A power grid monitoring bionic robot intelligent decision-making method supporting rule self-learning

Abstract

The present application relates to the technical field of smart grid monitoring, and discloses a power grid monitoring bionic robot intelligent decision-making method supporting rule self-learning, comprising the following steps: step S1, multi-source heterogeneous data sensing and fusion; step S2, dynamic rule self-evolution and knowledge modeling; step S3, multi-modal collaborative decision-making and closed-loop control. Through a visual acquisition unit, a voice interaction unit and various industrial communication protocols such as OPCUA and Modbus TCP, multi-source heterogeneous data such as meter reading, indicator light status, mechanical component displacement, monitor voice instructions and equipment operation noise of power grid equipment are comprehensively collected, wavelet denoising and semantic encoding conversion are performed on the data, a cross-system association is established by taking equipment code as the main key, a unified data integration view is constructed, and the problem of data acquisition limitation in the traditional monitoring mode is solved.

Description

Technical Field

[0001] This invention relates to the field of smart grid monitoring technology, specifically to a smart decision-making method for a bionic robot for power grid monitoring that supports rule self-learning. Background Technology

[0002] With the rapid development and intelligent transformation of power systems, the scale of power grids is constantly expanding and their structure is becoming increasingly complex. The safe and stable operation of these grids is becoming increasingly important for socio-economic development and people's daily lives. Traditional power grid monitoring mainly relies on fixed-installation sensor networks and manual inspections. However, these methods are gradually revealing many limitations when dealing with the complexity and dynamism of modern power grids.

[0003] A search revealed Chinese patent CN118081746A, which discloses a self-learning method and device for robot friction parameters. The self-learning method includes: acquiring the robot's real-time joint torque, position, velocity, and acceleration; inputting the real-time joint torque, position, velocity, and acceleration into a robot dynamics model, and calculating the real-time friction force acting on the robot through the robot dynamics model; inputting the real-time friction force and velocity into a Coulomb viscous friction model, and updating the friction parameters corresponding to the Coulomb viscous friction model in real time using a least squares method with a forgetting factor. Using the method disclosed in this invention, the problem that a single set of dynamic parameters cannot meet the requirements of all robots of the same model, nor can it meet the requirements of the same robot at different times, can be solved in the prior art.

[0004] However, the aforementioned technical solutions, in the context of power grid monitoring, do not address the rule-based self-learning requirements specific to the power grid monitoring field, such as the dynamic optimization of power grid accident anomaly judgment rules and multi-source data fusion strategies. This makes them ill-suited to the intelligent decision-making requirements of complex power grid operation scenarios. Furthermore, these solutions lack the ability to integrate heterogeneous data from multiple systems such as scheduling, centralized control, and operation and maintenance, failing to construct a unified knowledge rule base and real-time information pool. Since power grid monitoring requires real-time responses to massive data changes, existing methods suffer from lag in dynamic rule adjustment and self-evolution mechanisms, failing to meet real-time requirements such as root cause reasoning for power grid accidents and optimization of handling strategies. Therefore, this invention designs a rule-based self-learning bionic robot-based intelligent decision-making method for power grid monitoring to address the aforementioned problems. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent decision-making method for a bionic robot for power grid monitoring that supports rule self-learning, thereby solving the problem in the prior art that it cannot adapt to the complex operating scenarios of the power grid.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A method for intelligent decision-making in a biomimetic robot for power grid monitoring that supports rule self-learning includes the following steps:

[0008] Step S1, Multi-source Heterogeneous Data Perception and Fusion, includes the following specific processes: Step S1.1, Using a target detection algorithm, the visual acquisition unit identifies meter readings, indicator light status, and mechanical component displacement; Step S1.2, Using a ring microphone array in the voice interaction unit, the unit collects the monitor's voice commands and equipment operating noise, and uses WaveNet voiceprint separation technology to extract power terminology command features; Step S1.3, Using the OPC UA protocol to connect to the D5000 scheduling system to obtain real-time telemetry data, using the Modbus TCP protocol to connect to the E6000 centralized control system to obtain alarm signals, and using a direct database connection to obtain equipment ledger data from the OMS / PMS system; Step S1.4, The telemetry data undergoes wavelet denoising processing, the alarm text is semantically encoded and converted, and a cross-system association with the equipment code as the primary key is established;

[0009] Step S2, Dynamic Rule Self-Evolution and Knowledge Modeling: A dynamic weighted rule base is constructed based on the historical case library, and a neural network architecture is used to realize rule evolution; a knowledge graph of equipment health status is constructed to form a feature model of the entire equipment life cycle;

[0010] Step S3: Multimodal collaborative decision-making and closed-loop control. When an accident or abnormality is triggered, voice, vision, and motion collaborative decision-making are executed; the rule base is dynamically corrected based on on-site detection data to generate a reliable handling chain and update the equipment status prediction.

[0011] Preferably, the industrial communication protocol in step S1.3 includes: step S1.3.1, using the OPC UA protocol to connect to the D5000 scheduling system and collect telemetry data at a frequency of 10Hz; step S1.3.2, using the Modbus TCP protocol to connect to the E6000 centralized control system to achieve millisecond-level response to alarm signals; and step S1.3.3, connecting to the OMS system via JDBC to periodically synchronize equipment maintenance log data.

[0012] Preferably, the standardization process in step S1.4 includes: threshold denoising using a 5-level decomposition with the sym5 wavelet basis; generating a 768-dimensional semantic vector for the alarm text using the BERT model; and calculating the cosine similarity with historical cases.

[0013] Preferably, the rule evolution implementation in step S2 includes: step S2.1, constructing a dual neural network architecture, including a policy generation network that processes time-series data features based on LSTM and outputs an initial disposal policy; and a rule evaluation network that evaluates the policy effect based on DQN and generates weight adjustment amounts; step S2.2, setting a rule effectiveness decay function: Wt=W0×e^(-λt), triggering an elimination mechanism when Wt<0.3; step S2.3, establishing device-related impact nodes in the knowledge graph, including electrical topology connections and common-source fault propagation paths.

[0014] Preferably, the rule evaluation network adopts a dual reward mechanism, including: an immediate reward function R_imm = 0.6 × (reduction rate of processing time) + 0.4 × (reduction rate of equipment damage); and a long-term reward function R_long = 0.7 × equipment health score + 0.3 × change in rule base entropy.

[0015] Preferably, the collaborative decision-making in step S3 includes the following steps: Step S3.1, the voice layer generates a TTS alarm containing the risk level and receives a voice confirmation command; Step S3.2, the vision layer displays a three-dimensional anatomical model of the device and a heat map of abnormal parts on the AR interface; Step S3.3, the motion layer controls the robotic arm to perform infrared temperature measurement and partial discharge detection.

[0016] Preferably, the robotic arm detection includes: setting an end-effector contact force threshold F_contact≤0.2×F_device_max; automatically initiating redundant detection when the coefficient of variation of the detection data CV>15%; and the condition for triggering rule backtracking training based on feedback data is |detected value - predicted value|>3σ.

[0017] Preferably, step S3 further includes: generating an accident report chain with a digital signature, including: the original data hash value, a video summary of the handling process, and a rule version identifier; and updating the equipment remaining life prediction by adopting an improved Weibull model.

[0018] Preferably, the method further includes initialization configuration: calibrating the vision-robotic arm coordinate transformation matrix with an error controlled within ±0.1mm; setting the communication heartbeat cycle: ≤50ms for D5000 systems and ≤100ms for E6000 systems; and loading emergency response whitelist rules, including priority handling strategies for main transformer overload and busbar undervoltage.

[0019] Preferably, the method further includes a performance optimization and iterative upgrade module, the specific implementation of which includes:

[0020] Performance monitoring and evaluation involves real-time monitoring of key performance indicators such as execution efficiency, data processing speed, and decision accuracy at each step during method operation, and periodic generation of performance evaluation reports. Adaptive parameter adjustment automatically adjusts the parameters involved in the method based on the performance evaluation results. Model compression and optimization compresses the neural network architecture in step S2, using techniques such as pruning and quantization to reduce the model size without significantly reducing performance, thereby improving the model's running speed and resource utilization. Simultaneously, the training and inference algorithms of the model are optimized in a targeted manner.

[0021] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0022] 1. In this invention, through a visual acquisition unit, a voice interaction unit, and various industrial communication protocols such as OPC UA and Modbus TCP, multi-source heterogeneous data such as meter readings, indicator light status, mechanical component displacement, monitor voice commands, and equipment operating noise of power grid equipment are comprehensively collected. After processing such as wavelet denoising and semantic coding conversion, cross-system associations are established with equipment codes as the primary key to construct a unified data integration view, thereby solving the problem of data acquisition limitations in traditional monitoring methods.

[0023] 2. In this invention, a dynamic weighted rule base is constructed based on a historical case library. A dual neural network architecture of a policy generation network and a rule evaluation network is adopted to realize the dynamic evolution of rules. At the same time, a rule validity decay function is set. When the weight decays to a threshold, an elimination mechanism is triggered to ensure continuous optimization of the rule base.

[0024] 3. In this invention, when an accident is triggered abnormally, the voice layer generates a TTS alarm and receives a voice confirmation command, the vision layer displays a three-dimensional anatomical model of the device and a heat map of the abnormal part on the AR interface, the motion layer controls the robotic arm to perform infrared temperature measurement and partial discharge detection, and at the same time dynamically corrects the rule base based on the on-site detection data, generates a reliable handling chain, and updates the device status prediction. Attached Figure Description

[0025] Figure 1 This is a diagram of the overall system architecture of the present invention;

[0026] Figure 2 This is a flowchart of the multi-source data fusion process of the present invention;

[0027] Figure 3 This is a schematic diagram of the rule self-evolution mechanism of the present invention;

[0028] Figure 4 This is a flowchart of the multimodal collaborative decision-making process of the present invention;

[0029] Figure 5 This is a block diagram of the performance optimization and iterative upgrade module of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1;

[0032] Please see Figures 1-5 A method for intelligent decision-making in a bionic robot for power grid monitoring that supports rule self-learning includes the following steps: Step S1, multi-source heterogeneous data perception and fusion, including the following specific processes: Step S1.1, using a target detection algorithm through a visual acquisition unit to identify meter readings, indicator light status, and mechanical component displacement; Step S1.2, using a ring microphone array in a voice interaction unit to collect the monitor's voice commands and equipment operating noise, and using WaveNet voiceprint separation technology to extract power terminology command features; Step S1.3, using the OPC UA protocol to interface with the D5000 scheduling system to obtain real-time telemetry data, and using Modbus... The system uses TCP protocol to connect to the E6000 centralized control system to obtain alarm signals and uses direct database connection to obtain equipment ledger data from the OMS / PMS system. Step S1.4: Wavelet denoising is performed on the telemetry data, and semantic encoding is performed on the alarm text to establish a cross-system association with the equipment code as the primary key. Step S2: Dynamic rule self-evolution and knowledge modeling are implemented. A dynamic weighted rule base is built based on a historical case library, and a neural network architecture is used to achieve rule evolution. A knowledge graph of equipment health status is constructed to form a full lifecycle feature model of the equipment. Step S3: Multimodal collaborative decision-making and closed-loop control are implemented. When an accident or anomaly is triggered, voice, vision, and motion collaborative decisions are executed. The rule base is dynamically corrected based on on-site detection data to generate a reliable handling chain and update the equipment status prediction.

[0033] The industrial communication protocol in step S1.3 includes: Step S1.3.1, using the OPC UA protocol to connect to the D5000 scheduling system and collect telemetry data at a frequency of 10Hz; Step S1.3.2, using the Modbus TCP protocol to connect to the E6000 centralized control system to achieve millisecond-level response to alarm signals; Step S1.3.3, connecting to the OMS system via JDBC to synchronize equipment maintenance log data on a regular basis.

[0034] The rule evolution implementation in step S2 includes: Step S2.1, constructing a dual neural network architecture, including a policy generation network that processes time-series data features based on LSTM and outputs an initial disposal policy; and a rule evaluation network that evaluates the policy effect based on DQN and generates weight adjustment amounts; Step S2.2, setting a rule effectiveness decay function: Wt=W0×e^(-λt), triggering an elimination mechanism when Wt<0.3; Step S2.3, establishing equipment-related impact nodes in the knowledge graph, including electrical topology connections and common-source fault propagation paths. The rule evaluation network adopts a dual reward mechanism, including: an immediate reward function R_imm=0.6×(disposal time reduction rate)+0.4×(equipment damage reduction rate); and a long-term reward function R_long=0.7×equipment health score+0.3×rule base entropy change.

[0035] The working principle of this invention is as follows: In a power grid monitoring scenario, this method first comprehensively collects various key information through a multi-source heterogeneous data perception and fusion process. The visual acquisition unit uses a target detection algorithm to accurately identify meter readings, indicator light status, and mechanical component displacement of power grid equipment, providing visual basis for decision-making. The voice interaction unit's ring microphone array captures the monitor's voice commands and equipment operating noise, and uses WaveNet voiceprint separation technology to extract the characteristics of power terminology commands. Regarding data integration, the OPC UA protocol collects real-time telemetry data from the D5000 system at a frequency of 10Hz, the Modbus TCP protocol achieves millisecond-level response to alarm signals from the E6000 system, and simultaneously obtains equipment ledger data from the S / PMS system through direct database connection to OM. The collected telemetry data undergoes wavelet denoising processing, and the alarm text is semantically encoded and converted. Finally, a cross-system association is established using the equipment code as the primary key, forming a unified data integration view.

[0036] In the dynamic rule self-evolution and knowledge modeling stages, a dynamic weighted rule base is constructed based on a historical case library to provide rule support for power grid monitoring decisions. A dual neural network architecture is adopted to drive rule evolution. The policy generation network uses LSTM to process time-series data features and outputs initial treatment policies, while the rule evaluation network uses DQN to evaluate the effectiveness of the policies and generate and adjust the weights. Simultaneously, a rule effectiveness decay function Wt = W0 × e^(-λt) is set. When Wt < 0.3, an elimination mechanism is triggered to ensure continuous optimization of the rule base.

[0037] When a power grid anomaly occurs, the system enters a multimodal collaborative decision-making and closed-loop control phase. The voice layer generates a TTS alarm containing the risk level and receives voice confirmation commands. The vision layer displays a 3D anatomical model of the equipment and a heat map of the affected area on an AR interface. The motion layer controls a robotic arm to perform infrared thermometry and partial discharge detection, with the end-effector contact force threshold controlled within the range of F_contact ≤ 0.2 × F_device_max. Redundant detection is automatically initiated when the coefficient of variation (CV) of the detected data exceeds 15%. The condition for triggering rule backtracking training based on feedback data is |detected value - predicted value| > 3σ. Simultaneously, a digitally signed accident report chain is generated, containing the original data hash value, a video summary of the handling process, and a rule version identifier. An improved Weibull model is used to update the equipment's remaining life prediction.

[0038] Example 2;

[0039] Please see Figures 1-5 In this embodiment of the invention, the collaborative decision-making in step S3 includes the following steps: Step S3.1, the voice layer generates a TTS alarm containing the risk level and receives a voice confirmation command; Step S3.2, the vision layer displays a three-dimensional anatomical model of the device and a heat map of abnormal parts on the AR interface; Step S3.3, the motion layer controls the robotic arm to perform infrared temperature measurement and partial discharge detection. The robotic arm detection includes: setting the end-effector contact force threshold F_contact≤0.2×F_device_max; automatically starting redundant detection when the coefficient of variation (CV) of the detection data >15%; the condition for triggering rule backtracking training based on feedback data is |detection value - predicted value| >3σ. Step S3 also includes: generating an accident report chain with a digital signature, including: the original data hash value, a video summary of the handling process, and a rule version identifier; updating the device's remaining life prediction using an improved Weibull model.

[0040] The method also includes initialization configuration: calibrating the vision-robotic arm coordinate transformation matrix with an error controlled within ±0.1mm; setting the communication heartbeat cycle: ≤50ms for D5000 systems and ≤100ms for E6000 systems; and loading emergency response whitelist rules, including priority handling strategies for main transformer overload and bus voltage loss. The method also includes a performance optimization and iterative upgrade module, specifically implemented as follows: performance monitoring and evaluation, real-time monitoring of key performance indicators such as execution efficiency, data processing speed, and decision accuracy during method operation, and periodic generation of performance evaluation reports; adaptive parameter adjustment, automatically adjusting parameters involved in the method based on performance evaluation results; model compression and optimization, compressing the neural network architecture in step S2 using pruning, quantization, and other techniques to reduce model size without significantly reducing performance, thereby improving model running speed and resource utilization; and targeted optimization of the model's training and inference algorithms.

[0041] The working principle of this invention is as follows: During the collaborative decision-making phase, the voice layer not only generates TTS alarms containing risk levels but also receives voice confirmation commands, ensuring two-way interaction and confirmation of information transmission. The vision layer intuitively displays a three-dimensional anatomical model of the equipment and a heat map of abnormal parts on the AR interface, helping monitoring personnel quickly locate faults. The motion layer controls the robotic arm to perform infrared temperature measurement and partial discharge detection, strictly controlling the end-effector contact force threshold to avoid damage to the equipment.

[0042] Furthermore, this embodiment includes initialization configuration, such as calibrating the vision-robotic arm coordinate transformation matrix with an error control within ±0.1mm, and setting the communication heartbeat cycle to ensure communication stability with the D5000 and E6000 systems. It also loads an emergency response whitelist rule to prioritize handling situations such as main transformer overload and busbar undervoltage. The performance optimization and iterative upgrade module monitors the execution efficiency, data processing speed, and decision accuracy of each step in real time, regularly generates performance evaluation reports, and automatically adaptively adjusts the parameters in the method based on the evaluation results. Simultaneously, it compresses the neural network architecture, employing techniques such as pruning and quantization to reduce model size, thereby improving operating speed and resource utilization.

[0043] Example 3;

[0044] Please see Figures 1-5 In this embodiment of the invention, taking a 220kV substation monitoring scenario as an example, the intelligent decision-making method is applied, and the specific parameters and process are as follows:

[0045] In the multi-source heterogeneous data perception and fusion stage, the high-resolution camera equipped in the visual acquisition unit utilizes the YOLOv5 target detection algorithm to identify equipment meter readings 15 times per second, achieving an accuracy rate of 95%. It can accurately identify minute changes in meter readings within the range of 0-100MPa, while simultaneously monitoring indicator light status and mechanical component displacement, with millimeter-level accuracy in mechanical component displacement monitoring. The voice interaction unit's circular microphone array consists of eight microphones, collecting monitor voice commands and equipment operating noise at a sampling rate of 44.1kHz. After processing with WaveNet voiceprint separation technology, the accuracy rate of extracting power terminology commands exceeds 90%. For data integration, the OPC UA protocol is used to interface with the D5000 scheduling system, acquiring real-time telemetry data including voltage and current at a frequency of 10Hz, with a data transmission latency of less than 50ms. The Modbus TCP protocol is used to interface with the centralized control E6000 system to obtain alarm signals, achieving millisecond-level response with an alarm signal response time of no more than 200ms. The OMS system is connected via JDBC, synchronizing equipment maintenance log data every 15 minutes. The collected telemetry data underwent wavelet denoising processing, effectively removing over 90% of high-frequency noise. After semantic encoding conversion, alarm texts were linked across systems using equipment codes as the primary key, successfully integrating data from over 200 devices within the substation to form a unified data integration view.

[0046] In the dynamic rule self-evolution and knowledge modeling stage, a dynamic weighted rule base was constructed based on over 2000 historical cases accumulated from the substation over the past 5 years, initially containing 500 handling rules. The adopted dual neural network architecture uses an LSTM-based policy generation network with 128 hidden layer neurons and a sequence length of 20. After processing time-series data features, it outputs the initial handling policy, with policy generation taking approximately 0.5 seconds. The rule evaluation network is based on DQN, employing Experience Replay technology with a memory capacity of 10000 and 64 training samples per batch. After evaluating the policy effectiveness, it generates weight adjustments, with an evaluation cycle of once every 10 policy generation iterations. The rule effectiveness decay function parameter λ is set to 0.1. When the rule weight Wt falls below 0.3, an elimination mechanism is triggered. In the past 6 months of operation, over 30 inefficient rules have been eliminated. Equipment-related impact nodes are established in the knowledge graph, including the complete electrical topology connections of electrical equipment within the substation and common common-source fault propagation paths.

[0047] In the multimodal collaborative decision-making and closed-loop control phase, when a power grid anomaly is triggered, such as an abnormal rise in the oil temperature of a main transformer in a substation triggering an alarm, the voice layer immediately generates a TTS alarm containing "Main transformer oil temperature abnormal, high risk level," and receives the monitor's voice confirmation command at a 16kHz sampling rate. The vision layer displays a real-time 3D anatomical model of the main transformer equipment on the AR interface, clearly marking the location of the oil temperature anomaly and displaying the temperature distribution in the form of a heat map, with the heat map updating at a frequency of 10 frames per second. The motion layer controls the robotic arm to perform infrared thermometry and partial discharge detection. The contact force at the end of the robotic arm is controlled within 20% of the maximum allowable contact force of the equipment, with infrared thermometry accuracy reaching ±0.5℃ and partial discharge detection sensitivity of 0.1pC. During the detection process, if the coefficient of variation (CV) of the detection data exceeds 15%, or if the ratio of the standard deviation to the mean of five consecutive detection data exceeds 15%, redundant detection is automatically initiated, increasing the detection frequency and performing data comparison and verification.

[0048] In the past year of operation, this method has successfully handled more than 30 power grid accidents and anomalies, with the average handling time shortened by 40% compared with traditional methods, equipment damage reduced by 35%, and the level of intelligence in power grid monitoring and the efficiency of accident handling improved.

[0049] Working Principle: In power grid monitoring scenarios, the system first performs multi-source heterogeneous data perception and fusion. Key information is collected comprehensively through visual acquisition units, voice interaction units, and industrial communication protocols, including equipment meter readings, indicator light status, mechanical component displacement, monitor voice commands, equipment operating noise, telemetry data, alarm signals, and equipment ledger data. This data is then processed to form a unified integrated view. Next, a dynamic weighted rule base is built based on a historical case library. A dual-neural network architecture drives rule evolution, and a rule effectiveness decay function ensures continuous optimization of the rule base. When a power grid accident or anomaly occurs, the system enters a multimodal collaborative decision-making and closed-loop control phase. The voice layer, visual layer, and motion layer work collaboratively, generating a digitally signed accident report chain and updating the equipment remaining life prediction using an improved Weibull model. Furthermore, the system includes initialization configuration, performance optimization, and iterative upgrade modules to improve the overall performance and efficiency of the method.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent decision-making in a biomimetic robot for power grid monitoring that supports rule self-learning, characterized in that, Includes the following steps: Step S1, multi-source heterogeneous data perception and fusion, includes the following specific processes: Step S1.1: The visual acquisition unit uses a target detection algorithm to identify the meter readings, indicator light status, and mechanical component displacement of the equipment. Step S1.2: Collect the voice commands of the monitor and the operating noise of the equipment through the ring microphone array of the voice interaction unit, and extract the characteristics of power terminology commands using WaveNet voiceprint separation technology. Step S1.3: Obtain real-time telemetry data by connecting to the D5000 system via OPC UA protocol, obtain alarm signals by connecting to the E6000 centralized control system via Modbus TCP protocol, and obtain equipment ledger data from the OMS / PMS system via direct database connection. Step S1.4: Perform wavelet denoising on the telemetry data, perform semantic encoding conversion on the alarm text, and establish a cross-system association with the device code as the primary key; Step S2, Dynamic Rule Self-Evolution and Knowledge Modeling: A dynamic weighted rule base is constructed based on the historical case library, and a neural network architecture is used to realize rule evolution; a knowledge graph of equipment health status is constructed to form a feature model of the entire equipment life cycle; The rule evolution implementation methods in step S2 include: Step S2.1: Construct a dual neural network architecture, including a policy generation network that processes time-series data features based on LSTM and outputs an initial treatment policy; and a rule evaluation network that evaluates the policy effect based on DQN and generates weight adjustment amounts. Step S2.2: Set the rule validity decay function: Wt=W0×e^(-λt), and trigger the elimination mechanism when Wt<0.3; Step S2.3: Establish device-related impact nodes in the knowledge graph, including electrical topology connections and common-source fault propagation paths; Step S3: Multimodal collaborative decision-making and closed-loop control. When an accident or abnormality is triggered, voice, vision, and motion collaborative decision-making are executed; the rule base is dynamically corrected based on on-site detection data to generate a reliable handling chain and update the equipment status prediction. The collaborative decision-making in step S3 includes the following steps: Step S3.1: The voice layer generates a TTS alarm containing the risk level and receives a voice confirmation command; Step S3.2: The visual layer displays the device's three-dimensional anatomical model and heat map of abnormal areas on the AR interface; Step S3.3: The motion layer controls the robotic arm to perform infrared temperature measurement and partial discharge detection.

2. The intelligent decision-making method for a biomimetic robot supporting rule self-learning for power grid monitoring according to claim 1, characterized in that: The industrial communication protocol in step S1.3 includes: Step S1.3.1: Use the OPC UA protocol to interface with and schedule the D5000 system, and collect telemetry data at a frequency of 10Hz; Step S1.3.2: Use the Modbus TCP protocol to connect to the E6000 centralized control system to achieve millisecond-level response to alarm signals; Step S1.3.3: Connect to the OMS system via JDBC to periodically synchronize equipment maintenance log data.

3. The intelligent decision-making method for a biomimetic robot supporting rule self-learning for power grid monitoring according to claim 1, characterized in that, The standardization process in step S1.4 includes: threshold denoising using a 5-level decomposition based on the sym5 wavelet basis; generating a 768-dimensional semantic vector for the alarm text using the BERT model; and calculating the cosine similarity with historical cases.

4. The intelligent decision-making method for a biomimetic robot supporting rule self-learning for power grid monitoring according to claim 1, characterized in that, The rule evaluation network adopts a dual reward mechanism, including: an immediate reward function R_imm = 0.6 × (reduction rate of processing time) + 0.4 × (reduction rate of equipment damage); and a long-term reward function R_long = 0.7 × equipment health score + 0.3 × change in rule base entropy.

5. The intelligent decision-making method for a biomimetic robot supporting rule self-learning for power grid monitoring according to claim 1, characterized in that, The robotic arm detection includes: setting the end-effector contact force threshold F_contact≤0.2×F_device_max; automatically starting redundant detection when the coefficient of variation of the detection data CV>15%; and triggering rule backtracking training based on feedback data when |detected value - predicted value|>3σ.

6. The intelligent decision-making method for a biomimetic robot supporting rule self-learning for power grid monitoring according to claim 1, characterized in that, Step S3 further includes: generating an incident report chain with a digital signature, which includes: the original data hash value, a video summary of the handling process, and a rule version identifier; and updating the equipment remaining life prediction by using an improved Weibull model.

7. The intelligent decision-making method for a biomimetic robot supporting rule self-learning for power grid monitoring according to claim 1, characterized in that, The method also includes initialization configuration: calibrating the vision-robotic arm coordinate transformation matrix with an error controlled within ±0.1mm; setting the communication heartbeat cycle: ≤50ms for D5000 systems and ≤100ms for E6000 systems; and loading emergency response whitelist rules, including priority handling strategies for main transformer overload and busbar undervoltage.

8. The intelligent decision-making method for a biomimetic robot supporting rule self-learning for power grid monitoring according to claim 1, characterized in that, The method also includes a performance optimization and iterative upgrade module, the specific implementation of which includes: Performance monitoring and evaluation involves real-time monitoring of key performance indicators such as execution efficiency, data processing speed, and decision accuracy at each step during the method's operation, and the generation of performance evaluation reports periodically. Adaptive parameter adjustment: Based on performance evaluation results, the parameters involved in the method are automatically adjusted adaptively. Model compression and optimization: The neural network architecture in step S2 is compressed by using pruning and quantization techniques to reduce the model size without significantly reducing performance, thereby improving the model's running speed and resource utilization; at the same time, the training and inference algorithms of the model are optimized in a targeted manner.

Images