Multi-mode power grid operation, maintenance and first-aid repair intelligent auxiliary system based on domestic large model

Abstract

The invention relates to the technical field of power grid operation inspection and repair, and discloses a multi-modal power grid operation inspection and repair intelligent auxiliary system based on a localized large model, which comprises the steps of: constructing the localized large model by collecting power grid professional data and basic-level operation scene data, dynamically balancing professional precision and deploying lightweight, and establishing a multi-modal power grid operation inspection and repair intelligent auxiliary system; in combination with equipment and personnel operation characteristics of base-level operation scene data, collecting multi-modal data and performing time sequence calibration and environment noise reduction processing; a three-level safety protection system is constructed by combining a power grid APN / GRE special communication channel, and data safety and operation real-time requirements are balanced through dynamic encryption parameter adjustment; the method comprises the following steps: constructing intelligent assistance of three scenes by combining a large domestic model and processed multi-modal data, establishing a system optimization evaluation model through operation effect data and personnel feedback data, and realizing continuous adaptation of the system with basic-level operation scenes and personnel skills by dynamically adjusting model parameters and function configuration.

Description

Technical Field

[0001] This invention relates to the field of power grid operation, maintenance and emergency repair technology, specifically to a multimodal intelligent auxiliary system for power grid operation, maintenance and emergency repair based on a domestically developed large-scale model. Background Technology

[0002] The current intelligent auxiliary system for power grid operation and maintenance acquires equipment status information through drone inspections, infrared thermal imaging, and acoustic fingerprint detection. It then combines knowledge graphs and expert systems to generate emergency repair strategies, enabling rapid data processing and lightweight model deployment.

[0003] In existing technologies, general-purpose large models lack a deep understanding of power system topology and equipment characteristics, making it difficult to construct a complete power knowledge graph. This hinders accurate performance of core numerical analyses of power systems, such as power flow calculations and short-circuit current calculations, impacting fault location accuracy. Furthermore, they struggle to generate power optimization schemes incorporating specialized mathematical models like second-order cone relaxation, limiting advanced decision support capabilities. They cannot accurately represent the hybrid ring-and-radial topology of the power grid, leading to incorrect assessments of the fault impact range. Limited understanding of the dynamic characteristics of power flow in the power system makes it difficult to predict power flow transfers and cascading effects after a fault. The lack of cross-equipment fault correlation analysis and the absence of intelligent generation capabilities for power grid recovery strategies after a complete grid outage threaten the system's safety baseline.

[0004] Therefore, there is a need to provide a multimodal intelligent auxiliary system for power grid operation, maintenance and emergency repair based on a domestically developed large-scale model. Summary of the Invention

[0005] The purpose of this invention is to provide a multimodal intelligent auxiliary system for power grid operation, maintenance, and emergency repair based on a domestically developed large-scale model. To solve the aforementioned problems in the prior art, this invention achieves this through the following technical solution:

[0006] The first part, the multimodal power grid operation and maintenance intelligent auxiliary system based on a domestically developed large model provided by the embodiments of the present invention, specifically includes the following modules:

[0007] Data processing module: By collecting power grid professional data and grassroots operation scenario data, it adopts hierarchical knowledge distillation and professional knowledge graph alignment technology to build a domestic large model, dynamically balance professional accuracy and lightweight deployment, and collect multimodal data and perform time-series calibration and environmental noise reduction processing based on the equipment and personnel operation characteristics of grassroots operation scenario data.

[0008] Protection Adjustment Module: Combined with the power grid APN / GRE dedicated communication channel, a three-level security protection system of identity authentication, data encryption, and transmission monitoring is constructed. By dynamically adjusting encryption parameters, the system balances data security with the real-time requirements of operations.

[0009] Intelligent Assistance Module: Combining domestically produced large-scale models with processed multimodal data, it constructs intelligent assistance for three major scenarios: skills learning, on-site assistance, and remote collaboration. Through work performance data and personnel feedback data, it establishes a system optimization and evaluation model. By dynamically adjusting model parameters and function configurations, it achieves continuous adaptation of the system to grassroots work scenarios and personnel skills.

[0010] Furthermore, the method for constructing the domestically produced large-scale model is as follows:

[0011] Collect hardware parameters of commonly used terminals at the grassroots level and establish a library of various terminal adaptation models;

[0012] Industry standards and failure cases are transformed into a four-level knowledge graph of equipment type, failure characteristics, troubleshooting steps, and safety requirements, with entity relationships marked. In particular, high-frequency failures at the grassroots level are marked with enhanced annotations to avoid knowledge redundancy in the general large model. Grassroots high-frequency failures include, but are not limited to, line icing and transformer oil leakage.

[0013] Based on the domestically developed large-scale model, the upper layer distills general language understanding and multimodal recognition capabilities, the middle layer injects power grid professional knowledge graphs, and the lower layer optimizes the model structure for the hardware parameters of the grassroots terminals and removes redundant parameters.

[0014] Furthermore, the method for optimizing the model structure is as follows:

[0015] The total number of original parameters of the Guangming large model is obtained, and the total number of parameters of the module is dynamically and lightweightly calculated by combining the power grid professional knowledge. The optimized total number of model parameters is obtained by analysis, and the model complexity and terminal adaptability are dynamically balanced.

[0016] Furthermore, the method for performing timing calibration is as follows:

[0017] The smart safety helmet uses a built-in microphone to collect voice commands from on-site personnel and soundprints of equipment operation, and supports offline collection.

[0018] Equipment operating data is collected by vibration sensors, temperature sensors, and oil chromatography sensors deployed on the equipment;

[0019] The smart safety helmet collects work environment data through built-in temperature and humidity sensors and light sensors;

[0020] To address the timing misalignment issue between the vibration sensor and the video feed, the operator's actions are used as the time anchor point, and the formula is applied: Calculate the timing calibration offset ,in, For the timestamp of sensor data acquisition, For video frame capture timestamps, The sensor sampling frequency, For video frame rate;

[0021] Furthermore, the method for environmental noise reduction processing is as follows:

[0022] To address image interference from environments such as fog and backlighting, noise reduction parameters are adjusted based on ambient light intensity (L) and humidity (H) data. Image denoising employs adaptive median filtering, while speech denoising uses spectral subtraction to adjust the noise reduction intensity, using the formula: Analysis yields the actual noise reduction intensity ,in, Based on the noise reduction intensity, This represents the actual light intensity. For saturated light intensity, This represents the actual ambient humidity. The humidity is saturated.

[0023] Furthermore, the method for constructing the three-level security protection system is as follows:

[0024] A triple authentication mechanism combining hardware identification, biometrics, and job-related permissions is adopted. The authentication pass conditions are as follows:

[0025]

[0026] in, For the authentication results, For terminal hardware identification, For hardware identifiers in the device whitelist, The biometric matching degree has a value range of [0, 1]. This represents the actual authority level of the personnel, with values ​​ranging from [1, 5], where 1 represents the lowest authority and 5 represents the highest authority. The minimum permission level required for the current task;

[0027] Furthermore, the method for balancing data security and job real-time performance is as follows:

[0028] Dynamic encryption is performed using the AES-256 encryption algorithm. The size of the encryption block is dynamically adjusted based on the actual transmission bandwidth B of the transmission link and the real-time requirement coefficient R of the operational scenario, using the formula:

[0029]

[0030] Analysis yields the actual encrypted block size ,in, The base encryption block size, defaulting to 512 bytes. This is the actual transmission bandwidth. To preset standard bandwidth, This is the real-time demand coefficient;

[0031] Real-time monitoring of packet characteristics in the APN / GRE channel; if abnormal behavior is detected, the link is automatically disconnected and an alarm is sent to the management platform.

[0032] Furthermore, the method for constructing intelligent assistance for the three major scenarios is as follows:

[0033] By combining skills learning data, on-site operation data, and remote collaboration data, personalized skills learning is pushed out. Learning content is recommended based on personnel's skill gaps, job requirements, and equipment types. The priority of recommended learning content is calculated by integrating the matching degree of skill gaps, job requirements, and equipment types in the work area through a preset weighted formula.

[0034] The domestically developed large-scale model, based on multimodal fault characteristics, generates a four-step operational guide: safety measures, troubleshooting steps, tool selection, and precautions. The guide content is dynamically adjusted according to the following logic:

[0035] When the operator's operating procedures do not conform to the instructions, the system automatically pushes a correction prompt; when the fault characteristics change, the guidance plan is dynamically adjusted.

[0036] Multimodal data from the field is synchronized to remote expert terminals in real time. Experts generate structured guidance opinions through voice / text replies, and the system automatically links them to the on-site operation guidance process.

[0037] Furthermore, the method for establishing the system optimization evaluation model is as follows:

[0038] By combining operational performance data, personnel feedback data, and model operation data, we analyze and optimize evaluation indicators using the following formula: Calculate the overall optimization coefficient of the system The system's operational effectiveness is evaluated, with values ​​ranging from [0, 1]. A system comprehensive optimization coefficient > 0.8 is considered excellent, a coefficient between [0.6, 0.8] is considered good, and a coefficient < 0.6 indicates the system needs further optimization. To optimize weights, To optimize the metrics, and To optimize the index of indicator numbers, To improve work efficiency, To improve the accuracy of fault diagnosis, To reduce the safety risk rate, For staff satisfaction;

[0039] If the overall system optimization coefficient is less than 0.6, the knowledge injection coefficient, actual noise reduction intensity, actual encryption block size, and skill deficiency matching degree weight will be automatically adjusted.

[0040] Furthermore, the continuous adaptation method is as follows:

[0041] Model iteration is divided into core iteration and regular iteration: core iteration targets newly added niche fault cases, adopts small-batch incremental training, and updates the knowledge graph and feature association model corresponding to the fault type; regular iteration is carried out quarterly, integrating all newly added cases for incremental training to ensure iteration efficiency and deployment stability.

[0042] Parameter adjustment backtracking verification: After parameter adjustment, the system automatically records the running data within 72 hours after the adjustment and compares it with the data before the adjustment and the historical data of similar adjustments;

[0043] If the overall optimization coefficient of the system does not improve after adjustment or if stability issues occur, the backtracking mechanism will be automatically triggered to restore the optimal parameter configuration and mark the current parameter adjustment scheme as invalid for optimization algorithm improvement.

[0044] The second part, the intelligent auxiliary method for multimodal power grid operation, maintenance and emergency repair based on a domestically developed large model provided by the embodiments of the present invention, specifically includes the following steps:

[0045] Step 1: By collecting power grid professional data and grassroots operation scenario data, and using hierarchical knowledge distillation and professional knowledge graph alignment technology, a domestic large model is constructed to dynamically balance professional accuracy and lightweight deployment. Combining the equipment and personnel operation characteristics of grassroots operation scenario data, multimodal data is collected and time-series calibration and environmental noise reduction processing are performed.

[0046] Step 2: Combine the dedicated communication channel of the power grid APN / GRE to build a three-level security protection system of identity authentication, data encryption and transmission monitoring. By dynamically adjusting the encryption parameters, the system balances the requirements of data security and real-time operation.

[0047] Step 3: Combining the domestically produced large-scale model with the processed multimodal data, construct intelligent assistance for three major scenarios: skills learning, on-site assistance, and remote collaboration. By using operational performance data and personnel feedback data, establish a system optimization and evaluation model. By dynamically adjusting model parameters and functional configurations, achieve continuous adaptation of the system to grassroots operational scenarios and personnel skills.

[0048] The beneficial effects of this invention are:

[0049] 1. Enhance professional knowledge annotation for high-frequency faults at the grassroots level to avoid knowledge redundancy in general large models; ensure professional accuracy of the model and solve the problem of insufficient knowledge by coordinating the adjustment of model pruning coefficient, knowledge injection coefficient, and scenario adaptation coefficient; control model size and adapt to low-configuration terminals at the grassroots level to solve deployment difficulties; at the same time, match personnel skill gaps to solve the problem of inaccurate skill assistance; use personnel operation actions as anchor points to solve the problem of temporal misalignment caused by differences in the data collection frequency of different modes; adjust the noise reduction intensity in combination with real-time environmental data to improve data quality in complex environments; achieve deep correlation of multimodal features through power grid professional knowledge graphs to solve the shortcomings of existing technologies that only perform decision-level fusion.

[0050] 2. By combining power grid equipment asset numbers with personnel job permissions, unauthorized equipment / personnel access is prevented; a formula is used to balance encryption strength and transmission efficiency, resolving the conflict between security protection and real-time operation; the entire process of access-transmission-use is covered, addressing the shortcomings of existing technologies that only focus on data encryption; priority recommendations are calculated based on multi-dimensional matching degrees, addressing the lack of specificity in existing training methods; guidance content is adjusted based on real-time operation data to improve the accuracy of on-site operations; expert responses are transformed into standardized operation guidance, resolving the problem of information fragmentation during collaboration; core parameters are automatically adjusted based on comprehensive optimization coefficients to address the problem of insufficient system adaptability; incremental training is adopted to avoid the interruption of grassroots deployment caused by full-scale training. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of the structure of the intelligent auxiliary system for multimodal power grid operation, maintenance and emergency repair based on a domestically produced large model provided in Embodiment 1 of the present invention;

[0053] Figure 2 This is a flowchart of the steps of the intelligent auxiliary method for multimodal power grid operation, maintenance and emergency repair based on a domestically produced large model provided in Embodiment 2 of the present invention. Detailed Implementation

[0054] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0055] Example 1: As Figure 1 As shown in the figure, the multimodal power grid operation and maintenance intelligent auxiliary system based on a domestically developed large model provided in this embodiment of the invention specifically includes the following modules:

[0056] Data processing module: By collecting power grid professional data and grassroots operation scenario data, it adopts hierarchical knowledge distillation and professional knowledge graph alignment technology to build a domestic large model, dynamically balance professional accuracy and lightweight deployment, and collect multimodal data and perform time-series calibration and environmental noise reduction processing based on the equipment and personnel operation characteristics of grassroots operation scenario data.

[0057] In a specific embodiment, the basic capability interfaces of the mainstream domestic large-scale models are collected, and the standard texts of the power grid operation, maintenance and emergency repair industry, structural drawings of 35kV-220kV transformers, typical line equipment and historical fault cases are obtained.

[0058] Data on grassroots operations of frontline personnel was obtained through questionnaires and on-site filming: operating habits, skill gaps, and common work scenarios. Skill gaps include, but are not limited to, weakness in identifying faults in new equipment and complex topology analysis.

[0059] Collect hardware parameters of commonly used terminals at the grassroots level and establish a library of various terminal adaptation models;

[0060] Industry standards and failure cases are transformed into a four-level knowledge graph of equipment type, failure characteristics, troubleshooting steps, and safety requirements, with entity relationships marked. In particular, high-frequency failures at the grassroots level are marked with enhanced annotations to avoid knowledge redundancy in the general large model. Grassroots high-frequency failures include, but are not limited to, line icing and transformer oil leakage.

[0061] Based on the domestically developed large-scale model, the upper layer distills general language understanding and multimodal recognition capabilities, the middle layer injects power grid professional knowledge graphs, and the lower layer optimizes the model structure for the hardware parameters of the grassroots terminals and removes redundant parameters.

[0062] Specifically, the total number of original parameters of the Guangming large model is obtained, and the total number of parameters of the module is dynamically and lightweightly calculated by combining the power grid professional knowledge. The optimized total number of model parameters is then obtained through analysis, achieving a dynamic balance between model complexity and terminal adaptability.

[0063] For example, obtaining the total number of original parameters of the Guangming large model. Total number of parameters injected into the power grid expertise module Through the formula: The total number of optimized model parameters was calculated. ,in, This is the model clipping factor, with a value range of [0.3, 0.5], determined by the terminal memory size. The scene adaptation coefficient ranges from 0.4 to 0.8, and is determined by the complexity of the task scenario. The higher the complexity of the task scenario, the larger the scene adaptation coefficient. The knowledge injection coefficient ranges from 0.6 to 0.9 and is determined by the matching degree of personnel skill gaps. The greater the matching degree of personnel skill gaps, the greater the knowledge injection coefficient.

[0064] The system uses drones equipped with high-definition cameras and smart safety helmets worn by on-site workers to collect data on the appearance of the equipment and infrared thermal imaging videos / images. The data collection triggering method supports both manual shooting and automatic triggering, and automatically captures images when the equipment temperature is abnormal or the appearance is damaged.

[0065] The smart safety helmet uses a built-in microphone to collect voice commands from on-site personnel and soundprints of equipment operation, and supports offline collection.

[0066] Vibration sensors, temperature sensors, and oil chromatography sensors deployed on the equipment collect equipment operation data, which is then transmitted to the field terminal in real time via Bluetooth.

[0067] The smart safety helmet collects work environment data through built-in temperature and humidity sensors and light sensors. The work environment data includes temperature, humidity and light intensity, providing a basis for environmental adaptation for data processing.

[0068] To address the timing misalignment issue between the vibration sensor and the video feed, the operator's actions are used as the time anchor point, and the formula is applied: Calculate the timing calibration offset ,in, For the timestamp of sensor data acquisition, For video frame capture timestamps, The sensor sampling frequency, For video frame rate;

[0069] To address image interference from environments such as fog and backlighting, noise reduction parameters are adjusted based on ambient light intensity (L) and humidity (H) data. Image denoising employs adaptive median filtering, while speech denoising uses spectral subtraction to adjust the noise reduction intensity, using the formula: Analysis yields the actual noise reduction intensity ,in, The base noise reduction strength is set to 0.8 by default. This represents the actual light intensity. For saturated light intensity, This represents the actual ambient humidity. The humidity is saturated.

[0070] Based on the power grid professional knowledge graph, we associate equipment defect features in images, fault descriptions in voice, and data anomalies in sensors with the same fault event to build a feature-fault-solution association model.

[0071] Protection Adjustment Module: Combined with the power grid APN / GRE dedicated communication channel, a three-level security protection system of identity authentication, data encryption, and transmission monitoring is constructed. By dynamically adjusting encryption parameters, the system balances data security with the real-time requirements of operations.

[0072] In a specific embodiment, hardware identifiers and equipment registration information of on-site operation terminals, sensors, and drones are collected to establish a whitelist database of equipment identities; fingerprint and facial recognition data and job permission information of grassroots personnel are collected and updated synchronously with the power grid human resources system; and transmission bandwidth, latency, packet loss rate data of APN / GRE channels, as well as abnormal link behavior data, are collected in real time. Abnormal link behavior data includes abnormal access IP and abnormal data transmission frequency.

[0073] A triple authentication mechanism combining hardware identification, biometrics, and job-related permissions is adopted. The authentication pass conditions are as follows:

[0074]

[0075] in, For the authentication results, For terminal hardware identification, For hardware identifiers in the device whitelist, The biometric matching degree has a value range of [0, 1]. This represents the actual authority level of the personnel, with values ​​ranging from [1, 5], where 1 represents the lowest authority and 5 represents the highest authority. The minimum permission level required for the current task;

[0076] Dynamic encryption is performed using the AES-256 encryption algorithm. The size of the encryption block is dynamically adjusted based on the actual transmission bandwidth B of the transmission link and the real-time requirement coefficient R of the operational scenario, using the formula:

[0077]

[0078] Analysis yields the actual encrypted block size The value range is [128 bytes, 1024 bytes], where, The base encryption block size, defaulting to 512 bytes. This is the actual transmission bandwidth. To preset standard bandwidth, This is the real-time demand coefficient, with a value range of [0.8, 1.2].

[0079] Real-time monitoring of data packet characteristics in the APN / GRE channel, including data size and transmission frequency. If abnormal behavior is detected, such as three consecutive authentication failures or data transmission frequency exceeding twice the normal range, the link will be automatically disconnected and an alarm will be sent to the management platform.

[0080] Intelligent Assistance Module: Combining domestically produced large-scale models with processed multimodal data, it constructs intelligent assistance for three major scenarios: skills learning, on-site assistance, and remote collaboration. Through work performance data and personnel feedback data, it establishes a system optimization evaluation model. By dynamically adjusting model parameters and function configurations, it achieves continuous adaptation of the system to grassroots work scenarios and personnel skills.

[0081] In a specific embodiment, skills learning data is collected through questionnaires and analysis of historical learning behavior: learned courses, assessment scores, fault diagnosis accuracy and operation time; real-time on-site operation data is collected: fault characteristic data after multimodal processing, whether safety measures are implemented according to procedures, fault diagnosis time, and emergency repair completion status; remote collaboration data is collected: text, voice and video of remote consultation requests initiated by on-site personnel, as well as text, voice and video of expert replies, and data interaction records during the collaboration process are collected.

[0082] By combining skills learning data, on-site operation data, and remote collaboration data, personalized skills learning is pushed out. Learning content is recommended based on the personnel's skill gaps, job requirements, and equipment type. The priority of the recommended learning content is calculated by integrating the matching degree of skill gaps, job requirements, and equipment type in the work area through a preset weighted formula. The value range is [0,1], and the higher the value, the higher the priority.

[0083] Among them, the skill gap matching degree is the matching degree between personnel and fault diagnosis gap. If the personnel have a transformer fault diagnosis gap, then the skill gap matching degree between the personnel and transformer-related courses is 1.

[0084] The domestically developed large-scale model, based on multimodal fault characteristics, generates a four-step operational guide: safety measures, troubleshooting steps, tool selection, and precautions. The guide content is dynamically adjusted according to the following logic:

[0085] When the operator's operating steps do not conform to the instructions, the system automatically pushes a correction prompt; when the fault characteristics change, the guidance plan is dynamically adjusted; for example, if the sensor data continues to be abnormal, the guidance plan is regenerated.

[0086] The system synchronizes multimodal data from the field to remote expert terminals in real time. The multimodal data includes images, videos, and sensor data. Experts generate structured guidance opinions through voice / text replies. The system automatically links these opinions to the on-site operation guidance process, realizing a closed loop of on-site data collection, expert analysis, and guidance feedback.

[0087] The system collects operational performance data, personnel feedback data, and model operation data. Operational performance data includes, but is not limited to: troubleshooting time, fault diagnosis accuracy, emergency repair completion time, and safety accident incidence rate. Personnel feedback data includes, but is not limited to: evaluations of the system's auxiliary effect, ease of operation, and practicality of learning content collected through post-operation questionnaires and on-site voice feedback. Model operation data includes, but is not limited to: large model inference time, multimodal data processing accuracy, and safety protection response time.

[0088] By combining operational performance data, personnel feedback data, and model operation data, we analyze and optimize evaluation indicators using the following formula:

[0089]

[0090] Calculate the overall optimization coefficient of the system The system's operational effectiveness is evaluated, with values ​​ranging from [0, 1]. A system comprehensive optimization coefficient > 0.8 is considered excellent, a coefficient between [0.6, 0.8] is considered good, and a coefficient < 0.6 indicates the system needs further optimization. To optimize weights, To optimize the metrics, and To optimize the index of indicator numbers, To improve work efficiency, To improve the accuracy of fault diagnosis, To reduce the safety risk rate, For staff satisfaction;

[0091] If the overall system optimization coefficient is less than 0.6, the following parameters will be automatically adjusted:

[0092] The knowledge injection coefficient of the domestically produced large model was improved by 0.1; the actual noise reduction strength of multimodal data was improved by 0.1; the actual encrypted block size was increased by 0.1 times; and the weight of skill deficiency matching degree in the priority of learning content recommendation was increased to 0.6.

[0093] Model iteration is divided into core iteration and regular iteration: core iteration targets newly added niche fault cases, adopts small-batch incremental training, and updates the knowledge graph and feature association model corresponding to the fault type; regular iteration is carried out quarterly, integrating all newly added cases for incremental training to ensure iteration efficiency and deployment stability.

[0094] Parameter adjustment backtracking verification: After parameter adjustment, the system automatically records the running data within 72 hours after the adjustment and compares it with the data before the adjustment and the historical data of similar adjustments;

[0095] If the overall optimization coefficient of the system does not improve after adjustment or if stability issues occur, the backtracking mechanism will be automatically triggered to restore the optimal parameter configuration and mark the current parameter adjustment scheme as invalid for optimization algorithm improvement.

[0096] Example 2: Figure 2 As shown in the figure, the intelligent auxiliary method for multimodal power grid operation, maintenance and emergency repair based on a domestically developed large model provided in this embodiment of the invention specifically includes the following steps:

[0097] Step 1: By collecting power grid professional data and grassroots operation scenario data, and using hierarchical knowledge distillation and professional knowledge graph alignment technology, a domestic large model is constructed to dynamically balance professional accuracy and lightweight deployment. Combining the equipment and personnel operation characteristics of grassroots operation scenario data, multimodal data is collected and time-series calibration and environmental noise reduction processing are performed.

[0098] Step 2: Combine the dedicated communication channel of the power grid APN / GRE to build a three-level security protection system of identity authentication, data encryption and transmission monitoring. By dynamically adjusting the encryption parameters, the system balances the requirements of data security and real-time operation.

[0099] Step 3: Combining the domestically produced large-scale model with the processed multimodal data, construct intelligent assistance for three major scenarios: skills learning, on-site assistance, and remote collaboration. By using operational performance data and personnel feedback data, establish a system optimization and evaluation model. By dynamically adjusting model parameters and functional configurations, achieve continuous adaptation of the system to grassroots operational scenarios and personnel skills.

[0100] The above provides a detailed description of one embodiment of the present invention, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. The above formulas are all dimensionless numerical calculations, and the formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world situation. The preset parameters in the formulas are set by those skilled in the art based on actual conditions and historical experience, and can be adjusted according to actual conditions. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. All equivalent changes and improvements made in accordance with the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A multimodal power grid operation, maintenance, and emergency repair intelligent auxiliary system based on a domestically developed large-scale model, characterized in that: Includes the following modules: Data processing module: By collecting power grid professional data and grassroots operation scenario data, it adopts hierarchical knowledge distillation and professional knowledge graph alignment technology to build a domestic large model, dynamically balance professional accuracy and lightweight deployment, and collect multimodal data and perform time-series calibration and environmental noise reduction processing based on the equipment and personnel operation characteristics of grassroots operation scenario data. Protection and Adjustment Module: After calibration and noise reduction processing, combined with the power grid APN / GRE dedicated communication channel, a three-level security protection system is constructed. Through dynamic encryption parameter adjustment, the data security and real-time operation requirements are balanced. Intelligent Assistance Module: Combining domestically produced large-scale models with processed multimodal data, intelligent assistance is constructed for three major scenarios. Through operational performance data and personnel feedback data, a system optimization and evaluation model is established. By dynamically adjusting model parameters and functional configurations, the system can continuously adapt to grassroots operational scenarios and personnel skills.

2. The intelligent auxiliary system for multimodal power grid operation, maintenance, and emergency repair based on a domestically developed large-scale model as described in claim 1, characterized in that, The method for constructing the domestically produced large-scale model is as follows: Collect hardware parameters of commonly used terminals at the grassroots level and establish a library of various terminal adaptation models; Industry standards and failure cases are transformed into a four-level knowledge graph of equipment type, failure characteristics, troubleshooting steps, and safety requirements, with entity relationships marked. In particular, high-frequency failures at the grassroots level are marked with enhanced annotations to avoid knowledge redundancy in the general large model. Grassroots high-frequency failures include, but are not limited to, line icing and transformer oil leakage. Based on the domestically developed large-scale model, the upper layer incorporates general language understanding and multimodal recognition capabilities, the middle layer injects power grid professional knowledge graphs, and the lower layer optimizes the model structure for the hardware parameters of the grassroots terminals and removes redundant parameters.

3. The intelligent auxiliary system for multimodal power grid operation, maintenance, and emergency repair based on a domestically developed large-scale model as described in claim 2, characterized in that, The method for optimizing the model structure is as follows: The total number of original parameters of the Guangming large model is obtained, and dynamic lightweight calculation is performed by combining the total number of parameters of the module with power grid professional knowledge. The optimized total number of model parameters is obtained by analysis, and the model complexity and terminal adaptability are dynamically balanced.

4. The intelligent auxiliary system for multimodal power grid operation, maintenance, and emergency repair based on a domestically developed large-scale model as described in claim 1, characterized in that, The method for performing timing calibration is as follows: The smart safety helmet uses a built-in microphone to collect voice commands from on-site personnel and soundprints of equipment operation, and supports offline collection. Equipment operating data is collected by vibration sensors, temperature sensors, and oil chromatography sensors deployed on the equipment; The smart safety helmet collects work environment data through built-in temperature and humidity sensors and light sensors; To address the timing misalignment issue between the vibration sensor and the video feed, the operator's actions are used as the time anchor point, and the formula is applied: Calculate the timing calibration offset ,in, For the timestamp of sensor data acquisition, For video frame capture timestamps, The sensor sampling frequency, This refers to the video frame rate.

5. The intelligent auxiliary system for multimodal power grid operation, maintenance, and emergency repair based on a domestically developed large-scale model as described in claim 1, characterized in that, The method for environmental noise reduction is as follows: To address image interference from environments such as fog and backlighting, noise reduction parameters are adjusted based on ambient light intensity (L) and humidity (H) data. Image denoising employs adaptive median filtering, while speech denoising uses spectral subtraction to adjust the noise reduction intensity, using the following formula: Analysis yields the actual noise reduction intensity ,in, Based on the noise reduction intensity, This represents the actual light intensity. For saturated light intensity, This represents the actual ambient humidity. This represents saturated humidity.

6. The intelligent auxiliary system for multimodal power grid operation, maintenance, and emergency repair based on a domestically developed large-scale model as described in claim 1, characterized in that, The method for constructing a three-tier security protection system is as follows: A triple authentication mechanism combining hardware identification, biometrics, and job-related permissions is adopted. The authentication pass conditions are as follows: in, For the authentication results, For terminal hardware identification, For hardware identifiers in the device whitelist, The biometric matching degree has a value range of [0, 1]. This represents the actual authority level of the personnel, with values ​​ranging from [1, 5], where 1 represents the lowest authority and 5 represents the highest authority. This represents the minimum privilege level required for the current task.

7. The intelligent auxiliary system for multimodal power grid operation, maintenance, and emergency repair based on a domestically developed large-scale model as described in claim 1, characterized in that, The method for balancing data security and job real-time performance is as follows: Dynamic encryption is performed using the AES-256 encryption algorithm. The size of the encryption block is dynamically adjusted based on the actual transmission bandwidth B of the transmission link and the real-time requirement coefficient R of the operational scenario, using the formula: Analysis yields the actual encrypted block size ,in, The base encryption block size, defaulting to 512 bytes. This is the actual transmission bandwidth. To preset standard bandwidth, This is the real-time demand coefficient; The system monitors the characteristics of data packets in the APN / GRE channel in real time. If abnormal behavior is detected, it automatically disconnects the link and sends an alarm to the management platform.

8. The intelligent auxiliary system for multimodal power grid operation, maintenance, and emergency repair based on a domestically developed large-scale model as described in claim 1, characterized in that, The method for constructing intelligent assistance for the three major scenarios is as follows: By combining skills learning data, on-site operation data, and remote collaboration data, personalized skills learning is pushed out. Learning content is recommended based on personnel's skill gaps, job requirements, and equipment types. The priority of recommended learning content is calculated by integrating the matching degree of skill gaps, job requirements, and equipment types in the work area through a preset weighted formula. The domestically developed large-scale model, based on multimodal fault characteristics, generates a four-step operational guide: safety measures, troubleshooting steps, tool selection, and precautions. The guide content is dynamically adjusted according to the following logic: When the operator's operating procedures do not conform to the instructions, the system will automatically push a correction prompt; when the fault characteristics change, the guidance plan will be dynamically adjusted. The system synchronizes multimodal data from the field to remote expert terminals in real time. Experts then generate structured guidance opinions through voice / text replies, which are automatically linked to the on-site operation guidance process.

9. The intelligent auxiliary system for multimodal power grid operation, maintenance, and emergency repair based on a domestically developed large-scale model as described in claim 1, characterized in that, The method for establishing the system optimization evaluation model is as follows: By combining operational performance data, personnel feedback data, and model operation data, we analyze and optimize evaluation indicators using the following formula: Calculate the overall optimization coefficient of the system The system's operational effectiveness is evaluated, with values ​​ranging from [0, 1]. A system comprehensive optimization coefficient > 0.8 is considered excellent, a coefficient between [0.6, 0.8] is considered good, and a coefficient < 0.6 indicates the system needs further optimization. To optimize weights, To optimize the metrics, and To optimize the index of indicator numbers, To improve work efficiency, To improve the accuracy of fault diagnosis, To reduce the safety risk rate, For staff satisfaction; If the overall system optimization coefficient is less than 0.6, the knowledge injection coefficient, actual noise reduction intensity, actual encryption block size, and skill deficiency matching degree weight will be automatically adjusted.

10. The intelligent auxiliary system for multimodal power grid operation, maintenance, and emergency repair based on a domestically developed large-scale model as described in claim 1, characterized in that, The method for continuous adaptation is as follows: The model iteration is divided into core iteration and regular iteration: the core iteration targets newly added niche fault cases, adopts small-batch incremental training, and updates the knowledge graph and feature association model of the corresponding fault type. Regular iterations are conducted quarterly, integrating all newly added cases for incremental training to ensure iteration efficiency and deployment stability; Parameter adjustment backtracking verification: After parameter adjustment, the system automatically records the running data within 72 hours after the adjustment and compares it with the data before the adjustment and the historical data of similar adjustments; If the overall optimization coefficient of the system does not improve after adjustment or if stability issues occur, the backtracking mechanism will be automatically triggered to restore the optimal parameter configuration and mark the current parameter adjustment scheme as invalid for optimization algorithm improvement.

Images