Micro-grid power facility intelligent inspection robot cooperative inspection and remote guidance system

Abstract

The application belongs to the technical field of intelligent inspection of power facilities, and specifically discloses a micro-grid power facility intelligent inspection robot cooperative inspection and remote guidance system. The application strictly limits the processing of original inspection data and model training in the local edge nodes of each micro-grid, and only exchanges encrypted model parameter increments between networks, thereby eliminating the risk of sensitive power facility data being intercepted or leaked during transmission. By collecting model knowledge from multiple micro-grids in different geographical locations and different environmental conditions, each local model can learn the fault feature experience of other nodes, which is conducive to building a global intelligent model with strong adaptability and accurate diagnosis. Through seamless integration of the whole process of data collection, model training, global aggregation, local optimization, autonomous diagnosis and remote guidance, the dependence on artificial inspection is reduced, the operation and maintenance efficiency is improved, and the efficiency and reliability of the inspection task are ensured.

Description

Technical Field

[0001] This invention belongs to the field of intelligent inspection technology for power facilities, specifically relating to a collaborative inspection and remote guidance system for intelligent inspection robots of microgrid power facilities. Background Technology

[0002] As a crucial component of modern power systems, the safe and stable operation of microgrids is vital for ensuring regional energy supply and the overall reliability of the power grid. Power facilities are the core assets of microgrids, and their daily inspection and maintenance are key aspects of ensuring microgrid safety. With the development of artificial intelligence and robotics, adopting intelligent inspection robots to replace traditional manual inspections has become an important development direction for improving operation and maintenance efficiency and reducing safety risks.

[0003] In existing technologies, intelligent inspection of power facilities typically employs solutions based on centralized cloud platforms. This approach involves inspection robots deployed at various sites collecting data such as equipment images and infrared thermal images. The massive amounts of raw data are then transmitted via network to a remote central server. On the central server, all the collected data is used to train a unified, large-scale fault diagnosis AI model, which is then distributed to all robots for subsequent inspection tasks.

[0004] However, the above-mentioned inspection scheme based on centralized cloud computing has certain drawbacks: (1) It transmits highly sensitive raw data, including the location and operating status of power facilities, over long distances and stores them centrally in the cloud, which brings great risks to data security and privacy.

[0005] (2) Due to data security and business barriers, microgrid data from different operators or regions often form data silos, making it difficult to effectively share and integrate them, which limits the further improvement of diagnostic model performance.

[0006] (3) Continuously uploading high-resolution inspection data puts enormous pressure on network bandwidth and also affects the real-time performance of the system. Summary of the Invention

[0007] In view of this, in order to solve the problems mentioned in the background technology, a collaborative inspection and remote guidance system for intelligent inspection robots of microgrid power facilities is proposed.

[0008] The objective of this invention can be achieved through the following technical solution: This invention provides a collaborative inspection and remote guidance system for intelligent inspection robots of microgrid power facilities, comprising: deploying an edge training unit locally in each independent microgrid; collecting raw inspection data of the corresponding microgrid power facilities through inspection robots deployed in each microgrid; and obtaining a basic diagnostic model based on the training of the edge training unit.

[0009] Extract the model weight gradient, bias parameters, and feature mapping layer parameter information of the basic diagnostic model, encrypt them to obtain encrypted model incremental data, and upload them to the global aggregation node.

[0010] The global aggregation node receives encrypted incremental model data uploaded from multiple microgrids, decrypts it, performs secure aggregation calculations, and generates a global diagnostic model. At the same time, a hierarchical model access and data auditing mechanism is established on the global aggregation node.

[0011] The parameters of the global diagnostic model are distributed to the edge training units of each microgrid. Based on the distributed global parameters and combined with new local inspection data, the model is fine-tuned and optimized to form a local optimized diagnostic model.

[0012] Inspection robots in each microgrid use optimized diagnostic models to diagnose faults; through periodic execution, each microgrid can collaboratively improve the accuracy and generalization ability of the fault diagnosis model without sharing the original data.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention strictly restricts the processing of raw inspection data and model training to the local edge nodes of each microgrid, and only exchanges encrypted model parameter increments between networks, effectively eliminating the risk of sensitive power facility data being intercepted or leaked during transmission, and solving the data security bottleneck faced by traditional centralized intelligent inspection.

[0014] 2. This invention uses a federated learning mechanism to aggregate model knowledge from multiple microgrids with different geographical locations and environmental conditions, enabling each local model to learn from the fault characteristics and experiences of other nodes. In particular, the ability to identify rare or unique fault types is enhanced, which is conducive to building a more adaptable and more accurate global intelligent model.

[0015] 3. This invention reduces reliance on manual inspections and improves operational efficiency by seamlessly integrating the entire process of data acquisition, model training, global aggregation, local optimization, autonomous diagnosis, and remote guidance. Furthermore, its intelligent hierarchical diagnosis and alarm mechanism, as well as the logic of introducing expert remote guidance when the model is uncertain, ensures the high efficiency and reliability of inspection tasks. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1This is a schematic diagram of the system module connections of the present invention. Detailed Implementation

[0018] 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.

[0019] Please see Figure 1 As shown, the present invention provides a collaborative inspection and remote guidance system for intelligent inspection robots of microgrid power facilities. The specific module connections are as follows: an edge training unit is deployed locally in each independent microgrid. The inspection robot deployed in each microgrid collects the original inspection data of the corresponding microgrid power facilities, and the edge training unit obtains a basic diagnostic model based on the data.

[0020] As an exemplary embodiment of the present invention, the specific method for obtaining the basic diagnostic model includes: the inspection robot performs multimodal data collection on the power facilities in the microgrid according to the preset inspection route to obtain the original inspection dataset, wherein the original inspection data includes visible light images, infrared thermal imaging data, partial discharge ultrasonic signals and environmental temperature and humidity data.

[0021] Specifically, the inspection robot periodically captures raw inspection datasets by using a pre-set three-dimensional inspection route covering power facilities such as transformers, switchgear, and power transmission cables, and utilizing its onboard visible light camera, infrared thermal imager, and ultrasonic sensor.

[0022] The edge training unit preprocesses and extracts features from the original inspection dataset, and establishes a feature association model. It uses environmental temperature and humidity data as an external correction factor and jointly labels it with visible light images, infrared thermal imaging data, and partial discharge ultrasonic signals to build a training label library.

[0023] Specifically, the preprocessing and feature extraction include: (1) For visible light images: performing histogram equalization to redistribute the pixel gray levels of the image, thereby enhancing the global contrast of the image and making low-contrast defects such as small cracks and rust on the surface of the facility easier for the neural network to identify. At the same time, to ensure the quality of data acquisition, an edge detection operator is applied to identify the core outline of the power facility in the image and calculate the offset distance between its centroid and the center of the image. If the distance exceeds the preset tolerance threshold of 10% of the image area, it is determined to be a target deviation, and the gimbal adaptive compensation logic is automatically triggered to adjust the pitch and yaw angles of the camera until the main body of the target facility is re-stabilized in the center area of ​​the image.

[0024] (2) For infrared thermal imagers: The collected temperature dot matrix data is mapped into a standardized temperature matrix and compared with the ambient reference temperature obtained by the ambient temperature and humidity sensor. All pixels with temperature values ​​higher than the sum of the ambient reference temperature and the preset safe temperature rise threshold are selected to form abnormal hotspot candidate areas. Furthermore, through an affine transformation algorithm, the coordinates of the abnormal hotspot candidate areas are aligned and fused with the synchronously acquired visible light image at the pixel level, so that the invisible thermal anomaly information is intuitively superimposed on the physical entity of the power facility equipment.

[0025] (3) For partial discharge ultrasound signals: Perform fast Fourier transform to convert the time-domain signal into a frequency-domain spectrum, thereby analyzing the distribution of its energy in different frequency ranges. By setting a dynamic background noise baseline, filter out environmental noise interference with relatively constant frequency, and extract suspected discharge signals with energy concentrated in a specific high-frequency band and significant sudden pulse characteristics.

[0026] The edge training unit calls the deep convolutional neural network operator, takes the preprocessed raw inspection data as input, and updates the weight values ​​of each layer of the network through the backpropagation algorithm until the preset loss function converges, thereby training a basic diagnostic model to identify surface defects, abnormal heating and insulation degradation of power facilities.

[0027] Specifically, the edge training unit calls the deep convolutional neural network operator, takes the preprocessed raw inspection data as input, calculates the prediction result through forward propagation, calculates the deviation between the predicted value and the label using the cross-entropy loss function, updates the weight values ​​of each layer of the network through the backpropagation algorithm until the loss function converges, and obtains a basic diagnostic model that can identify surface defects, abnormal heating and insulation degradation of power facilities.

[0028] During training iterations, the model calculates the predicted classification of the input samples through a forward propagation process and uses the cross-entropy loss function to quantify the deviation between the predicted result and the true label. The specific formula for calculating the cross-entropy loss function is as follows: ,in Let cross-entropy be the loss function. Independent encoded vectors representing the true labels of the samples. This represents the model's predicted probability vectors for each category of the sample. Based on the calculated loss function, a backpropagation algorithm is executed to calculate the gradient of the loss with respect to the weight parameters of each layer of the network, and the network weights are updated accordingly. This iterative process continues until the value of the loss function converges to a value below a preset convergence threshold. The stable state indicates that a basic diagnostic model capable of initially identifying faults such as surface defects, abnormal heating, and insulation deterioration in power facilities has been successfully trained locally.

[0029] Extract the model weight gradient, bias parameters, and feature mapping layer parameter information of the basic diagnostic model, encrypt them to obtain encrypted model incremental data, and upload them to the global aggregation node.

[0030] It should be noted that during this process, the original inspection data is strictly encapsulated within the local microgrid firewall and is not transmitted across regions or microgrids.

[0031] As an exemplary embodiment of the present invention, the specific method for obtaining the incremental model data includes: the edge training unit accesses the basic diagnostic model, locates the fully connected layer and feature mapping layer therein, and extracts the weight vector gradient matrix and bias term parameters corresponding to each layer.

[0032] Specifically, after the local basic diagnostic model completes one round of training, the edge training unit accesses the model structure programmatically through its internal interface, traversing to the feature mapping layer responsible for high-level feature abstraction and the fully connected layer that performs the final classification, accurately extracting the weight gradient matrix and bias term parameters of each layer.

[0033] Redundant parameters with weight change rates below a preset fluctuation threshold are removed by parameter filtering rules, while core incremental data reflecting common characteristics of facility failures are retained.

[0034] Specifically, to optimize communication efficiency, a parameter filtering process is initiated. This process does not transmit all parameters, but rather filters out the core incremental data that contributes most to model performance based on a preset fluctuation threshold. The parameter filtering rules are specifically expressed by the following formula: ,in A boolean result indicating whether the parameter is selected; The Euclidean norm of the gradient vector of a specific parameter layer, i.e. its magnitude or strength, is calculated by taking the square root of the sum of the squares of the elements in the gradient vector, and reflects the learning strength of the parameter layer in this round of local training. It is a pre-set fluctuation threshold used to measure the significance of learning, a scalar value set based on empirical data. It only applies when the parameter update magnitude... Exceeding the threshold Only then are they identified as core incremental data that carries the common characteristics of key faults and are retained.

[0035] The local encryption module is invoked to encrypt the filtered core incremental data bit by bit using the public key of the global aggregation node, and a unique identification code and timestamp information of the microgrid node are added to the header of the encrypted data packet.

[0036] Specifically, after the selection process is complete, the edge training unit invokes the local encryption module to perform asymmetric encryption on the selected core incremental data using the pre-distributed global aggregation node public key, generating an encrypted model incremental data packet. The header of this data packet automatically appends the microgrid node's unique identifier and the current system timestamp for global node authentication and data synchronization.

[0037] It should be noted that before data is uploaded, the edge training unit probes the quality of the local network link to ensure that its bandwidth and latency meet the preset transmission service level requirements before sending the data packets to the global aggregation node through the established TLS encrypted tunnel. During this process, the firewall deployed at the microgrid egress performs deep packet inspection on all outbound data streams. Using a predefined set of rules that match the original inspection data file format or stream characteristics, it intercepts any illegal data packets containing the original data fingerprint that attempt to be transmitted in real time, thus forming a solid barrier to ensure that data does not leave the local microgrid.

[0038] This invention effectively eliminates the risk of sensitive power facility data being intercepted or leaked during transmission by strictly limiting the processing of raw inspection data and model training to the local edge nodes of each microgrid and exchanging encrypted model parameter increments only between networks, thus solving the data security bottleneck faced by traditional centralized intelligent inspection.

[0039] The global aggregation node receives encrypted incremental model data uploaded from multiple microgrids, decrypts it, performs secure aggregation calculations, and generates a global diagnostic model. At the same time, a hierarchical model access and data auditing mechanism is established on the global aggregation node.

[0040] As an exemplary embodiment of the present invention, the specific generation method of the global diagnostic model includes: after receiving the encrypted model incremental data, the global aggregation node detects the contribution index of each edge training unit.

[0041] Specifically, after receiving the incremental model data uploaded and decrypted by each edge training unit, the global aggregation node does not directly perform aggregation, but instead initiates contribution evaluation, that is, it maintains a dynamic contribution index for each edge unit. It is derived from a comprehensive quantitative assessment of its historical data quality and participation stability, and its specific calculation formula is as follows: ,in This represents the average accuracy feedback of the historical uploaded model from the edge training unit on the global validation set, and is a value between 0 and 1. The normalized online duration of the edge training unit in the most recent evaluation period is the ratio of the actual online duration to the total duration of the period, and it is also between 0 and 1. The preset weighting coefficients, and This is used to adjust the relative importance of accuracy and stability.

[0042] The global aggregation nodes use a federated averaging algorithm, assigning different aggregation weight coefficients based on contribution indicators.

[0043] Specifically, after calculating the contribution index of all edge training units participating in this round of aggregation, the global aggregation node adopts the federated averaging algorithm to directly link the aggregation weight coefficient of each edge training unit with its contribution index. Edge training units with higher contribution have greater say in generating the global model.

[0044] At the same time, a contribution threshold such as 0.6 is set. If the contribution index of a certain edge training unit is lower than this threshold, the weight of its uploaded model increment in the aggregation process will be logically reduced or even set to zero to punish low-quality or unstable participants.

[0045] The global aggregation node performs abnormal gradient filtering.

[0046] Specifically, the cosine similarity between gradient vectors uploaded by every two edge training units is calculated. For any gradient vector, if its average cosine similarity with more than 50% of the gradient vectors of other edge training units is lower than a preset configurable deviation threshold such as 0.5, it is determined to be an abnormal gradient packet, and the abnormal gradient packet is automatically removed and does not participate in this round of aggregation calculation.

[0047] The aggregated weight gradient, bias parameters, and feature mapping layer parameters are loaded into the preset global model base framework to generate the initial global diagnostic model.

[0048] It should be noted that the global model's basic framework is consistent with the network structure of the basic diagnostic model, such as an 8-layer deep convolutional neural network structure containing convolutional layers, pooling layers, fully connected layers, and feature mapping layers.

[0049] Verify the integrity of the network structure and the matching of parameter dimensions in the initial draft of the global diagnostic model; if structural anomalies are detected, re-execute the allocation of aggregate weight coefficients until the model structure is compliant and generate the global diagnostic model.

[0050] In one specific example, the structural integrity includes the connection relationships between convolutional layers, pooling layers, and fully connected layers.

[0051] The structural anomalies include, but are not limited to, missing layer parameters and dimension mismatch.

[0052] As an exemplary embodiment of the present invention, the specific implementation of the model access and data auditing mechanism includes: defining a three-layer access control architecture within the global aggregation node, including a super administrator layer, a microgrid management layer, and a terminal device execution layer, and configuring a corresponding digital certificate for each layer.

[0053] Specifically, a three-tier access control architecture is first defined and instantiated within the global aggregation node. This architecture strictly distinguishes the operational boundaries of different roles: the top layer is the super administrator layer, the middle layer is the microgrid management layer, and the bottom layer is the terminal device execution layer. A unique digital certificate is issued to each legitimate user or device entity at each level, serving as the sole credential for identity authentication and permission binding.

[0054] It should be noted that the super administrator layer has the highest privileges and can perform core management operations such as directly modifying global model parameters, dynamically configuring audit policies, and approving the access of new microgrid nodes.

[0055] The microgrid management layer was granted limited permissions, allowing it only to access optimized diagnostic models issued for its region, query historical audit trails related to its region, and initiate aggregation requests to join or leave federated learning.

[0056] The function of the terminal device execution layer is strictly limited to downloading the specified model parameters after authentication and securely injecting them into the onboard computing platform of the inspection robot.

[0057] Establish a real-time audit pipeline to intercept and analyze all requests accessing the global aggregation node and record them as operation logs.

[0058] It should be noted that the operation log includes the identity of the request initiator, the request time, the operation type, the model version number involved, and the operation result.

[0059] Specifically, a real-time audit pipeline will be established as the mandatory gateway for all access requests, performing in-depth analysis and recording of each interaction. The pipeline will capture the identity of the request initiator, the request time accurate to milliseconds, the specific operation type (e.g., read or write), the model version number involved, and the final result of the operation (success or failure). The operation logs will be stored in encrypted form.

[0060] The audit module regularly compares the matching degree between the operation log and the permission table; when it detects unauthorized access or abnormally frequent model downloads, it immediately triggers an alarm and automatically blocks the access credentials of the corresponding node, while generating a security audit report for remote guidance personnel to review.

[0061] Specifically, the audit module automatically executes a consistency check procedure at a preset period, such as every hour, comparing recent operation logs with the fixed permission table. The consistency check can be expressed by the following formula: ,in The result is a Boolean verification result. This represents a specific request record vector, containing information such as the operator's identity and the type of operation. This represents the default permission configuration table.

[0062] function Its function is to determine the request. Is it in the permissions table? Within the scope of authorization. Once discovered. If the error is false, meaning unauthorized access is detected, or if a node initiates more than a preset frequency threshold for model download requests within a short period of time (e.g., 30 minutes), an alarm will be triggered immediately, and the node's access credentials will be automatically added to the blocking list. At the same time, relevant logs will be integrated to generate a detailed security audit report, which will be pushed to the remote supervisor's terminal for analysis.

[0063] The parameters of the global diagnostic model are distributed to the edge training units of each microgrid. Based on the distributed global parameters and combined with new local inspection data, the model is fine-tuned and optimized to form a local optimized diagnostic model.

[0064] As an exemplary embodiment of the present invention, the specific method for forming the locally specific optimized diagnostic model includes: the global aggregation node pushes the global diagnostic model parameters to the cache area of ​​each node through an asynchronous distribution strategy based on the online status of each microgrid edge training unit.

[0065] Specifically, the global aggregation node actively monitors the online status of each microgrid edge training unit and adopts an asynchronous distribution strategy to push the latest generated global diagnostic model parameter package to the dedicated cache of each target node in a non-blocking manner.

[0066] After receiving the update instruction, the edge training unit starts the integrity verification logic, compares the hash value of the parameter packet, and after confirming that the transmission process is correct, loads the global parameters into the parameter dictionary of the local diagnostic model.

[0067] Specifically, once the edge training unit receives an update instruction, it immediately initiates the integrity verification logic for the model parameters. This integrity verification logic involves rigorously comparing the hash digest of the received parameter packet with the original hash value sent by the global aggregation node along with the packet. If the two are completely identical, it indicates that the integrity verification logic has passed, confirming that the data has not been tampered with or corrupted during transmission. After successful verification, the global parameters are loaded and overwritten into the parameter dictionary of the local diagnostic model.

[0068] The edge training unit detects newly added inspection data in the local database since the previous training cycle. If the amount of data reaches the scale threshold that triggers fine-tuning, the transfer learning process is started.

[0069] Specifically, the background service of the edge training unit continuously monitors the local database and counts the amount of newly collected and labeled inspection data since the end of the previous training cycle. If the amount of new data reaches a preset threshold sufficient to drive an effective learning cycle, such as more than 500 new samples, the transfer learning process is automatically triggered.

[0070] During the fine-tuning process, the parameters of the shallow feature extraction convolutional layer in the global diagnostic model are frozen, and only the parameters of the deep semantic recognition layer and the output layer are updated. Iterative training is performed using newly added local data to adjust the model's sensitivity to local facility models and lighting backgrounds, thereby generating a local-specific optimized diagnostic model.

[0071] Specifically, during fine-tuning, to preserve the universally applicable low-level features already learned by the global model, the parameters of the convolutional layers responsible for extracting basic visual elements such as edges and textures are frozen and kept unchanged during training. Simultaneously, learning resources are focused on updating the parameters of the deep semantic recognition layer and the final output layer, which are highly relevant to the semantics of specific scenes. Iterative training is performed in small batches using locally added data and a relatively low learning rate. This approach aims to fine-tune the model to make it more sensitive to the specific power facility models, lighting conditions, and background environments unique to this microgrid, ultimately generating a high-performance and highly customized locally optimized diagnostic model.

[0072] As an exemplary embodiment of the present invention, the fine-tuning and optimization process further includes a local verification step, which is specifically implemented as follows: after generating the optimized diagnostic model, the edge training unit automatically retrieves the reserved local verification set to evaluate the inference performance of the model and calculates the accuracy, recall and false alarm rate of the model in defect identification.

[0073] Specifically, after the optimized diagnostic model is generated, the edge training unit immediately and automatically retrieves data from a locally reserved validation set that is completely isolated from the training set, and performs a comprehensive inference performance evaluation. This evaluation process calculates the model's key performance indicators in identifying various power facility defects, including accuracy, recall, and false alarm rate.

[0074] Compare the performance of the optimized diagnostic model with the old local model on the validation set. If the performance gain exceeds the preset confidence threshold, the model is marked as the current available version and synchronized to the model execution library of the inspection robot.

[0075] If the performance does not meet the target, the fine-tuning process is repeated by adjusting the learning rate step size or increasing the weight of local samples until the model output tends to stabilize.

[0076] Specifically, performance metrics such as accuracy, recall, and false alarm rate are quantitatively compared with the historical performance of the old local model deployed on the robot on the same validation set. A performance gain confidence threshold is set, and the new model is only considered successful if its performance metrics meet the requirements. Greater than The model is considered qualified only when the following conditions are met: The performance gain of the new model compared to the old model can be expressed by the formula: The calculation shows that, among which These are the core evaluation metrics, such as the F1 score, measured on the validation set for the new and old models. This is a preset minimum performance improvement percentage, such as 5%, that the new model must achieve. Once verified, the optimized diagnostic model is marked as the current best available version and seamlessly synchronized to the inspection robot's onboard model execution library, ready for actual inspection tasks.

[0077] Conversely, if the performance gain fails to reach the preset confidence threshold, the fine-tuning is deemed invalid and automatically rolled back, while a new round of fine-tuning optimization is initiated. During this retry process, the system strategically adjusts hyperparameters, such as decreasing the learning rate step size by 10% or appropriately increasing the weight of misclassified local samples in the loss function calculation to guide the model to focus more on difficult samples. This loop continues until the newly generated model's output on the validation set stabilizes and meets the performance gain requirements.

[0078] This invention uses a federated learning mechanism to aggregate model knowledge from multiple microgrids with different geographical locations and environmental conditions, enabling each local model to learn from the fault characteristics and experiences of other nodes. In particular, the ability to identify rare or unique fault types is enhanced, which is conducive to building a more adaptable and more accurate global intelligent model.

[0079] Inspection robots in each microgrid use optimized diagnostic models to diagnose faults; through periodic execution, each microgrid can collaboratively improve the accuracy and generalization ability of the fault diagnosis model without sharing the original data.

[0080] As an exemplary embodiment of the present invention, the specific implementation of the fault diagnosis includes: the inspection robot receives the optimized diagnostic model synchronized by the edge training unit and loads it into the vehicle-mounted embedded computing platform.

[0081] Specifically, the inspection robot receives the optimized diagnostic model marked as the latest available version through its dedicated communication link with the edge training unit and loads it into its built-in high-performance vehicle-mounted embedded computing platform.

[0082] During the inspection process, the inspection robot acquires real-time status data of power facilities through sensors, inputs it into the diagnostic model, and outputs a classification probability distribution vector of the status of the power facilities.

[0083] Specifically, during the execution of the preset inspection route, the robot uses its sensor array to capture the status data of the power facilities in real time and feeds it into the diagnostic model as input. After forward propagation calculation, the model outputs a classification probability distribution vector describing the status of the facilities in real time.

[0084] If the highest value in the classification probability is greater than the preset safety threshold, the corresponding fault type or normal state conclusion is output; if the highest probability value is lower than the preset safety threshold, it is judged as a suspected fault, triggering a high-definition capture command and uploading it to the edge training unit for manual assistance in judgment.

[0085] Specifically, a diagnostic logic judgment mechanism is built in, the core of which is to compare the maximum probability value in the classification probability distribution vector. With a preset security threshold This threshold is typically set above 0.95 to ensure high reliability. If Greater than or equal to If the probability is high enough, the diagnosis is considered to have sufficient confidence, and the fault type or normal state conclusion corresponding to the maximum probability value is directly output. Conversely, if... Below If the state is identified as a suspected malfunction, the high-definition camera unit on the robot is immediately triggered to execute precise focusing and high-definition capture commands, and the data packet containing the original data and the model's preliminary judgment results is uploaded to the edge training unit for manual assistance in the judgment.

[0086] For the determined fault status, the inspection robot calculates the fault severity level according to the preset logic rules and sends an alarm message containing the fault coordinates, fault type and on-site evidence map to the remote management terminal via wireless link.

[0087] Specifically, for conditions identified as faults by models or manual inspection, the inspection robot will automatically calculate a quantified fault severity level based on its built-in fault knowledge base, combined with the fault type and key parameters. This quantified level will then be pushed in real-time to a remote centralized monitoring and management terminal via a 4G or 5G wireless link. The message is comprehensive, encapsulating the precise geographical coordinates of the faulty facility, the identified fault type, the assessed severity level, and high-resolution images of the scene.

[0088] As an exemplary embodiment of the present invention, the specific implementation of the collaborative improvement of the accuracy and generalization capability of the fault diagnosis model includes: the global aggregation node is equipped with a global synchronization timer, and when the set aggregation period is reached, it sends a model reporting invitation to all registered microgrid edge training units.

[0089] Specifically, a configurable global synchronization timer is deployed within the global aggregation node. This timer, based on a preset aggregation cycle, typically 24 hours, serves as the core metronome for the entire federated learning system. When the timer completes a full cycle, it automatically broadcasts a standardized model invitation command to all online microgrid edge training units in its registration list via a secure channel.

[0090] The edge training unit responds to the invitation, assesses the local data update status, and if the update conditions are met, it automatically executes the parameter extraction, encryption, and upload process, starting a new round of global aggregation.

[0091] Specifically, upon receiving an invitation, the edge training unit executes a self-evaluation procedure for its local data update status. The core of this procedure is determining whether the amount of newly added, labeled, and valid inspection data since the last successful participation in aggregation has reached the minimum contribution threshold sufficient to drive a meaningful model update. This upload decision condition... This can be expressed by the following formula: ,in This represents the number of newly added valid samples in the local database. This is a configurable threshold parameter used to ensure that uploaded model increments are based on sufficiently new knowledge. If If the value is 1, it means that the update condition is met, and the edge training unit will automatically start the complete parameter extraction, encryption and uploading process, and officially participate in the new round of global aggregation.

[0092] It should be noted that by continuously repeating this closed-loop process of synchronous invitation and responsive upload at a set frequency, a continuous knowledge-sharing flow is constructed. Each microgrid edge node can thus indirectly learn and integrate fault characteristic knowledge from other heterogeneous environment nodes without exposing its original local inspection images and sensitive power operation parameters. This mechanism enables the local model to handle fault samples that are rare or have never appeared in its region, thereby significantly improving its overall accuracy and robustness in intelligent inspections in complex and variable power grid environments.

[0093] As an exemplary embodiment of the present invention, the system further includes a remote guidance step, which is specifically implemented as follows: when the highest probability value in the classification probability distribution vector output by the optimized diagnostic model loaded by the inspection robot is in a preset uncertainty range, the remote guidance process is automatically triggered, and a real-time streaming media channel with the global aggregation node is automatically established.

[0094] It should be noted that the uncertainty interval is defined as: ignoring the threshold < < Safety threshold; where the safety threshold ≥ 0.95 is used to determine high-confidence diagnostic results; the ignore threshold ≤ 0.7 is used to determine no fault or invalid data, both of which are configurable parameters.

[0095] The global aggregation node pushes the abnormal feature data uploaded by the inspection robot to the expert diagnosis terminal, and simultaneously retrieves the case analysis results of similar features from the knowledge base as decision support.

[0096] It should be noted that the abnormal feature data includes the raw inspection data of power facilities collected in real time.

[0097] The knowledge base is composed of anonymized historical fault cases from each microgrid. Three to five highly similar cases are selected using feature similarity matching algorithms, such as cosine similarity, to serve as an auxiliary basis for expert decision-making.

[0098] The historical fault cases include abnormal characteristics, manual judgment results, and handling solutions.

[0099] Remote experts issue secondary inspection commands through an interactive interface and receive manual judgment results after the inspection.

[0100] It should be noted that the types of detection commands include: (1) attitude adjustment command: control the tilt / yaw angle of the robot gimbal, move the inspection position, and optimize the shooting angle of the fault area; (2) high-precision sampling command: start the high-resolution mode of the sensor to improve the image clarity, temperature sampling accuracy or ultrasonic signal sampling frequency.

[0101] The manual judgment results are labeled as standard tags and sent back to the local edge training unit for subsequent incremental model learning.

[0102] This invention reduces reliance on manual inspections and improves operational efficiency by seamlessly integrating the entire process of data acquisition, model training, global aggregation, local optimization, autonomous diagnosis, and remote guidance. Furthermore, its intelligent hierarchical diagnosis and alarm mechanism, as well as the logic of introducing remote expert guidance when the model is uncertain, ensures the high efficiency and reliability of inspection tasks.

[0103] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0104] Those skilled in the art will recognize that the algorithmic steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0105] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0106] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0107] Finally, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A collaborative inspection and remote guidance system for intelligent inspection robots of microgrid power facilities, characterized in that: include: Edge training units are deployed locally in each independent microgrid. Inspection robots deployed in each microgrid collect raw inspection data of the corresponding microgrid power facilities, and the edge training units train a basic diagnostic model based on this data. Extract the model weight gradient, bias parameters, and feature mapping layer parameter information of the basic diagnostic model, encrypt them to obtain encrypted model incremental data, and upload them to the global aggregation node; The global aggregation node receives encrypted incremental model data uploaded from multiple microgrids, decrypts it, performs secure aggregation calculations, and generates a global diagnostic model. At the same time, a hierarchical model access and data auditing mechanism is established on the global aggregation node. The parameters of the global diagnostic model are distributed to the edge training units of each microgrid, and the model is fine-tuned and optimized based on the distributed global parameters and combined with new local inspection data to form a local exclusive optimized diagnostic model. Inspection robots in each microgrid use optimized diagnostic models to diagnose faults; through periodic execution, each microgrid can collaboratively improve the accuracy and generalization ability of the fault diagnosis model without sharing the original data.

2. The microgrid power facility intelligent inspection robot collaborative inspection and remote guidance system according to claim 1 is characterized in that: The specific methods for obtaining the basic diagnostic model include: The inspection robot collects multimodal data on the power facilities in the microgrid according to the preset inspection route to obtain the raw inspection dataset, which includes visible light images, infrared thermal imaging data, partial discharge ultrasonic signals, and environmental temperature and humidity data. The edge training unit preprocesses and extracts features from the original inspection dataset and establishes a feature association model. It uses environmental temperature and humidity data as an external correction factor and jointly labels it with visible light images, infrared thermal imaging data, and partial discharge ultrasonic signals to build a training label library. The edge training unit calls the deep convolutional neural network operator, takes the preprocessed raw inspection data as input, and updates the weight values ​​of each layer of the network through the backpropagation algorithm until the preset loss function converges, thereby training a basic diagnostic model to identify surface defects, abnormal heating and insulation degradation of power facilities.

3. The intelligent inspection robot collaborative inspection and remote guidance system for microgrid power facilities according to claim 1, characterized in that: The specific methods for obtaining the incremental data of the model include: The edge training unit accesses the basic diagnostic model, locates the fully connected layer and feature mapping layer, and extracts the weight vector gradient matrix and bias term parameters corresponding to each layer. Redundant parameters with weight change rates lower than a preset fluctuation threshold are removed by parameter filtering rules, while core incremental data reflecting common characteristics of facility failures are retained. The local encryption module is invoked to encrypt the filtered core incremental data bit by bit using the public key of the global aggregation node, and a unique identification code and timestamp information of the microgrid node are added to the header of the encrypted data packet.

4. The intelligent inspection robot collaborative inspection and remote guidance system for microgrid power facilities according to claim 1, characterized in that: The specific methods for generating the global diagnostic model include: After receiving incremental data from the encrypted model, the global aggregation node detects the contribution index of each edge training unit. The global aggregation nodes use a federated averaging algorithm, assigning different aggregation weight coefficients based on contribution indicators; Global aggregation nodes perform abnormal gradient filtering; The aggregated weight gradient, bias parameters, and feature mapping layer parameters are loaded into the preset global model base framework to generate the initial global diagnostic model. Verify the integrity of the network structure and the matching of parameter dimensions in the initial draft of the global diagnostic model; if structural anomalies are detected, re-execute the allocation of aggregate weight coefficients until the model structure is compliant and generate the global diagnostic model.

5. The intelligent inspection robot collaborative inspection and remote guidance system for microgrid power facilities according to claim 4, characterized in that: The specific implementation methods of the model access and data auditing mechanism include: A three-tier access control architecture is defined within the global aggregation node, including the super administrator layer, the microgrid management layer, and the terminal device execution layer, and a corresponding digital certificate is configured for each layer. Establish a real-time audit pipeline to intercept and analyze all requests that access the global aggregation node, and record them as operation logs; The audit module regularly compares the matching degree between the operation log and the permission table; when it detects unauthorized access or abnormally frequent model downloads, it immediately triggers an alarm and automatically blocks the access credentials of the corresponding node, while generating a security audit report for remote guidance personnel to review.

6. The intelligent inspection robot collaborative inspection and remote guidance system for microgrid power facilities according to claim 1, characterized in that: The specific methods for forming the locally specific optimized diagnostic model include: The global aggregation node pushes the global diagnostic model parameters to the cache of each node through an asynchronous distribution strategy based on the online status of each microgrid edge training unit. After receiving the update instruction, the edge training unit starts the integrity verification logic, compares the hash value of the parameter packet, and after confirming that the transmission process is correct, loads the global parameters into the parameter dictionary of the local diagnostic model. The edge training unit detects newly added inspection data in the local database since the previous training cycle. If the amount of data reaches the scale threshold that triggers fine-tuning, the transfer learning process is started. During the fine-tuning process, the parameters of the shallow feature extraction convolutional layer in the global diagnostic model are frozen, and only the parameters of the deep semantic recognition layer and the output layer are updated. Iterative training is performed using newly added local data to adjust the model's sensitivity to local facility models and lighting backgrounds, thereby generating a local-specific optimized diagnostic model.

7. The intelligent inspection robot collaborative inspection and remote guidance system for microgrid power facilities according to claim 6, characterized in that: The fine-tuning and optimization process also includes a local verification step, the specific implementation of which includes: After generating the optimized diagnostic model, the edge training unit automatically retrieves the reserved local validation set to evaluate the inference performance of the model and calculates the model's accuracy, recall, and false alarm rate in defect identification. Compare the performance of the optimized diagnostic model with the old local model on the validation set. If the performance gain exceeds the preset confidence threshold, the model is marked as the current available version and synchronized to the model execution library of the inspection robot. If the performance does not meet the target, the fine-tuning process is repeated by adjusting the learning rate step size or increasing the weight of local samples until the model output tends to stabilize.

8. The intelligent inspection robot collaborative inspection and remote guidance system for microgrid power facilities according to claim 1, characterized in that: The specific implementation methods of the fault diagnosis include: The inspection robot receives the optimized diagnostic model synchronized by the edge training unit and loads it into the vehicle-mounted embedded computing platform; During the inspection process, the inspection robot acquires real-time data on the status of power facilities through sensors, inputs it into the diagnostic model, and outputs a classification probability distribution vector of the status of the power facilities. If the highest value in the classification probability is greater than the preset safety threshold, the corresponding fault type or normal state conclusion is output; if the highest probability value is lower than the preset safety threshold, it is judged as a suspected fault, triggering a high-definition capture command and uploading it to the edge training unit for manual assistance in judgment. For the determined fault status, the inspection robot calculates the fault severity level according to the preset logic rules and sends an alarm message containing the fault coordinates, fault type and on-site evidence map to the remote management terminal via wireless link.

9. The intelligent inspection robot collaborative inspection and remote guidance system for microgrid power facilities according to claim 8, characterized in that: The specific implementation methods for collaboratively improving the accuracy and generalization ability of the fault diagnosis model include: The global aggregation node has a built-in global synchronization timer. When the set aggregation period is reached, it sends a model reporting invitation to all registered microgrid edge training units. The edge training unit responds to the invitation, assesses the local data update status, and if the update conditions are met, it automatically executes the parameter extraction, encryption, and upload process, starting a new round of global aggregation.

10. The intelligent inspection robot collaborative inspection and remote guidance system for microgrid power facilities according to claim 1, characterized in that: The system also includes a remote guidance step, the specific implementation of which includes: When the highest probability value in the classification probability distribution vector output by the optimized diagnostic model loaded by the inspection robot is in the preset uncertainty range, the remote guidance process is automatically triggered, and a real-time streaming media channel with the global aggregation node is automatically established. The global aggregation node pushes the abnormal feature data uploaded by the inspection robot to the expert diagnosis terminal, and simultaneously retrieves the case analysis results of similar features in the knowledge base as decision support; Remote experts issue secondary inspection commands through an interactive interface and receive manual judgment results after the inspection.

Images