Photovoltaic inspection robot intelligent operation and maintenance method and system and electronic equipment

Abstract

The present application relates to the field of data processing, and provides a photovoltaic inspection robot intelligent operation and maintenance method and system and electronic equipment, through differentiated extraction of data of different sensors, only local key information is retained, transmission of invalid information is reduced, protocol unification and timestamp alignment are carried out to avoid invalidity of association between data, then on the basis of analysis and identification, an inspection path is generated through a dynamic inspection priority queue and a dual-mode path planning algorithm, and finally, according to the path and an operation and maintenance task queue, inspection and operation and maintenance are carried out in sequence. The present application can reduce transmission and association of invalid information, reduce calculation pressure and communication occupation of system hardware, and improve system response speed, on the basis of which, the system can realize instant response to different situations by using a dynamic inspection priority queue and a dual-mode path planning, and improve inspection efficiency, accuracy and effectiveness.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to intelligent operation and maintenance methods, systems and electronic devices for photovoltaic inspection robots. Background Technology

[0002] In the traditional field of photovoltaic (PV) inspection, taking the common track-mounted inspection robot as an example, its body is equipped with multiple sensors to achieve real-time monitoring and perception of equipment and the environment inside the power distribution room. It can detect indoor temperature and humidity, and effectively monitor toxic and harmful gases. The infrared thermal imager on the robot can detect the temperature of the power distribution equipment, automatically issuing alarms in cases of overheating or high temperatures. The robot can also perform partial discharge detection on the power distribution equipment in the power distribution room, and the detection results are uploaded to the AIoT robot inspection platform in real time. Image acquisition devices capture images of PV modules, and then the data is transmitted back to the control center for simple analysis to determine whether there are obvious damages or other problems with the PV modules. In terms of movement, it relies on a preset track or a simple path planning algorithm to move within the PV power station.

[0003] In existing technologies, to ensure the accuracy and effectiveness of inspections, real-time inspection path planning may be employed, typically achieved by improving pathfinding algorithms to fully consider the actual inspection paths required in different situations. However, this approach is ineffective in real-time pathfinding during emergencies. When faced with complex environments and large amounts of data, the poor targeting of data processing in existing technologies leads to the repeated processing and transmission of a large amount of invalid information, impacting system response speed. For example, directly transmitting and analyzing all sensor data between modules, directly integrating different types of data, or directly determining inspection paths based on path planning algorithms all waste significant hardware resources, resulting in a slow system response. Even with highly accurate information about the external situation, a slow system processing speed makes it difficult to adjust in real-time according to the actual situation, thus affecting the accuracy and effectiveness of identification.

[0004] Therefore, to ensure the accuracy and effectiveness of inspections, it is necessary to address the problem from multiple dimensions simultaneously. Summary of the Invention

[0005] To address the problem of low accuracy and effectiveness in photovoltaic inspections due to the repeated processing and transmission of a large amount of invalid information in existing technologies, this invention provides an intelligent operation and maintenance method, system, and electronic equipment for photovoltaic inspection robots. By differentially extracting data from different sensors, only key local information is retained, reducing the transmission of invalid information. Protocol standardization and timestamp alignment are used to avoid data correlation failures. Then, based on analysis and identification, an inspection path is generated using a dynamic inspection priority queue and a dual-mode path planning algorithm. Finally, inspections and maintenance are performed sequentially according to the path and the maintenance task queue. This invention reduces the transmission and correlation of invalid information, lowers the computational pressure and communication overhead of the system hardware, and improves system response speed. Based on this, the system can utilize the dynamic inspection priority queue and dual-mode path planning to achieve real-time responses to different situations, improving inspection efficiency, accuracy, and effectiveness.

[0006] The following is the technical solution of the present invention.

[0007] This invention provides an intelligent operation and maintenance method for photovoltaic inspection robots, comprising:

[0008] Image data, temperature data, and environmental data of photovoltaic modules in photovoltaic power plants are collected and extracted differentially, retaining only key local information;

[0009] Standardized data is obtained by unifying the protocols and aligning the timestamps of various types of key information from the data acquisition module through a standardized data interface.

[0010] Defect identification and operational status analysis are performed on standardized data to obtain analysis results;

[0011] Acquire photovoltaic power plant layout data, combine analysis results, environmental data and time synchronization information, and generate inspection paths through dynamic inspection priority queue and dual-mode path planning algorithm;

[0012] An operation and maintenance task queue is generated based on the analysis results, and the entire process data is stored at the same time;

[0013] The robot body performs movement and maintenance operations based on the inspection path and maintenance task queue.

[0014] In this invention, data dimensionality is reduced by differential extraction, and transmission volume is compressed synchronously to reduce communication load; data spatiotemporal consistency and parsing efficiency are ensured by protocol unification and timestamp alignment to avoid association failure; on this basis, path reconstruction is achieved in special cases through dynamic inspection priority queue and dual-mode path planning algorithm to improve inspection efficiency and targeting.

[0015] As one possible implementation, the process involves collecting image data, temperature data, and environmental data of the photovoltaic modules in the photovoltaic power station, and then performing differentiated extraction to retain only key local information, including:

[0016] By using the feature saliency scoring function, we can extract the Region of Interest (ROI) by selecting regions with higher weights and retaining only local images containing potential defects. At the same time, we can use a sliding window filtering algorithm to extract temperature and environmental data.

[0017] In this invention, data from different sensor sources is extracted in different ways, retaining only key local information and reducing the processing and transmission of invalid data.

[0018] As one possible implementation, the standardized data is obtained by unifying the protocols and aligning the timestamps of various types of key information from the data acquisition module through a standardized data interface, including:

[0019] Define a pre-defined data structure for image, temperature, and environmental data. The structure includes data type labels, acquisition device ID, raw values, and quality factors, where the quality factors reflect the completeness and reliability of the current data.

[0020] By connecting to a precision clock synchronization system, the sampling time is dynamically calibrated using the difference between the master clock broadcast signal and the local timing of the slave nodes, ensuring that the time error of all sampling nodes in the system is controlled within the preset error. Each frame of data is bound with a unique timestamp when it is generated. By minimizing the time alignment cost function between different data sequences, the distance matrix is ​​traversed through dynamic programming to backtrack to find the optimal alignment path. After alignment, the result is constrained with the target error window to ensure the consistency of multi-source data in the time dimension.

[0021] This invention addresses the heterogeneity of multi-source data protocols by standardizing data interfaces and using timestamp alignment mechanisms to uniformly define a data format containing "data type label, device ID, original value, and quality factor," thereby improving the data parsing efficiency of each module. Furthermore, by quantifying data integrity through quality factors, the system can automatically filter low-reliability data, preventing invalid data from entering the processing flow and improving data reliability from the source, reducing the risk of subsequent misjudgments. Through a master-slave node dynamic calibration mechanism, the Dynamic Time Warping (DTW) algorithm minimizes the time alignment cost, controlling the sampling time error across the entire system and ensuring strict alignment of multi-source data (such as images, temperature, and environmental parameters) in the time dimension. Ultimately, this avoids data correlation failures caused by frequency differences, improving the accuracy of subsequent analysis of multi-physics coupling faults.

[0022] As one possible implementation, the defect identification and operational status analysis of standardized data to obtain analysis results includes:

[0023] A lightweight convolutional neural network model optimized based on knowledge distillation is used to identify defects in image data of photovoltaic modules.

[0024] A fusion structure of gated recurrent unit and Transformer encoder is adopted to extract long-term dependency information. Dynamic weights are assigned to different feature dimensions through multi-head attention mechanism for runtime state analysis.

[0025] Among them, defect identification and operation status analysis are based on a distributed computing architecture. Image recognition tasks and status analysis tasks are processed by independent computing nodes. The task allocation between nodes is dynamically adjusted in real time through a load balancing algorithm. The load balancing algorithm improves processing efficiency by minimizing the overall task time.

[0026] As one possible implementation, the step of acquiring photovoltaic power plant layout data, combining analysis results, environmental data, and time synchronization information, and generating inspection paths through a dynamic inspection priority queue and a dual-mode path planning algorithm includes:

[0027] A dynamic inspection priority queue is determined based on the defect level and environmental risk level; the priority of each photovoltaic module is generated based on the defect level, environmental risk level and corresponding weights, and the weight of the defect level is greater than the weight of the environmental risk level.

[0028] Configure a dual-mode path planning engine, where:

[0029] In normal mode, an improved ant colony algorithm is used to construct a global path by combining the distribution of defect hotspots with future weather forecast data. In this algorithm, each ant makes decisions based on heuristic functions and pheromone intensity during the path selection process, and periodically and automatically re-evaluates the global path to ensure coverage of all potentially high-risk areas.

[0030] When at least one of the defect level or environmental risk level is at the highest level, switch to emergency mode. In emergency mode, a cost function is constructed based on real-time environmental data and component location information. The comprehensive cost of the current path node is calculated based on the actual cost from the starting node to the current node and the heuristic cost from the current node to the target node. Node expansion is performed with the goal of minimizing the comprehensive cost, and path planning is completed.

[0031] In this invention, the normal mode and emergency mode dynamically adjust the path to deal with different situations, and rely on low-latency data with high-precision timestamps to ensure the spatiotemporal consistency of path planning, prioritize the inspection of high-risk components, significantly enhance preventive maintenance capabilities, and reduce the robot's ineffective movement distance.

[0032] As one possible implementation, the step of obtaining the defect level includes:

[0033] Obtain the defect type and confidence level obtained from defect identification, and determine the defect level based on the confidence level.

[0034] As one possible implementation, if the generation of inspection paths is blocked, alternative inspection paths are generated with reference to prior knowledge, including: pre-selection by using the minimum path cost function based on the defect distribution density function. In the path cost function, the higher the defect density, the greater the path cost, thereby guiding path planning to avoid or prioritize the coverage of high-risk areas.

[0035] As one possible implementation, the step of generating an operation and maintenance task queue based on the analysis results, while storing full-process data, includes:

[0036] An operation and maintenance task queue is generated based on the defect level and processing time. The priority of each operation and maintenance task is generated based on the defect level, processing time and corresponding weight, with the weight of the defect level being greater than the weight of the processing time.

[0037] This invention also provides an intelligent operation and maintenance system for photovoltaic inspection robots, used to execute any of the above-mentioned intelligent operation and maintenance methods for photovoltaic inspection robots, including:

[0038] The data acquisition module is used to collect image data, temperature data, and environmental data of photovoltaic modules in the photovoltaic power station.

[0039] The data fusion and processing module is used to unify the protocols and align the timestamps of various types of key information from the data acquisition module through a standardized data interface to obtain standardized data.

[0040] The deep learning processing module is used to identify defects and analyze operational status on standardized data to obtain analysis results.

[0041] The path optimization module is used to acquire photovoltaic power plant layout data, combine analysis results, environmental data and time synchronization information, and generate inspection paths through a dynamic inspection priority queue and a dual-mode path planning algorithm.

[0042] The operations and maintenance management module is used to generate an operations and maintenance task queue based on the analysis results, and to store the entire process data.

[0043] The robot itself is used to perform movement and maintenance operations based on the inspection path and maintenance task queue.

[0044] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-described intelligent operation and maintenance method for photovoltaic inspection robots when it calls the computer program in the memory.

[0045] The present invention also provides a storage medium storing computer-executable instructions, which, when loaded and executed by a processor, implement the steps of the above-described intelligent operation and maintenance method for photovoltaic inspection robots.

[0046] The substantial effects of this invention include:

[0047] This invention reduces data dimensionality by differential extraction, compresses transmission volume synchronously, and reduces communication load; it ensures data spatiotemporal consistency and parsing efficiency by unifying protocols and aligning timestamps, avoiding association failures; and on this basis, it improves inspection efficiency and targeting by implementing path reconstruction in special cases through dynamic inspection priority queues and dual-mode path planning algorithms. Attached Figure Description

[0048] Figure 1 This is a flowchart of an embodiment of the present invention;

[0049] Figure 2 This is a system block diagram of an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions will be clearly and completely described below in conjunction with the embodiments. 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.

[0051] It should be understood that in the various embodiments of the present invention, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0052] It should be understood that in this invention, "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0053] It should be understood that in this invention, "multiple" refers to two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "Contains A, B, and C", "Contains A, B, and C" means that all three A, B, and C are contained; "Contains A, B, or C" means that one of A, B, and C is contained; "Contains A, B, and / or C" means that any one, two, or three of A, B, and C are contained.

[0054] The technical solution of the present invention will be described in detail below with reference to specific embodiments. Embodiments may be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0055] Example: Intelligent operation and maintenance method for photovoltaic inspection robots, such as Figure 1 As shown, it includes:

[0056] Image data, temperature data, and environmental data of photovoltaic modules in photovoltaic power plants are collected and extracted differentially, retaining only key local information;

[0057] Standardized data is obtained by unifying the protocols and aligning the timestamps of various types of key information from the data acquisition module through a standardized data interface.

[0058] Defect identification and operational status analysis are performed on standardized data to obtain analysis results;

[0059] Acquire photovoltaic power plant layout data, combine analysis results, environmental data and time synchronization information, and generate inspection paths through dynamic inspection priority queue and dual-mode path planning algorithm;

[0060] An operation and maintenance task queue is generated based on the analysis results, and the entire process data is stored at the same time;

[0061] The robot body performs movement and maintenance operations based on the inspection path and maintenance task queue.

[0062] This embodiment also provides an intelligent operation and maintenance system for photovoltaic inspection robots, used to execute the above-described intelligent operation and maintenance method for photovoltaic inspection robots, such as... Figure 2 As shown, it includes:

[0063] The data acquisition module is used to collect image data, temperature data, and environmental data of photovoltaic modules in the photovoltaic power station.

[0064] The data fusion and processing module is used to unify the protocols and align the timestamps of various types of key information from the data acquisition module through a standardized data interface to obtain standardized data.

[0065] The deep learning processing module is used to identify defects and analyze operational status on standardized data to obtain analysis results.

[0066] The path optimization module is used to acquire photovoltaic power plant layout data, combine analysis results, environmental data and time synchronization information, and generate inspection paths through a dynamic inspection priority queue and a dual-mode path planning algorithm.

[0067] The operations and maintenance management module is used to generate an operations and maintenance task queue based on the analysis results, and to store the entire process data.

[0068] The robot itself is used to perform movement and maintenance operations based on the inspection path and maintenance task queue.

[0069] Another communication module enables data transmission between modules, integrating a data buffer and a hardware acceleration interface.

[0070] As one possible implementation, the process involves collecting image data, temperature data, and environmental data of photovoltaic modules in the photovoltaic power station, and then performing differentiated extraction to retain only key local information, including:

[0071] By employing multiple types of sensors working collaboratively, efficient acquisition and preliminary processing of key operational data from photovoltaic power plants are achieved. The image acquisition unit consists of multiple high-definition cameras, and the edge computing unit integrates an attention mechanism to enhance the intelligence of image data processing and transmission efficiency. This attention mechanism focuses on potential defect areas in the image through a feature saliency scoring function, and its core is a weighted feature extraction formula:

[0072] ;

[0073] in, This represents the attention weight of the i-th region. Let be the feature response value of the region, and n be the number of regions in the image. The system extracts Regions of Interest (ROIs) by selecting regions with higher weights, retaining only local images containing potential defects, significantly reducing image data redundancy and transmission overhead. The temperature acquisition unit uses an infrared temperature sensor to capture surface thermal information of the photovoltaic modules, while the environmental acquisition unit includes sensors for light, wind speed, and humidity to collect key environmental factors affecting photovoltaic operation. The edge computing unit uses a sliding window filtering algorithm on temperature and environmental data to improve data stability; its mathematical expression is:

[0074] ;

[0075] in, This represents the smoothed output value at time t. Here, represents the original sampled values, and w is the window width. The filtered data is further subjected to principal component analysis (PCA) for dimensionality compression to retain key information and improve subsequent processing efficiency. The core transformation formula for PCA is:

[0076] ;

[0077] Where X is the standardized data matrix, W is the principal component direction matrix obtained from eigenvalue decomposition, and Z is the dimensionality-reduced feature matrix. To further constrain the compression ratio, the edge computing unit determines the number of principal components to be selected by accumulating the variance contribution rate, and selects the minimum number of principal components k that satisfies the following conditions:

[0078] ;

[0079] in, This represents the i-th feature value, where m is the total number of original dimensions. The variance contribution rate threshold is set between 0.2 and 0.3 to ensure that the compressed data retains its core features. All data that has undergone the above preprocessing is sent to the communication module via a standardized data bus, and then forwarded by the communication module to the data fusion processing module for protocol matching and time alignment to ensure the real-time performance and consistency of the system-level data stream.

[0080] In this embodiment, the attention-based ROI extraction technique has significant advantages for image data. Traditional image data transmission often contains a large amount of redundant information, while this embodiment only transmits potentially defective local images, greatly reducing invalid data transmission. This not only reduces the pressure on data transmission bandwidth but also allows subsequent processing modules to focus on key areas, improving processing efficiency. Taking a medium-sized photovoltaic power station as an example, the amount of image data collected daily can reach several gigabytes. After adopting this technology, the amount of data transmitted can be reduced by about 80%, effectively alleviating the burden on network transmission. For temperature and environmental data, the sliding window filtering combined with principal component analysis can reduce data dimensionality while retaining key information. The sliding window filtering algorithm removes noise interference through smoothing of time series data, making the data more stable and reliable. Principal component analysis further extracts the main features of the data, reducing data dimensionality by more than 50% while ensuring data quality, thus improving the speed and efficiency of subsequent data processing.

[0081] As one possible implementation, the standardized data is obtained by unifying the protocols and aligning the timestamps of various types of key information from the data acquisition module through a standardized data interface, including:

[0082] The data fusion processing module undertakes the core task of standardizing and aligning multi-source heterogeneous data in terms of time. Through the collaborative work of the protocol conversion unit and the time synchronization unit, it ensures that the data received by the subsequent deep learning module remains consistent in structure and timing. The protocol conversion unit defines a unified JSON-formatted data structure for image, temperature, and environmental data. This structure includes data type tags, acquisition device IDs, original values, and a quality factor. The quality factor reflects the completeness and reliability of the current data, with a value range of [0,1]. This unified data structure improves the data parsing efficiency between processing units within the system and optimizes the transmission structure through field-level indexing. The time synchronization unit connects to the IEEE 1588 precision clock synchronization system, dynamically calibrating the sampling time using the difference between the master clock broadcast signal and the local timing of the slave nodes, ensuring that the time error of all sampling nodes in the system is controlled within 1 second. Within.

[0083] This embodiment, by adding high-precision timestamps to image, temperature, and environmental data and utilizing a synchronized clock system, can accurately ensure the consistency of multi-source data across the time dimension. In actual photovoltaic power plant operation, different types of sensors may collect data at different times. If the times are not synchronized, it can lead to serious errors in the analysis of photovoltaic module status. For example, when analyzing the relationship between temperature changes and power generation efficiency of photovoltaic modules, if the timestamps of temperature data and power generation data are inconsistent, incorrect conclusions may be drawn, affecting the operation and maintenance decisions of the power plant. This embodiment, through a time synchronization mechanism, controls time errors within a very small range, avoiding status analysis errors caused by time misalignment, and providing a solid guarantee for the stable operation of the system.

[0084] In this embodiment, each frame of data is bound with a unique timestamp at the time of generation, and the timestamp is represented as:

[0085] ;

[0086] in, This represents the timestamp of the i-th frame of data. The initial time of the master clock. The synchronization offset between this frame of data and the master clock is continuously updated using the bidirectional messaging mechanism in the IEEE 1588 protocol to maintain synchronization accuracy. To eliminate timing drift caused by different sampling frequencies of different types of sensors, the system introduces a Dynamic Time Warping (DTW) algorithm, the core of which is to minimize the time alignment cost function between different data sequences. The recursive formula for the DTW cumulative distance is:

[0087] ;

[0088] in, and Let represent the sampled values ​​of two different types of sensors at times i and j, respectively. To minimize the cumulative cost of alignment to this point, the formula uses dynamic programming to traverse the distance matrix and backtrack to find the optimal alignment path. To ensure strict control over time alignment errors, the DTW alignment result is constrained against the target error window after completion, satisfying the following conditions:

[0089] ;

[0090] in, and These are the timestamps of the two types of data after alignment. To mitigate time alignment errors, the system is set to 5ms, ensuring high consistency of multi-source data across time. Standardized data is pushed to the deep learning processing module via a message queue mechanism using the MQTT protocol, guaranteeing low-latency transmission while providing high-concurrency processing capabilities to support the real-time computational needs of the image recognition and state analysis units.

[0091] In this embodiment, a standardized data interface and timestamp alignment mechanism are used to uniformly define a data format containing "data type label, device ID, original value, and quality factor," resolving the heterogeneity issue of multi-source data protocols and improving the data parsing efficiency of each module. Furthermore, by quantifying data integrity through quality factors, the system can automatically filter low-reliability data, preventing invalid data from entering the processing flow, thus improving data reliability from the source and reducing the risk of subsequent misjudgments. Through a master-slave node dynamic calibration mechanism, the Dynamic Time Warping (DTW) algorithm is used to minimize the time alignment cost, controlling the sampling time error of the entire system and ensuring strict alignment of multi-source data (such as images, temperature, and environmental parameters) in the time dimension. Ultimately, this avoids data correlation failures caused by frequency differences, improving the accuracy of subsequent analysis of multi-physics coupling faults.

[0092] For example, in the event of a hot spot fault in a photovoltaic module, the time difference between infrared temperature data and image data can be precisely matched, avoiding fault location errors caused by time misalignment. Alternatively, when a wind speed sensor (high-frequency sampling) and a camera (low-frequency sampling) simultaneously capture module vibrations caused by strong winds, the system can accurately align their timing, avoiding data correlation failures caused by frequency differences and improving the accuracy of deep learning models in analyzing multi-physics coupled faults.

[0093] As one possible implementation, the defect identification and operational status analysis of standardized data to obtain analysis results includes:

[0094] The deep learning processing module, centered on parallel computing, achieves accurate analysis of image and time-series data through the collaborative work of the image recognition unit and the state analysis unit. The image recognition unit employs a lightweight convolutional neural network model optimized by knowledge distillation. This model uses ResNet-50 as the teacher model and constructs a loss function by minimizing the difference between the output distributions of the student and teacher models, compressing the number of parameters by over 60% and adapting to embedded GPU platforms. Its distillation loss function is defined as:

[0095] ;

[0096] in, This represents the student model's predicted labels. For real labels, and These are the logits outputs for the student model and the teacher model, respectively, where T is the temperature coefficient. For loss weighting coefficients, For cross-entropy loss, The model uses Kullback-Leibler divergence. It achieves real-time performance with a single-image inference latency of no more than 20ms on embedded platforms such as NVIDIA Jetson. The state analysis unit employs a fusion structure of a gated recurrent unit (GRU) and a Transformer encoder, enhancing the global awareness and multi-dimensional data expressiveness of time series modeling. The state update process of the GRU is as follows:

[0097] ;

[0098] ;

[0099] in, Currently in a hidden state. For the current input, To update the door, To reset the door, and This is the weight matrix. The Transformer encoder is used to extract long-term dependency information, and it assigns dynamic weights to different feature dimensions through a multi-head attention mechanism. Its calculation formula is:

[0100] ;

[0101] Where Q, K, and V are the query, key, and value matrices, respectively. Using the key vector dimension, this mechanism significantly enhances the model's selective attention to different data dimensions such as temperature, current, and humidity, achieving a state analysis latency of no more than 10ms. The entire deep learning processing module runs on a distributed computing architecture, with image recognition and state analysis tasks processed by independent computing nodes. Task allocation between nodes is dynamically adjusted in real time using a load balancing algorithm. The load balancing objective function is:

[0102] ;

[0103] in, Let be the computational complexity of the task currently pending at the i-th node. For the computing resources of this node, the load balancing mechanism improves processing efficiency by minimizing the overall task time. This parallel architecture increases the overall system throughput by more than 3 times, meeting the requirements of real-time data processing and large-scale deployment in photovoltaic inspection while ensuring identification accuracy.

[0104] This embodiment employs knowledge distillation technology for the CNN model, a crucial method for achieving model lightweighting while maintaining recognition accuracy. Taking the simplification of ResNet-50 to MobileNet-V3 as an example, knowledge distillation transfers knowledge from the teacher model (ResNet-50) to the student model (MobileNet-V3), maintaining high recognition accuracy even with a compression of over 60% of model parameters. This lightweight model is compatible with embedded GPU deployments, significantly reducing hardware costs and energy consumption. In practical applications, embedded devices have limited computing resources, making it difficult for traditional large CNN models to run efficiently on these devices. The optimized lightweight model, however, achieves real-time performance with a single-image inference latency of no more than 20ms on embedded GPUs, meeting the stringent real-time requirements of photovoltaic inspection. For the time-series analysis task of RNNs, a gating mechanism using Long Short-Term Memory (LSTM) networks is introduced for optimization, effectively reducing the gradient vanishing problem. Simultaneously, the parallel processing of multi-dimensional environmental data using a time-series prediction model significantly improves the speed of state analysis. When analyzing the operating status of photovoltaic modules, it can quickly and accurately capture the changing trends of data such as temperature, current, and environment, detect potential faults in advance, and provide a basis for timely maintenance.

[0105] Furthermore, by dynamically allocating computing resources through a load balancing algorithm, overload of a single node is avoided, thus improving the overall processing capacity of the system. In a large-scale photovoltaic power plant, the amount of data generated daily is enormous. If processing relies solely on a single computing node, the processing speed will be slow, failing to meet real-time requirements. Adopting a distributed computing architecture increases the system's throughput by more than three times, enabling the processing and analysis of large amounts of data in a short time, ensuring efficient photovoltaic inspections.

[0106] As one possible implementation, the step of acquiring photovoltaic power plant layout data, combining analysis results, environmental data, and time synchronization information, and generating inspection paths through a dynamic inspection priority queue and a dual-mode path planning algorithm includes:

[0107] A dynamic inspection priority queue is determined based on the defect level and environmental risk level; the priority of each photovoltaic module is generated based on the defect level, environmental risk level and corresponding weights, and the weight of the defect level is greater than the weight of the environmental risk level.

[0108] Configure a dual-mode path planning engine, where:

[0109] In normal mode, an improved ant colony algorithm is used to construct a global path by combining the distribution of defect hotspots with future weather forecast data. In this algorithm, each ant makes decisions based on heuristic functions and pheromone intensity during the path selection process, and periodically and automatically re-evaluates the global path to ensure coverage of all potentially high-risk areas.

[0110] When at least one of the defect level or environmental risk level is at the highest level, switch to emergency mode. In emergency mode, a cost function is constructed based on real-time environmental data and component location information. The comprehensive cost of the current path node is calculated based on the actual cost from the starting node to the current node and the heuristic cost from the current node to the target node. Node expansion is performed with the goal of minimizing the comprehensive cost, and path planning is completed.

[0111] In this embodiment, the path optimization module is one of the key components for achieving efficient inspection and risk avoidance. It integrates a dynamic inspection priority queue and a dual-mode path planning engine, relying on the output data from the deep learning processing module and the data fusion processing module to achieve dynamic scheduling. The dynamic inspection priority queue (data priority queue) first performs a weight evaluation on the photovoltaic modules. This evaluation is based on a comprehensive priority index constructed from defect level and environmental risk level. Emergency defect information has a priority level five higher than regular defects. Furthermore, if the environmental risk level is "dangerous," the inspection weight of the corresponding modules is further increased. (Comprehensive priority weighting) Expressed as:

[0112] ;

[0113] in, This represents the defect level code of the i-th component (5 for critical defects and 1 for regular defects). This indicates the environmental risk level code (hazard is 3, warning is 2, and safety is 1). and For the corresponding weighting coefficients, the system defaults to , This ensures that defects influence the dominant path ranking. During path planning, the normal mode employs an improved ant colony algorithm that combines defect hotspot distribution with 72-hour weather forecasts to construct a global path. In this algorithm, each ant makes path selection decisions based on a heuristic function and pheromone intensity, with the probability function defined as:

[0114] ;

[0115] in, Let be the transition probability of the k-th ant from node i to node j. Indicates the intensity of path pheromones. The heuristic function values ​​are constructed based on defect density and weather impact, with α and β controlling their weights. This mode automatically reassesses the global path every 30 seconds to ensure coverage of all potentially high-risk areas. In special circumstances, such as real-time wind speeds exceeding level 10 or the detection of an emergency defect, the system immediately switches to emergency mode via A. The algorithm generates an obstacle avoidance path. This path is based on a cost function constructed using real-time environmental data and component location information.

[0116] ;

[0117] in, The total cost of the current path node n. The actual cost from the starting node to the current node. For the heuristic cost estimation from the current node to the target node, path planning aims to minimize... The target node is expanded. The path optimization module interfaces with the data fusion and processing module to obtain environmental parameters such as wind speed and humidity with high-precision timestamps, ensuring that the environmental data used in path planning and the component status data are strictly consistent in the time dimension, ultimately achieving a simultaneous improvement in inspection efficiency and system responsiveness.

[0118] This embodiment assigns the highest weight to photovoltaic module data marked as "urgent defects," such as cracks or overheating, enabling real-time path planning. This means that when an urgent defect is detected in a photovoltaic module, the robot can quickly adjust its inspection path and reach the problematic module immediately for inspection and maintenance, effectively reducing power generation losses due to delayed fault handling. For example, in the actual operation of a photovoltaic power station, when an overheating emergency defect is detected in a photovoltaic module, the system uses a priority queue mechanism to plan a path to the module within one minute, allowing the robot to handle the issue promptly and preventing further damage to the module. For routine inspection data, a batch processing mode is used, balancing real-time performance and computational efficiency, ensuring both timely handling of urgent defects and efficient completion of routine inspection tasks.

[0119] In this embodiment, the normal mode and emergency mode dynamically adjust the path to deal with different situations, and rely on low-latency data with high-precision timestamps to ensure the spatiotemporal consistency of the path planning, prioritize the inspection of high-risk components, significantly enhance preventive maintenance capabilities, and reduce the robot's ineffective movement distance.

[0120] For example, in normal mode, an improved ant colony algorithm is used to generate a globally optimal path by combining historical data and predicted environmental parameters. This approach fully utilizes historical data and environmental prediction information, enabling the robot to perform inspections along the optimal path under normal conditions, improving inspection efficiency and coverage. For instance, based on historical defect distribution and weather forecast data for the next 72 hours, the robot can plan in advance to avoid potentially high-risk areas while ensuring comprehensive inspection of all components. When the deep learning module detects a sudden defect or real-time environmental parameters exceeding a safety threshold, the system immediately switches to emergency mode, using the Dijkstra algorithm to generate a local obstacle avoidance path to ensure robot safety. In the event of sudden severe weather, the robot can quickly adjust its path to avoid the impact of strong winds and other adverse environmental conditions, while continuing to complete the inspection task, ensuring the system's stability and adaptability.

[0121] As one possible implementation, the step of obtaining the defect level includes:

[0122] Obtain the defect type and confidence level obtained from defect identification, and determine the defect level based on the confidence level.

[0123] For example, the image recognition unit uses a lightweight CNN model (such as MobileNet-V3 based on knowledge distillation) to infer the image of a photovoltaic module, outputting the defect type (such as crack, dust accumulation, damage, overheating) and the corresponding confidence value C (range: 0≤C≤1). When the confidence value C≥0.95, the defect level is 5; when 0.8≤C<0.95, the defect level is 3; and in other cases, the defect level is 1. It should be noted that the threshold for classification can be other values, and more levels can be classified.

[0124] As one possible implementation, if the generation of inspection paths is blocked, alternative inspection paths are generated with reference to prior knowledge, including: pre-selection by using the minimum path cost function based on the defect distribution density function. In the path cost function, the higher the defect density, the greater the path cost, thereby guiding path planning to avoid or prioritize the coverage of high-risk areas.

[0125] For example, pre-selection can be performed using a minimum path cost criterion based on the defect distribution density function, where the path cost function is defined as:

[0126] ;

[0127] in, Let p be the total cost. Let be the Euclidean distance between the i-th node and the (i+1)-th node in the path. γ represents the historical defect density of the region corresponding to the node, and γ is the defect penalty coefficient. The higher the defect density, the greater the path cost, thereby guiding path planning to avoid or prioritize coverage of high-risk areas.

[0128] In this embodiment, the communication module, as a key connection unit of the intelligent operation and maintenance system, integrates an efficient caching mechanism and a low-latency transmission channel. While ensuring the overall real-time performance and robustness of the system, it achieves high-speed data forwarding and autonomous recovery capabilities during network outages. The module's built-in data cache has a capacity of ≥2GB, pre-loaded with photovoltaic power plant layout grid data, a historical defect coordinate library, and optimal path templates generated through ant colony optimization. These templates are used by the path optimization module to generate alternative inspection paths based on prior knowledge during network interruptions.

[0129] The hardware acceleration interface connects to a dedicated AI chip in the robot's body module. A dedicated bus channel enables low-latency communication in image recognition output and path command transmission, with end-to-end latency strictly controlled to within 1ms. The data throughput T is expressed as:

[0130] ;

[0131] Where S represents the amount of data transmitted in a single transmission, in bytes, and Δt represents the total transmission time, in seconds, ensuring that the real-time delivery of single-frame image results and path data commands meets the accuracy requirements of dynamic inspection. Wireless communication employs a dual-link collaborative mechanism of Wi-Fi 6 and 4G LTE. In-station communication uses a Wi-Fi module to achieve low-latency command synchronization of ≤10ms, while remote communication uses a 4G module to complete data return and status reporting with the cloud platform. AES-256 high-strength encryption algorithm is used during transmission, and its encryption function is expressed as:

[0132] ;

[0133] Where C represents the ciphertext data and P represents the original plaintext data. This indicates that AES encryption is performed using key k, with a key length of 256 bits. This encryption strategy ensures data privacy and integrity during public network transmission, avoids the risk of leakage of critical computation results during operation and maintenance, and builds a secure, low-latency, and fault-tolerant communication guarantee system.

[0134] As one possible implementation, the step of generating an operation and maintenance task queue based on the analysis results, while storing full-process data, includes:

[0135] An operation and maintenance task queue is generated based on the defect level and processing time. The priority of each operation and maintenance task is generated based on the defect level, processing time and corresponding weight, with the weight of the defect level being greater than the weight of the processing time.

[0136] For example, the task scheduling unit of the operations and maintenance management module generates an operations and maintenance task queue based on the time synchronization data from the data fusion processing module and the priority paths from the path optimization module. The priority of task scheduling consists of "defect level + processing time," where the defect level is derived from the image recognition results of the deep learning processing module, and the processing time is dynamically adjusted according to the urgency of the task. The response delay for urgent tasks must be controlled within 1 minute to ensure rapid response to system failures or equipment malfunctions. The task queue generation mechanism evaluates task priority using the following cost function:

[0137] ;

[0138] in, Let i be the priority of the i-th task. The defect level corresponding to the task. To determine the processing timeliness of tasks, λ1 and λ2 are weighting coefficients for defect level and processing timeliness. The system defaults to λ1=0.7 and λ2=0.3, ensuring that tasks with urgent defects are processed first. The task scheduling system continuously updates the task queue, adjusting priority weights to ensure that high-priority tasks are completed in the shortest possible time. The data storage unit uses a distributed time-series database (such as InfluxDB) to store multi-dimensional sensor data, setting different storage periods according to the data's lifecycle. Raw sensor data is stored as a time series with a storage period of 1 year, preprocessed feature data is stored for 3 years, while maintenance logs are stored permanently. The storage and retrieval of time-series data are achieved through efficient indexing mechanisms and query algorithms, supporting multi-dimensional data retrieval with second-level precision, ensuring data access efficiency and accuracy. The query cost function for data retrieval can be expressed as:

[0139] ;

[0140] in, The cost of querying The weight of the i-th query field. The system dynamically assesses the retrieval cost based on the data volume and the queried field. The human-computer interaction unit provides a real-time visual interface displaying the data flow status of each module, including key metrics such as data throughput, processing latency, and synchronization accuracy. Users can manually adjust the ROI extraction threshold and priority queue weight parameters of the edge computing unit through the interface, thereby optimizing system performance and response speed according to on-site operational needs. This interface supports flexible scheduling of tasks and parameters, adapting to the adjustment requirements of different working environments and operating states.

[0141] As one implementation method, this embodiment also incorporates hardware-software co-optimization, including heterogeneous computing acceleration: integrating a dedicated AI chip (such as a TPU / NPU) into the robot body module to achieve hardware acceleration of image recognition and path optimization algorithms, reducing data transmission latency between modules. Caching and prefetching strategies: setting up a data cache in the communication module to prefetch photovoltaic power station layout data and historical defect distribution patterns, enabling the path optimization module to generate candidate paths based on prior knowledge even without real-time data input, thus improving response speed.

[0142] This embodiment also provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it implements the steps of the above-described intelligent operation and maintenance method for photovoltaic inspection robots.

[0143] This embodiment also provides a storage medium storing computer-executable instructions. When the computer-executable instructions are loaded and executed by a processor, the steps of the above-described intelligent operation and maintenance method for photovoltaic inspection robots are implemented.

[0144] Through the above description of the embodiments, those skilled in the art will understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the specific device can be divided into different functional modules to complete all or part of the functions described above.

[0145] In the embodiments provided in this application, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another structure, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between structures or units, and may be electrical, mechanical, or other forms.

[0146] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units; that is, it can be located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0147] Furthermore, in the embodiments of this application, the functional units can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0148] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0149] 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.

Claims

1. A method for intelligent operation and maintenance of photovoltaic inspection robots, characterized in that, include: Image data, temperature data, and environmental data of photovoltaic modules in a photovoltaic power station are collected and differentially extracted, retaining only key local information. This includes: using a feature saliency scoring function to select regions with high weights for ROI extraction, focusing on potential defect areas in the image data by retaining only local images containing potential defects; simultaneously, a sliding window filtering algorithm is used to extract temperature and environmental data; the filtered data is further compressed using principal component analysis to retain the main information of the data. Standardized data is obtained by unifying the protocols and aligning the timestamps of various types of key information from the data acquisition module through a standardized data interface. Defect identification and operational status analysis are performed on standardized data to obtain analysis results. This includes: using a lightweight convolutional neural network model optimized by knowledge distillation to identify defects in photovoltaic module image data; employing a fusion structure of gated recurrent units and Transformer encoders to extract long-term dependency information; and assigning dynamic weights to different feature dimensions through a multi-head attention mechanism for operational status analysis. The defect identification and operational status analysis are based on a distributed computing architecture, with image recognition and status analysis tasks processed by independent computing nodes. Task allocation among nodes is dynamically adjusted in real-time using a load balancing algorithm, which improves processing efficiency by minimizing overall task time. The system acquires photovoltaic power plant layout data and combines analysis results, environmental data, and time synchronization information to generate inspection paths using a dynamic inspection priority queue and a dual-mode path planning algorithm. This includes: determining a dynamic inspection priority queue based on defect level and environmental risk level; where the priority of each photovoltaic module is generated based on its defect level and environmental risk level, along with their corresponding weights, with the defect level having a greater weight than the environmental risk level; and setting up a dual-mode path planning engine. In normal mode, an improved ant colony algorithm is used, combined with defect hotspot distribution and future weather forecast data, to construct a global path. In this algorithm, each ant makes decisions based on a heuristic function and pheromone intensity during path selection and periodically re-evaluates the global path to ensure coverage of all potentially high-risk areas. When either the defect level or environmental risk level is at least the highest level, the system switches to emergency mode. In emergency mode, a cost function is constructed based on real-time environmental data and module location information. The comprehensive cost of the current path node is calculated based on the actual cost from the starting node to the current node and the heuristic cost from the current node to the target node. The goal is to minimize the comprehensive cost by expanding nodes and completing path planning. An operation and maintenance task queue is generated based on the analysis results, and the entire process data is stored at the same time; The robot body performs movement and maintenance operations based on the inspection path and maintenance task queue.

2. The intelligent operation and maintenance method for photovoltaic inspection robots according to claim 1, characterized in that, The process involves unifying the protocols and aligning the timestamps of various key information types from the data acquisition module through a standardized data interface to obtain standardized data, including: Define a pre-formatted data structure for image, temperature, and environmental data. The structure includes data type labels, acquisition device ID, raw values, and quality factors, where the quality factors reflect the completeness and reliability of the current data. By connecting to a precision clock synchronization system, the sampling time is dynamically calibrated using the difference between the master clock broadcast signal and the local timing of the slave nodes, ensuring that the time error of all sampling nodes in the system is controlled within the preset error. Each frame of data is bound with a unique timestamp when it is generated. By minimizing the time alignment cost function between different data sequences, the distance matrix is ​​traversed through dynamic programming to backtrack to find the optimal alignment path. After alignment, the result is constrained with the target error window to ensure the consistency of multi-source data in the time dimension.

3. The intelligent operation and maintenance method for photovoltaic inspection robots according to claim 1, characterized in that, The steps for obtaining the defect level include: Obtain the defect type and confidence level obtained from defect identification, and determine the defect level based on the confidence level.

4. The intelligent operation and maintenance method for photovoltaic inspection robots according to claim 1, characterized in that, If the generation of inspection paths is blocked, alternative inspection paths are generated by referring to prior knowledge, including: pre-selection by using the minimum path cost function based on the defect distribution density function. In the path cost function, the higher the defect density, the greater the path cost, thereby guiding the path planning to avoid or prioritize the coverage of high-risk areas.

5. The intelligent operation and maintenance method for photovoltaic inspection robots according to claim 1, characterized in that, The process involves generating an operation and maintenance task queue based on the analysis results, and simultaneously storing full-process data, including: An operation and maintenance task queue is generated based on the defect level and processing time. The priority of each operation and maintenance task is generated based on the defect level, processing time and corresponding weight, with the weight of the defect level being greater than the weight of the processing time.

6. A photovoltaic inspection robot intelligent operation and maintenance system, used to execute the photovoltaic inspection robot intelligent operation and maintenance method as described in any one of claims 1 to 5, characterized in that, include: The data acquisition module is used to collect image data, temperature data, and environmental data of photovoltaic modules in the photovoltaic power station. The data fusion processing module is used to unify protocols and align timestamps for various types of key information from the data acquisition module through a standardized data interface to obtain standardized data. This includes: using a feature saliency scoring function, selecting regions with high weights for ROI extraction, and focusing on potential defect areas in image data by retaining only local images containing potential defects; simultaneously, using a sliding window filtering algorithm to extract temperature and environmental data; and further using principal component analysis to compress the dimensions of the filtered data to retain the main information. The deep learning processing module is used to perform defect identification and operational status analysis on standardized data to obtain analysis results. This includes: using a lightweight convolutional neural network model optimized based on knowledge distillation to identify defects in photovoltaic module image data; employing a fusion structure of gated recurrent units and Transformer encoders to extract long-term dependency information; and assigning dynamic weights to different feature dimensions through a multi-head attention mechanism for operational status analysis. The defect identification and operational status analysis are based on a distributed computing architecture, with image recognition and status analysis tasks processed by independent computing nodes. The nodes dynamically adjust task allocation in real time through a load balancing algorithm, which improves processing efficiency by minimizing overall task time. The path optimization module acquires photovoltaic power plant layout data and, combined with analysis results, environmental data, and time synchronization information, generates inspection paths through a dynamic inspection priority queue and a dual-mode path planning algorithm. This includes: determining a dynamic inspection priority queue based on defect level and environmental risk level; where the priority of each photovoltaic module is generated based on its defect level, environmental risk level, and corresponding weights, with the defect level having a greater weight than the environmental risk level; and setting up a dual-mode path planning engine. In normal mode, an improved ant colony algorithm is used to construct a global path by combining defect hotspot distribution and future weather forecast data. In this algorithm, each ant makes decisions based on a heuristic function and pheromone intensity during path selection and periodically re-evaluates the global path to ensure coverage of all potentially high-risk areas. When either the defect level or environmental risk level is at least the highest level, the system switches to emergency mode. In emergency mode, a cost function is constructed based on real-time environmental data and module location information. The comprehensive cost of the current path node is calculated based on the actual cost from the starting node to the current node and the heuristic cost from the current node to the target node. The goal is to minimize the comprehensive cost by expanding nodes and completing path planning. The operations and maintenance management module is used to generate an operations and maintenance task queue based on the analysis results, and to store the entire process data. The robot itself is used to perform movement and maintenance operations based on the inspection path and maintenance task queue.

7. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor, when calling the computer program in the memory, implements the steps of the intelligent operation and maintenance method for photovoltaic inspection robots as described in any one of claims 1 to 5.

Images