and multimodal technologies, and a knowledge graph service for storing relationships and historical data. Containerization technology is used to achieve loose coupling and flexible deployment between modules, forming a microservice architecture that supports independent development, deployment, and expansion.

[0180] S5.3: Based on the microservice architecture, design a layered data storage strategy for different types of data to achieve separation of hot and cold data, build a multi-level security mechanism including data encryption to protect sensitive information, access control to restrict operation permissions, and operation audit records for system use, develop standardized API interfaces to support data exchange and function integration with other substation systems such as SCADA system and asset management system, design a flexible plug-in mechanism that allows third parties to develop extended functions, and realize a full lifecycle management system.

[0181] like Figure 2 As shown, the present invention also provides a full lifecycle management system for optical fiber smart tags, comprising:

[0182] The data parsing module 601 is used to extract structured data, including descriptions of logical nodes, logical devices, communication connections, and physical devices, based on the intelligent substation configuration file, which includes intelligent substation SCD files and SPCD files, through an XML parsing engine. It uses natural language processing technology to perform semantic analysis on unstructured text, combines the symbolic quantile regression method to quantify text features, and constructs a knowledge graph containing the relationship between devices, ports, optical cables, and virtual terminals.

[0183] The tag generation module 602 is used to dynamically adjust the text size, QR code position and fault tolerance according to the optical cable type and pasting scenario based on the knowledge graph and generative adversarial network technology to generate optical cable smart tags adapted to different scenarios.

[0184] The tag recognition module 603 is used to collect tag images of the optical cable on site based on the optical cable smart tag, and use multimodal fusion computer vision technology to convert them into structured tag recognition data containing optical cable type, connection relationship and transmission signal type;

[0185] The anomaly warning module 604 is used to rely on the structured tag recognition data and the knowledge graph to check the consistency between the tag information and the expected connection relationship through a real-time data comparison algorithm, analyze potential anomalies, evaluate the safety impact, reliability impact and functional impact of the anomalies using a multi-objective evolutionary algorithm, and generate graded warning information and targeted handling suggestions.

[0186] The system integration module 605 is used for constructing a modular system based on a micro-service architecture, and includes a data analysis engine, an AI generation engine, an intelligent identification engine and a knowledge graph service. The system integration module 605 supports data exchange and function integration with other systems of the smart substation through a standardized API interface, and realizes intelligent management of the optical cable label from generation, identification to operation and maintenance.

[0187] It should be noted that the above specific embodiments are only exemplary descriptions of the present application, not limitations of the present application. Those skilled in the art can make many forms of modifications and changes under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims, and these all belong to the protection scope of the present application.

Claims

1. A method for full lifecycle management of optical fiber smart tags, characterized in that, Includes the following steps: Based on the intelligent substation configuration file, which includes SCD and SPCD files, a structured data including logical nodes, logical devices, communication connections, and physical device descriptions is extracted using an XML parsing engine. Natural language processing (NLP) techniques are used to perform semantic analysis on the unstructured text, and the text features are quantified using the symbolic quantile regression method. A knowledge graph containing the relationships between devices, ports, optical cables, and virtual terminals is then constructed. Specifically, the quantification of text features using the symbolic quantile regression method includes: segmenting and tagging the unstructured text containing cabinet names, port descriptions, and virtual terminal signals using a power industry-specific thesaurus and rule base, and then performing symbolic processing to represent the professional... Terms and descriptions are mapped to symbolic representations to generate symbolic text. Based on the symbolic text, a quantile-based regression model is established to map the semantic features of the text to different quantiles, analyze the correlation strength between symbols, calculate the conditional quantiles of different features to determine the importance of symbols in optical cable tags, and determine the importance weights in optical cable tags. Based on the importance weights, optical cable types including trunk optical cables, pigtail cables, and single-mode / multimode optical fibers are extracted; the connection relationships of signal flow and logical relationships between devices are extracted; the transmission signal types including GOOSE signals, SV sample values, and MMS messages are extracted; and priority information including critical protection signals and non-critical monitoring signals is extracted to obtain quantized text features. Based on the knowledge graph, generative adversarial network (GAN) technology is used to dynamically adjust the text size, QR code position, and fault tolerance rate according to the fiber optic cable type and pasting scenario to generate smart fiber optic cable labels adapted to different scenarios. Specifically, the GAN technology is used to optimize the label design, including: using a large number of historical label samples containing successful cases, failed cases, and expert annotations, along with corresponding usage effect evaluations, as training data to train a GAN model that includes a generator responsible for generating label design schemes and a discriminator evaluating whether the generated label designs meet expected standards. Content constraints are added to ensure key information is clearly visible; QR code constraints automatically adjust the QR code size, position, and fault tolerance rate according to the usage environment; spatial constraints optimize the overall layout based on the available space at the pasting location; and visual constraints ensure text-background contrast to improve readability, forming a GAN model under these constraints. Based on the aforementioned smart optical cable tags, tag images of the optical cables on site are collected, and multimodal fusion computer vision technology is used to convert them into structured tag recognition data containing optical cable type, connection relationship and transmission signal type. Based on the structured tag recognition data and the knowledge graph, the consistency between tag information and expected connection relationships is checked through real-time data comparison algorithms, potential anomalies are analyzed, and the security impact, reliability impact, and functional impact of anomalies are evaluated using multi-objective evolutionary algorithms to generate graded early warning information and targeted handling suggestions. A modular system based on a microservice architecture is constructed, including a data parsing engine, an AI generation engine, an intelligent recognition engine, and a knowledge graph service. Through standardized API interfaces, it supports data exchange and functional integration with other systems in smart substations, realizing intelligent management of the entire lifecycle of optical cable tags from generation and recognition to operation and maintenance.

2. The method according to claim 1, characterized in that, The intelligent substation configuration file, based on intelligent substation SCD and SPCD files, extracts structured data including descriptions of logical nodes, logical devices, communication connections, and physical devices using an XML parsing engine. Natural language processing techniques are used to perform semantic analysis on the unstructured text, and symbolic quantile regression is used to quantify text features. This constructs a knowledge graph containing relationships between devices, ports, optical cables, and virtual terminals, including: Based on the intelligent substation configuration file, graph neural network technology is used to mine the implicit connection relationships between devices, identify the indirect associations across optical cables, and establish a multi-level graph structure association network with an entity layer containing physical devices, ports, optical cables and virtual terminals, a relationship layer containing connection relationships, transmission relationships and attribution relationships, and an attribute layer containing device models, port protocols and the number of optical fiber cores. By using reservoir computing technology to predict potential optical cable connection patterns based on historical connection patterns, and employing Lyapunov index analysis to dynamically adjust the reliability of the prediction strategy, missing optical cable usage labels and spare fiber core information are automatically completed to obtain the completed label information. Based on the completed tag information and historical operation and maintenance data including historical operation and maintenance records, alarm logs, optical cable test data and tag usage feedback, a multi-objective evolutionary algorithm with diverse decision space is used to perform multi-dimensional data fusion while considering data integrity, consistency and timeliness. Based on the uniformity metric of Hamming distance, connection relationship conflicts, outdated information and abnormal patterns are detected and corrected to obtain the verified and optimized knowledge graph.

3. The method according to claim 1, characterized in that, The generation of fiber optic smart tags adapted to different scenarios includes: The knowledge graph is used to analyze the physical space constraints, environmental factors and viewing conditions of different application scenarios, including narrow cabinet spaces, outdoor environments and high-frequency viewing areas. The reservoir computing technology is used to learn from historical tag usage data to predict tag usage environment parameters and generate an initial tag template candidate set including compact tags, waterproof and weather-resistant tags and wear-resistant tags. Based on the initial candidate set of label templates, label content data containing fiber optic cable ID and connection information, and target scene parameters, generative adversarial network technology is used to generate label design schemes according to fiber optic cable type and pasting scene. A discriminator evaluates the information integrity, readability and recognition rate of the label design based on historical label samples. Content constraints are added to ensure that key information is clearly visible, QR code constraints are added to adjust the size, position and error tolerance according to the usage environment, spatial constraints are added to optimize the overall layout according to the pasting position, and visual constraints are added to ensure the contrast between text and background, thus generating an optimized label layout scheme. Based on the optimized label layout scheme and historical label usage data including label durability and scanning success rate, a reinforcement learning model is constructed. The Lyapunov index is used to analyze the impact of parameter adjustments on label performance, adaptively optimizing QR code encoding density, error correction level, and UV resistance encoding capability, adaptively optimizing text font size, thickness, and contrast, and adaptively optimizing material parameters including label thickness, adhesive type, and protective layer processing, to obtain the optical cable smart label.

4. The method according to claim 3, characterized in that, Generative adversarial network (GAN) technology is used to optimize label design, including: Based on the GAN model under the aforementioned constraints, special adjustments were made to different types of ODF tags that simplify information display and highlight port numbers and optical cable routes, device connection tags that emphasize information at both ends of the equipment and clearly indicate signal types, and tail cable tags that optimize information layout under small size to ensure recognizability. The resolution limitations of printing equipment, the characteristics of commonly used label materials with different ink absorption properties, possible deformation during the pasting process, and visual changes during the aging process were all taken into account to generate a variety of label design schemes. Based on the various label design schemes, through multiple rounds of optimization and iteration, the optimized label layout scheme is obtained by simultaneously considering functional requirements to meet label recognition and information transmission requirements, aesthetic enhancement to improve visual effects, and practicality to ensure ease of use.

5. The method according to claim 1, characterized in that, The conversion into structured tag identification data, which includes fiber optic cable type, connection relationship, and transmission signal type, includes: Based on the aforementioned optical cable smart tag, the tag image is collected by a mobile terminal, combined with supplementary environmental data obtained in areas with tag damage, obstructed view, and high-density optical cable, and the light intensity, collection angle, and collection distance are recorded simultaneously. The image enhancement algorithms of adaptive histogram equalization, nonlocal mean filtering, and super-resolution reconstruction are applied, and the geometric correction algorithm and illumination compensation algorithm are applied to generate a standardized tag image with uniform brightness, contrast, viewing angle, and resolution. Based on the standardized label image, the QR code region is located and extracted through gradient analysis, morphological verification, and localization point detection. A super-resolution reconstruction technique combining residual learning network and perceptual loss function is used to process low-resolution or blurry QR codes. Edge repair techniques that preserve structural integrity, context-aware filling, and topology are used to process partially worn or broken QR codes. An attention mechanism that combines spatial attention to focus on the effective region, channel attention to dynamically adjust feature weights, and multi-scale fusion to capture local and global features is used. QR code recognition is achieved through adaptive binarization, perspective correction, and error correction enhancement decoding to obtain the QR code recognition result. Based on the QR code recognition results, a multi-objective evolutionary algorithm with decision space diversity is introduced, which encodes key parameters and strategies of the recognition process into decision vectors. An initial solution set with diversity is generated based on the uniformity metric of Hamming distance. Multiple complementary recognition strategies are obtained through a two-layer selection process using Pareto advantage and decision space diversity. Simultaneously, optical character recognition is performed on the text information on the label image. A multi-modal fusion technique, employing information consistency verification, complementary information integration, and redundant information utilization, is used to fuse the QR code information with the text recognition results. Natural language processing technology is then used to convert this data into a standardized data structure, resulting in structured label recognition data containing information about the optical cable type, connection relationship, and transmission signal type.

6. The method according to claim 5, characterized in that, The process of acquiring tag images of the optical cables based on the smart tags, and converting them into structured tag recognition data containing optical cable type, connection relationship, and transmission signal type using multimodal fusion computer vision technology, also includes: The key parameters and strategies in the identification process are encoded into decision vectors, the Hamming distance between each decision vector in the solution set is calculated, and a density-based diversity evaluation function is designed to use the diversity index as an additional optimization objective. Based on the diversity evaluation function, an initial solution set with diversity is generated. New candidate solutions are generated through mutation and crossover to explore the decision space. A two-layer selection is performed based on Pareto advantage and then based on the diversity of the decision space. The diversity weight is dynamically adjusted according to the convergence of the optimization process, the learning rate is adjusted according to the sensitivity of different parameters, and an adaptive adjustment mechanism is used to dynamically balance exploration and utilization to obtain a variety of complementary recognition strategies. Based on the aforementioned complementary recognition strategies, a strategy that focuses more on illumination compensation is selected in strong light interference environments, a more aggressive image restoration strategy is selected in severe damage situations, and a simplified strategy with higher computational efficiency is selected when computational resources are limited. The recognition method most suitable for the current situation is dynamically selected or combined according to the environmental conditions detected in real time, thereby enhancing the diversity of the decision space.

7. The method according to claim 1, characterized in that, The process involves using the structured tag recognition data and the knowledge graph, employing a real-time data comparison algorithm to check the consistency between tag information and expected connection relationships, analyzing potential anomalies, and utilizing a multi-objective evolutionary algorithm to assess the security, reliability, and functional impacts of the anomalies, generating tiered early warning information and targeted handling suggestions, including: Based on the structured label recognition data and the knowledge graph, multi-dimensional feature matching is used to compare topological relationships and logical functions. A fuzzy matching strategy combining character similarity calculation, semantic similarity evaluation, and structural similarity analysis is adopted. Design document data, historical scanning records, and related optical cable data are integrated for multi-source data collaborative verification. The effective prediction time method of reservoir computing technology is used to learn from historical anomaly cases and configure anomaly patterns including mismatch, label error, unauthorized changes, and potential risks. The Lyapunov index is calculated to evaluate the time scale and scope of anomaly impact. The system detects discrepancies between physical connections and design drawings and obtains anomaly detection results. Based on the anomaly detection results and tagged operation and maintenance history data including fault history, maintenance records, and performance trends, a multi-objective evolutionary algorithm is used to assess the severity and potential impact of the anomaly from multiple perspectives, including the degree of security threat, reliability impact, functional impact (assessing the scope of functional loss), and economic impact (assessing repair costs and resources). A balance is sought across multiple assessment dimensions, taking into account load conditions, seasonal factors, and the contextual sensitivity of concurrent events. Based on the severity of the assessment, tiered warning information is generated, including emergency warnings, important warnings, general warnings, and attention prompts. Case reasoning based on similar anomalies is retrieved from a historical case library, and rule reasoning based on application domain expert rules is used to provide targeted handling suggestions, including emergency response plans, temporary mitigation plans, fundamental solutions, and prevention strategies. Based on the execution results of the targeted handling suggestions and the actual connection status confirmed on-site, a change source identification and multi-level change confirmation process is established, including planned changes, on-site discovery and fault handling. Graph neural network technology is used to add, modify or delete entity nodes and connection relationships affected by changes using an incremental update strategy. The direct impact, indirect impact and redundancy assessment of network topology changes are re-evaluated. The connection relationship data in the knowledge graph is dynamically updated and the change version history is maintained to achieve dynamic association and anomaly early warning.

8. The method according to claim 1, characterized in that, The modular system built on a microservice architecture includes a data parsing engine, an AI generation engine, an intelligent recognition engine, and a knowledge graph service. It supports data exchange and functional integration with other systems in the smart substation through standardized API interfaces, enabling intelligent management of the entire lifecycle of optical fiber tags from generation and recognition to operation and maintenance. The design includes a mobile terminal equipped with a GPU to support AI inference, a high-definition camera with image stabilization and fill light, an RFID identification module for auxiliary identification, and a hardware system that supports label printing equipment for various materials. By using edge computing technology to optimize the terminal's processing capabilities to support AI inference and image processing, the design ensures the availability of data processing and label identification in weak network or offline environments, resulting in a usable hardware system. Based on the available hardware system, a core module is developed, including a data parsing engine for processing the configuration files of the smart substation using NLP and GNN technologies, an AI generation engine for label optimization and generation using GAN and reinforcement learning technologies, an intelligent recognition engine for multimodal parsing using CV and multimodal technologies, and a knowledge graph service for storing relationships and historical data. Containerization technology is used to achieve loose coupling and flexible deployment between modules, forming a microservice architecture that supports independent development, deployment, and expansion. Based on the aforementioned microservice architecture, a layered data storage strategy is designed for different types of data to achieve separation of hot and cold data. A multi-level security mechanism is constructed, including data encryption to protect sensitive information, access control to restrict operation permissions, and operation audit records for system use. Standardized API interfaces are developed to support data exchange and functional integration with other substation systems such as SCADA systems and asset management systems. A flexible plug-in mechanism that allows third parties to develop extended functions is designed to achieve full lifecycle management.

9. A full lifecycle management system for optical fiber smart tags, characterized in that, include: The data parsing module is used to extract structured data, including logical nodes, logical devices, communication connections, and physical device descriptions, from the intelligent substation configuration files (including SCD and SPCD files) using an XML parsing engine. It then performs semantic analysis on the unstructured text using natural language processing techniques and quantifies text features using symbolic quantile regression to construct a knowledge graph containing relationships between devices, ports, optical cables, and virtual terminals. Specifically, the quantification of text features using symbolic quantile regression includes: segmenting and tagging the unstructured text containing cabinet names, port descriptions, and virtual terminal signals using a power industry-specific thesaurus and rule base, and performing symbolic processing. The method involves mapping technical terms and descriptions to symbolic representations to generate symbolic text. Based on this symbolic text, a quantile-based regression model is established to map the semantic features of the text to different quantiles. The correlation strength between symbols is analyzed, and the conditional quantiles of different features are calculated to determine the importance of the symbols in the optical cable tag and to determine the importance weight in the optical cable tag. Based on the importance weight, the optical cable types, including trunk optical cables, pigtail cables, and single-mode / multimode optical fibers, are extracted. The connection relationships of signal flow and logical relationships between devices are extracted. The transmission signal types, including GOOSE signals, SV sample values, and MMS messages, are extracted. Priority information, including critical protection signals and non-critical monitoring signals, is extracted to obtain the quantified text features. The tag generation module is used to dynamically adjust the text size, QR code position, and fault tolerance based on the knowledge graph and the fiber optic cable type and pasting scenario using generative adversarial network (GAN) technology to generate smart fiber optic cable tags adapted to different scenarios. Specifically, the tag design optimization using GAN technology includes: using a large number of historical tag samples containing successful cases, failed cases, and expert annotations, along with corresponding usage effect evaluations, as training data to train a GAN model that includes a generator responsible for generating tag design schemes and a discriminator evaluating whether the generated tag designs meet expected standards. This incorporates content constraints to ensure key information is clearly visible, QR code constraints to automatically adjust QR code size, position, and fault tolerance based on the usage environment, spatial constraints to optimize the overall layout based on available space at the pasting location, and visual constraints to ensure text-background contrast and improve readability, forming a GAN model under these constraints. The tag recognition module is used to collect tag images of the optical cable on site based on the optical cable smart tag, and use multimodal fusion computer vision technology to convert them into structured tag recognition data containing optical cable type, connection relationship and transmission signal type; The anomaly warning module is used to check the consistency between the tag information and the expected connection relationship by relying on the structured tag recognition data and the knowledge graph through real-time data comparison algorithm, analyze potential anomalies, evaluate the safety impact, reliability impact and functional impact of the anomalies using multi-objective evolution algorithm, and generate graded warning information and targeted handling suggestions. The system integration module is used to build a modular system based on a microservice architecture, including a data parsing engine, an AI generation engine, an intelligent recognition engine, and a knowledge graph service. It supports data exchange and functional integration with other systems in smart substations through standardized API interfaces, enabling intelligent management of the entire lifecycle of optical cable tags from generation and recognition to operation and maintenance.

Images