Power equipment monitoring system and method

Abstract

The invention provides a power equipment monitoring system and method, and relates to the field of power equipment monitoring. The system comprises a large language model module, an edge device processing module, a device portrait module and a result merging module, the large language model module generates a blocking instruction according to the image size recorded in a device portrait, the size of a processable sub-block and historical fault information, and can output customized image processing mode instructions for different risk areas; the edge equipment processing module only needs to perform block-by-block cutting and multi-mode detection on the partitioned image, so that the single-time processing burden is reduced; and the result merging module is used for fusing the overlapped regions based on the sub-block analysis result and the address mapping information, and finally obtaining fault identification or feature distribution of the complete high-resolution power equipment image. By means of the system, efficient block analysis and personalized depth detection of the images of the power equipment can be achieved, and the defect detection rate is increased while resource investment is reduced.

Description

Technical Field

[0001] The present disclosure relates to the field of power equipment monitoring, and more particularly, to a power equipment monitoring system and method. Background Art

[0002] In a power system, power equipment (such as transformers, circuit breakers, insulators, switchgear, etc.) operates under high voltage, high temperature or complex environments for a long time, and is prone to failures due to various reasons such as insulation aging, partial discharge, dirt accumulation, and mechanical fatigue. If these failures are not detected and handled in time, large-scale power outages or serious safety accidents may occur. Therefore, stable and accurate condition monitoring and fault prediction of power equipment have always been the core requirements in the field of power operation and maintenance.

[0003] With the development of digital and intelligent technologies, fault detection based on image analysis has gradually become an important means for power equipment monitoring. For example, visible light images or infrared thermal images are collected at the operation site, and by identifying surface cracks, discharge traces, and temperature abnormal areas on the equipment, it helps operation and maintenance personnel to detect potential hidden dangers early. Compared with the traditional manual inspection method, image analysis has higher detection efficiency and can achieve remote or batch monitoring. However, the existing technologies still face the following deficiencies in practical applications. For example, power equipment images generally have high resolutions. Especially in ultra-high voltage or extra-high voltage scenarios, in order to clearly observe local details, a single image may reach the level of thousands to tens of thousands of pixels. If processed as a whole at one time, it often places too high requirements on hardware computing power and memory, and problems such as too long processing time or insufficient memory are likely to occur. In addition, the fault modes, operating environments, and historical conditions of different power equipment vary significantly. If a single general image analysis process is adopted, it is often impossible to take into account diverse scenarios, and the missed detection or misjudgment rate is difficult to control.

[0004] Although some studies have introduced advanced algorithms such as deep learning, they lack refined configuration for requirements such as the level of fault risk, edge region processing, and image multi-modal fusion, and still cannot automatically adapt to different equipment or fault types within a unified system. Most of the existing solutions rely on manual settings or fixed scripts for the configuration of the image analysis process. Once the equipment working conditions change or the acquisition resolution increases, the algorithms and parameters need to be manually updated, and it is difficult to perform dynamic adjustment in a timely and efficient manner. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present application provides a power equipment monitoring system and method.

[0006] In a first aspect, the present application provides a power equipment monitoring system, including:

[0007] The large language model module analyzes the chunking strategy of the power equipment image with the target resolution based on the image size of the power equipment image obtained from the device portrait and the size of the sub-chunks that can be processed by the edge device, and obtains the chunking instruction; wherein, the large language model module includes: a large model trained and / or fine-tuned based on the power equipment scenario data;

[0008] The edge device processing module performs chunking processing on the power equipment image based on the chunking instruction, and only analyzes the image chunks in the area defined by the size of the sub-chunks that can be processed each time, and obtains the sub-chunk analysis result;

[0009] The device portrait module is used to store the image meta-information, historical fault records, and the size information of the sub-chunks that can be processed of the power equipment, and provide the image size and the size of the sub-chunks that can be processed to the large language model module;

[0010] The result merging module performs result stitching and fusion based on the sub-chunk analysis result and the address mapping corresponding to the sub-chunk in the chunking instruction, and obtains the fault identification and / or feature extraction result for the complete power equipment image with the target resolution;

[0011] Among them, the large language model module determines the chunking processing method of the edge device processing module for the power equipment image according to the image size and the size of the sub-chunks that can be processed provided by the device portrait module, and the result merging module maps the sub-chunk analysis result after chunking back to the original image coordinates, and finally records the overall detection result in the device portrait module.

[0012] As an optional implementation manner, the large language model module calculates the number of chunks according to the image size and the size of the sub-chunks that can be processed, and fills zeros or splices the edge parts that cannot be divided evenly to generate the corresponding chunking instruction.

[0013] As an optional implementation manner, the device portrait module is used to store at least one of the size of the sub-chunks that can be processed by the edge device, the distribution of historical fault hot spots, the image resolution, and the multi-modal sensing data;

[0014] The large language model module determines whether to use overlapping chunking for the edge area based on the size of the sub-chunks that can be processed and the distribution of historical fault hot spots.

[0015] As an optional implementation manner, the large language model module generates the chunking instruction including:

[0016] In response to the image size being evenly divisible by the size of the sub-chunks that can be processed, directly divide the entire power equipment image into multiple image chunks of the same size;

[0017] In response to the image size not being divisible by the integer of the processable sub-block size, zero-padding is performed on the non-divisible edge part or it is spliced with adjacent blocks, and the effective area information of the edge sub-blocks is recorded in the chunking instruction.

[0018] As an alternative implementation, the edge device processing module includes:

[0019] An image processing unit for cropping or splicing image blocks of the processable sub-block size according to the chunking instruction;

[0020] An image analysis unit for performing defect detection, texture analysis, and segmentation operations on each image block;

[0021] A communication unit for uploading the output data of the image analysis unit to the result merging module.

[0022] As an alternative implementation, the result merging module performs a maximum value or mean value process on the output results of adjacent sub-blocks with overlapping regions according to the fusion method specified in the chunking instruction. When different detection results appear at the same target pixel coordinate for multiple sub-blocks, a preset weighted merging strategy is used to obtain the final result.

[0023] As an alternative implementation, the large language model module is deployed on a cloud server or a computing platform, and the edge device processing module is deployed in the edge hardware at the power site. The two perform instruction issuance and sub-block analysis result feedback through network communication.

[0024] As an alternative implementation, the large language model module automatically adjusts the image chunking strategy based on the industry specifications and historical fault cases provided by the device profiling module. When the processable sub-block size or the image size changes, new chunking instructions are dynamically generated.

[0025] As an alternative implementation, the large language model module is also used to generate an image processing method instruction for each sub-block and send the image processing method instruction to the edge device processing module after determining the method of chunking the power device image based on the personalized information related to the historical fault records, operating environment, and sensing data of the power device obtained from the device profiling module. The image processing method instruction includes:

[0026] The target algorithm type, which is used to indicate the image processing, defect detection, and texture analysis algorithms to be used for each sub-block;

[0027] The parameter configuration, which is used to mark whether additional filtering, enhancement, or segmentation thresholds are performed under the target spectrum or temperature distribution;

[0028] Priority information is used to instruct the edge device processing module to execute a preset second detection process for sub - blocks with a failure risk greater than or equal to a preset risk value, and a preset first detection process is adopted for sub - blocks with a failure risk less than the preset risk value;

[0029] A multi - modal fusion strategy is used to, in response to the mapping relationship between visible light images and infrared images recorded in the device portrait module, and according to the image processing method instruction, instruct the edge device processing module to perform cross - spectral or multi - channel data fusion on the corresponding sub - blocks.

[0030] In a second aspect, the present application provides a power equipment monitoring method, including:

[0031] Using a large language model, based on the image size of the power equipment image obtained from the device portrait and the size of the sub - blocks that can be processed by the edge device, perform a block - division strategy analysis on the power equipment image with the target resolution to obtain a block - division instruction; wherein, the large language model includes: a large model trained and / or fine - tuned based on power equipment scenario data;

[0032] Using the edge device, based on the block - division instruction, perform block - division processing on the power equipment image, and only analyze the image blocks in the area defined by the size of the sub - blocks that can be processed each time to obtain sub - block analysis results;

[0033] Using the device portrait, store the image meta - information, historical failure records, and the size information of the sub - blocks that can be processed of the power equipment, and provide the image size and the size of the sub - blocks that can be processed to the large language model;

[0034] Based on the sub - block analysis results and the address mapping of the sub - blocks in the block - division instruction, perform result stitching and fusion to obtain a failure identification and / or feature extraction result for the complete power equipment image with the target resolution;

[0035] Wherein, the large language model determines the way of block - division processing of the power equipment image by the edge device according to the image size and the size of the sub - blocks that can be processed provided by the device portrait, maps the sub - block analysis results after block - division back to the original image coordinates, and records the overall detection results.

[0036] Compared with the prior art, by introducing the concepts of a large language model and a device portrait, the present application proposes an adaptive and scalable image analysis process for the generation of block - division strategies and failure risk assessment of high - resolution power equipment images. By combining image block - division with personalized detection instructions, it not only reduces the dependence on hardware resources but also enhances the processing ability for special failure modes and multi - modal data. Based on the existing image processing and power equipment operation and maintenance technologies, the present invention realizes more flexible and accurate failure monitoring and automated decision - making, which has important value for ensuring the safe and stable operation of the power system. Brief Description of the Drawings

[0037] Figure 1 It is a schematic diagram of the power equipment monitoring system provided by the embodiment of the present application;

[0038] Figure 2 It is a flowchart of the power equipment monitoring method provided by the embodiment of the present application;

[0039] Figure 3 It is a schematic diagram of the maximum sub-block size in the overall image analysis provided by the embodiment of the present application;

[0040] Figure 4 It is a schematic diagram of the position area of the historical fault in the overall image provided by the embodiment of the present application. Detailed Description of the Embodiment

[0041] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0042] See Figure 1 As shown, the power equipment monitoring system provided by the embodiment of the present application includes:

[0043] The large language model module 10 analyzes the chunking strategy of the power equipment image with the target resolution based on the image size of the power equipment image obtained from the device portrait and the size of the sub-chunks that can be processed by the edge device, and obtains a chunking instruction; wherein, the large language model module includes: a large model trained and / or fine-tuned based on the power equipment scenario data;

[0044] The edge device processing module 20 performs chunking processing on the power equipment image based on the chunking instruction, and only analyzes the image chunks in the area defined by the size of the sub-chunks that can be processed each time, and obtains the sub-chunk analysis result;

[0045] The device portrait module 30 is used to store the image meta-information, historical fault records and the size information of the sub-chunks that can be processed of the power equipment, and provide the image size and the size of the sub-chunks that can be processed to the large language model module;

[0046] The result merging module 40 performs result splicing and fusion based on the sub-chunk analysis result and the address mapping corresponding to the sub-chunk in the chunking instruction, and obtains the fault identification and / or feature extraction result for the complete power equipment image with the target resolution;

[0047] Among them, the large language model module 10 determines the way for the edge device processing module 20 to perform block processing on the power device image according to the image size and the size of the processable sub-blocks provided by the device portrait module 30, and the result merging module 40 maps the analysis results of the sub-blocks after block processing back to the original image coordinates, and finally records the overall detection result in the device portrait module 30.

[0048] As an alternative implementation, the large language model module 10 calculates the number of blocks according to the image size and the size of the processable sub-blocks, and pads zeros or splices the non-divisible edge parts to generate corresponding block instructions.

[0049] As an alternative implementation, the device portrait module 30 is used to store at least one of the size of the processable sub-blocks of the edge device, the distribution of historical fault hot areas, the image resolution, and multi-modal sensing data;

[0050] The large language model module 10 determines whether to use overlapping block division for the edge area based on the size of the processable sub-blocks and the distribution of historical fault hot areas.

[0051] As an alternative implementation, the large language model module 10 generating the block instructions includes:

[0052] In response to the image size being divisible by the size of the processable sub-blocks as an integer, directly dividing the entire power device image into several image blocks of the same size;

[0053] In response to the image size not being divisible by the size of the processable sub-blocks as an integer, padding zeros or splicing with adjacent blocks for the non-divisible edge parts, and recording the effective area information of the edge sub-blocks in the block instructions.

[0054] Regarding the above large language model module 10:

[0055] In specific implementation, the large language model module 10 can obtain the image size (such as the target resolution is 4096×4096) of the current power device image and the size of the processable sub-blocks of the edge device (such as 512×512) from the device portrait module 30 through means such as API / database query. If the device portrait also provides data such as historical fault hot spots or industry specifications, they can also be read together. This module can include a large model based on the Transformer structure, which has been pre-trained and / or fine-tuned on general language materials and power industry fault cases in the early stage to make it familiar with common fault modes of power devices, image processing requirements, and scenarios with limited edge computing power.

[0056] Exemplarily, during inference, prompts can be input to the large model, including information such as "image resolution = 4096×4096", "processable sub-block size = 512×512", "historical failures concentrated in the lower right corner", etc. The model will comprehensively process these contexts and output a chunking strategy.

[0057] It can be understood that the above prompts can be automatically extracted and executed by a script or other computer programs.

[0058] Exemplarily, please refer to Figure 4 , Figure 4 which is a schematic diagram of the position area of historical failures in the overall image provided by an embodiment of the present application.

[0059] Among them, the historical failure area is presented as solid circles in the sub-blocks concentrated in the lower right corner. The maximum sub-block size is 512×512, and the overall image size is 1536×1536.

[0060] In a specific implementation, the large language model module 10 is set to determine whether to pad with zeros or splice according to whether the image can be divisible by the processable sub-block; when it is found that the historical failure area is located in the lower right corner, the model can preferentially suggest more refined chunking of this area or setting a higher detection priority; through prompts or internal rules, the model generates information about the coordinates, sizes, and edge processing methods of several sub-blocks.

[0061] The model outputs the above analysis results in a structured format (such as JSON), for example, including: if there are cases where the row and column of edge pixels cannot be divided evenly, it will be marked as "edge_handling": "padding" or "merge_with_block": "X".

[0062] Exemplarily, in an embodiment of the present application, when the large language model module 10 generates a chunking instruction, if it recognizes that the image size cannot be divisible by the processable sub-block size, it will attach an "edge_handling" field to each non-divisible sub-block in the chunking instruction sent to the edge device processing module 20 to indicate the specific processing method. Among them, if it is written as "edge_handling": "padding", it means that when there is a shortage of edge pixels in the sub-block compared to the standard size (such as 512×512), the remaining area of the sub-block needs to be automatically filled with zeros to unify the sub-block size. For example, if the image width is exactly 4 columns short of reaching 512 columns, these 4 columns will be filled with 0 pixels (or invalid pixels) to ensure that the downstream image processing algorithm can analyze according to a fixed size.

[0063] If it is stated that “edge_handling”: “merge_with_block”: “X”, it means that the large language model module 10 determines that this edge part is in a close or adjacent state to another already partitioned sub-block (with the BlockID identified as X). Therefore, it instructs the edge device processing module 20 to directly merge this remaining edge area with the sub-block numbered X into a larger sub-block when cropping and loading the image, so as to avoid generating overly small or zero-padded data blocks. At this time, “X” is a block_id value defined in a partitioning instruction, used to uniquely identify the target sub-block to be merged. For example, when there is only a narrow strip 4 pixels wide in the lower right corner of the image, which forms a close relationship with the adjacent sub-block 7, the instruction will specify “edge_handling”: “merge_with_block”: “7”, indicating that this 4-pixel-wide image strip should be cropped, loaded, and processed in the same sub-block analysis process together with sub-block 7.

[0064] Through the flexible setting of the above “edge_handling” field, the system can clearly indicate the two types of edge area processing methods, “zero-padding” or “stitching with adjacent sub-blocks”, in the case of non-divisibility. After parsing the corresponding field, the edge device processing module 20 will automatically perform the corresponding operations: when it is “padding”, zero-padding is performed on the missing pixel segment; when it is “merge_with_block”: “X”, adjacent pixels are first extracted and combined from the image areas of this sub-block and sub-block X, and then loaded into the memory for image analysis. This can ensure that the edge area can also be included in the normal detection process, reducing the interference with computing power and algorithm adaptability, and enabling all potential fault points in high-resolution images to be covered and analyzed.

[0065] The finally generated partitioning instruction is sent to the edge device processing module 20, and the latter partitions and loads the image for analysis according to this instruction.

[0066] In another embodiment, the model is fine-tuned using a specialized dataset containing thousands of fault images of power equipment such as cracks, partial discharges, and overheating, as well as their text descriptions.

[0067] In a specific implementation, by repeatedly performing reinforcement learning or supervised learning on issues such as “whether to partition” and “how to set high-priority partitions”, the model gradually masters the decision-making thinking of balancing detection accuracy under limited computing power.

[0068] When performing partitioning strategy analysis, the system constructs a Prompt example, such as: Based on this, the model generates a partitioning instruction and can append details such as “The seventh sub-block needs padding” and “Set high detection priority for the lower right corner sub-block”.

[0069] In this way, the present application can implement an efficient block strategy analysis according to the image size and sub-block size and output block instructions.

[0070] Regarding the above-mentioned edge device processing module 20:

[0071] To address the problem of insufficient computing power or memory on edge devices for high-resolution power device images, based on the block instructions issued by the large language model module 10, the present invention designs an edge device processing module 20 for performing per-block analysis on the sub-block size that can be processed only once at a time.

[0072] In a feasible embodiment of the present application, the edge device processing module 20 includes an image processing unit, an image analysis unit, and a communication unit to complete the entire process of cropping and loading sub-block data, image processing, and result output.

[0073] In a specific implementation, the edge device processing module 20 receives the block instructions generated by the large language model module 10 through a network interface or local shared memory. The block instructions can adopt JSON or other structured formats, and the key fields may include:

[0074] block_i: used to uniquely identify a certain sub-block;

[0075] x_start, y_start: indicating the starting row and column positions of the sub-block in the original image coordinate system;

[0076] width, height: the block size, such as 512×512;

[0077] edge_handling: whether to pad with zeros or splice with adjacent sub-blocks when the image is not divisible;

[0078] detection_level or priority_merge: indicating the detection level, subsequent fusion priority, etc. that should be adopted for this sub-block.

[0079] After parsing the block instructions, the edge device processing module 20 generates a set of sub-block queues to be processed internally. Each item in the queue contains the above fields and possible fault risk prompts (if the large language model determines that the fault probability in a certain area is higher based on the device profile).

[0080] The image processing unit in the edge device processing module 20 first reads the entire power equipment image (or loads it in segments from a high-resolution image file or stream in actual implementation), and crops out the corresponding sub-block image data according to the coordinate data such as (x_start, y_start, width, height) specified by the instruction. If the edge_handling field indicates that zero-padding is required, the system will fill zero pixels at the lower right edge or the specified area to make up the size suitable for algorithm adaptation; if it indicates "stitching with block ID=XX", two sub-blocks need to be loaded first and merged at the boundary before being sent to the image analysis unit.

[0081] Through this block-by-block loading, the edge device only needs to retain image data of a processable size (such as 512×512) in memory once, avoiding excessive memory occupation caused by loading the entire image.

[0082] The image analysis unit can perform various processes on the loaded sub-block images, mainly including:

[0083] Defect detection: Using detection algorithms based on thresholding or deep learning (such as CNN / Transformer) to find surface cracks, partial discharge traces, or foreign object fouling;

[0084] Texture analysis: If a crack risk is detected or the surface material of the device is special, texture features can be extracted through Gabor filtering or GLCM methods;

[0085] Segmentation operation: Using models such as GrabCut and U-Net to segment infrared thermal images or visible light images to locate hot spots or possible fault areas;

[0086] Fusion processing: If the device portrait or the block instruction prompts that the infrared band and the visible light of the same sub-block need to be aligned, image registration and multi-channel fusion can be performed at this stage.

[0087] If the detection_level indicates in the block instruction that this block is in a high-fault-risk area, the image analysis unit switches to a higher-precision detection model (such as a deeper CNN) and outputs a more detailed segmentation Mask or probability distribution map.

[0088] In an alternative embodiment of the present invention, in addition to outputting block instructions, the large language model module can also output image processing method instructions, which will be elaborated in detail below.

[0089] After completing the analysis of the sub-block images, the edge device processing module 20 will generate "sub-block analysis results". For example, it includes:

[0090] block_id: Corresponding to the block instruction;

[0091] defect_coordinates or defect_mask: If cracks, hot spots, etc. are detected, record their specific positions in the sub-block coordinate system or the segmentation contours.

[0092] confidence_scores: Confidence levels for the outputs of the detection algorithms.

[0093] valid_area: If zero-padding or stitching is used, indicate the actual valid pixel range.

[0094] Other algorithm outputs (such as texture feature vectors, temperature distribution data, etc.).

[0095] If cloud / local collaboration is set up, the communication unit packages and sends this "sub-block analysis result" back to the result merging module 40 or the cloud server for subsequent global image stitching and comprehensive fault judgment.

[0096] In addition, when the sub-block size does not meet the algorithm requirements (for example, it must be a multiple of 2), the image processing unit can automatically adjust or throw an exception for handling during the instruction parsing stage; in a parallel environment, multiple sub-blocks can be loaded at once, and multiple threads or multiple GPUs can be called for acceleration; if the device portrait shows particular attention to hot spot anomalies in a specific season or working condition, the edge device processing module 20 can also preferentially analyze relevant sub-blocks based on the hints in the instructions and push real-time alarms.

[0097] In this way, through the above-mentioned block-based image analysis method, the edge device processing module 20 only needs to meet the memory and computing power requirements within the block size range, without having to process the entire high-resolution image at once, thus effectively reducing hardware investment and ensuring detection accuracy.

[0098] Regarding the above-mentioned device portrait module 30:

[0099] To enable the large language model module 10 to accurately obtain the image size of the power equipment, the processable sub-block size, and more historical fault information, the present invention centrally stores relevant information in the device portrait module 30.

[0100] In a feasible implementation of this application, the device portrait module 30 can be regarded as a set of "digital device archives" or "digital twin" databases, recording historical data, real-time monitoring information, and operation and maintenance conditions since the device was installed.

[0101] The content stored in the device portrait module 30 may include:

[0102] Image meta-information: Record the resolution, shooting time, shooting device ID, shooting angle, etc. of each captured image.

[0103] Historical fault records: corresponding to the occurrence time, type (such as crack, overheat, contamination) of each fault event, fault location (which may be in the form of coordinates or text description), repair results, etc.;

[0104] Processable sub-block size information: reflecting the maximum sub-block size that the edge device can load for each image analysis (for example, 512×512); it can also be extended to record specific available video memory, bandwidth and other limitations;

[0105] Exemplarily, please refer to Figure 3 , which is a schematic diagram of the maximum sub-block size in the overall image analysis provided by the embodiment of the present application. Among them, the maximum sub-block size is 512×512, and the overall image size is 1536×1536.

[0106] Multi-modal sensing data: If there is an infrared thermal image or vibration signal, its key features or file links can be stored;

[0107] Basic device attributes: such as device model, service life, manufacturer, installation location, and operating conditions such as ambient temperature and humidity in some scenarios.

[0108] In terms of technical implementation, a relational database (MySQL / PostgreSQL) or a NoSQL document library (MongoDB, ElasticSearch) can be used to manage these fields, or the key information can be saved in the form of a JSON file on the cloud server.

[0109] In addition, the device portrait module 30 is not a static configuration library, but can be continuously updated according to the results of subsequent fault detection or operation and maintenance inspections. For example, when the edge device processing module 20 detects a new crack or hot spot, it will package the detection time, coordinate area and severity and write them back to the device portrait. This process can be uniformly collected by the result merging module 40 and then written into the database through the API, so that the "device portrait" can keep the latest description of the current power device status. At the same time, the large language model module 10 can query this module again before the next inference to obtain the latest historical data and sub-block processable limits, thus forming a closed loop.

[0110] When the large language model module 10 needs to perform block strategy analysis or image processing method decision-making, it first sends a query request to the device portrait module 30 according to the device ID. The latter queries the record item corresponding to the device in the table (or document) and returns the required fields.

[0111] The large language model module 10 can then parse the specified fields therein, make a block strategy for the current image, and can also judge whether there are specific areas that need high-precision detection according to special fields such as historical maintenance records.

[0112] In actual deployment, the device portrait module 30 can be combined with a cloud server to provide data query and update operations externally through RESTful or RPC interfaces. If the system scale is large, a cache layer can also be added to accelerate frequent queries.

[0113] For power equipment, usually a substation or line contains multiple devices, and each device retains independent data records in the device portrait module, distinguished by device_id or other unique identifiers. This can meet the management requirements for scenarios with multiple devices, multiple images, and high concurrency.

[0114] Through the unified storage and dynamic maintenance method of this module, the large language model module 10 no longer needs to obtain device information manually or temporarily, greatly reducing the configuration burden. At the same time, as the number of detections increases, the device portrait continuously enriches the fault modes and maintenance information, thereby providing data support for the personalized generation of block strategies or processing strategies, enabling the system to continuously improve the accuracy and efficiency of fault monitoring during long-term operation.

[0115] Regarding the result merging module 40:

[0116] To integrate the analysis results of each sub-block into a fault detection or feature extraction output consistent with the original high-resolution image coordinate system, the present invention sets up a result merging module 40. This module receives multiple "sub-block analysis results" from the edge device processing module 20, and completes the splicing and fusion of the complete image with reference to the address mapping information in the block instruction generated by the large language model module 10, and finally writes the overall detection result into the device portrait module 30.

[0117] After the edge device processing module 20 completes the detection of each sub-block, it will produce a data structure (such as JSON), including:

[0118] block_id: Identifies the sub-block;

[0119] local_defect_mask or local_features: Records the detection results or features in the local coordinate system (0width - 1, 0height - 1) of this sub-block;

[0120] confidence_scores: Confidence levels of each pixel or detection target;

[0121] valid_area: If the sub-block is padded with zeros or spliced, it indicates the actual valid pixel area.

[0122] Meanwhile, in the chunk instructions generated by the large language model module 10, the global starting coordinates (x_start, y_start) corresponding to the block_id, as well as information such as width and height, are recorded. Based on this, the result merging module 40 can map the local coordinates back to the original image coordinates.

[0123] When adjacent sub-chunks are exactly continuous and non-overlapping in the horizontal or vertical direction, the result merging module 40 can directly place the local_defect_mask or local_features at the corresponding global coordinate positions. For example, place sub-chunk ID=1 at (0511,0511), sub-chunk ID=2 at (5121023,0511), etc., thereby stitching together the entire high-resolution detection result map.

[0124] Among them, if the coordinate range of sub-chunk ID=1 is x = 0 to 511, y = 0 to 511 (denoted as (0511, 0511)), then the range of sub-chunk ID=2 can be immediately after it, such as x = 512 to 1023, y = 0 to 511 (denoted as (5121023, 0511)), thereby stitching together a local part of the image with a full width of 1024 pixels in the horizontal direction.

[0125] It can be understood that the above markings are only examples, and actual character segmentation can be performed, which will not be elaborated here.

[0126] If a certain sub-chunk has a "zero-padding" or "edge_handling = padding" field, only the valid_area part is stitched; the zero-padding area can be automatically ignored or overwritten.

[0127] If the model instruction specifies that there is an N-pixel-wide overlapping area between sub-chunks (or the edge parts of adjacent sub-chunks are overlapped to improve the detection accuracy), the result merging module 40 needs to fuse the overlapping area. The following strategies can be adopted:

[0128] Maximum value fusion: Take the maximum value of the detection values obtained for the same pixel in the two sub-chunks;

[0129] Mean value fusion: Take the average of the same pixel;

[0130] Confidence weighting: If the detection confidence of one sub-chunk is higher, give priority to using its result;

[0131] Priority strategy: If the priority_merge field of a certain sub-chunk is high, overwrite the result of the other sub-chunk.

[0132] When the result merging module 40 detects overlapping pixels, it will consult the confidence level or priority field of the sub-block analysis results, and then generate the final pixel value (or the final defect mark) accordingly. This process can be flexibly configured according to the operation and maintenance needs.

[0133] After coordinate mapping and fusion, the result merging module 40 can finally obtain a defect detection / splitting map or other feature map with the same resolution as the original image (such as 4096×4096). If the output of the detection algorithm is a segmentation Mask, multiple sub-block Masks can be spliced into the same two-dimensional matrix; if the output is a list of coordinates, it can be converted into a global coordinate list and marked in the device portrait.

[0134] After the merging is completed, the result merging module 40 can write the whole-image detection results (or the fault coordinates and fault types extracted therefrom) back to the device portrait module 30. When the image analysis or manual maintenance is performed again, this global result can be used as a historical reference to help the large language model module 10 perform the next round of block division strategy fine-tuning or risk assessment.

[0135] In this way, through the result merging module 40, the system can accurately map the detection information scattered in each sub-block back to the original image coordinates, realizing the overall fault visualization or feature representation of the high-resolution image. Compared with the global stitching difficulties easily caused by single-block algorithms or manual repeated annotation, the fusion mechanism of the present invention takes into account the sub-block features, overlapping processing and priority configuration, ensuring the accuracy and consistency of the whole-image detection results.

[0136] In another alternative embodiment of the present application, when the large language model module 10 performs the block division strategy analysis, it focuses on deciding whether to use "padding with zeros" or "splicing with adjacent blocks" for the non-divisible edge part.

[0137] For example, when the image size provided by the device portrait module 30 , and the size of the sub-block that can be processed , the model can calculate 8 complete sub-blocks, with 4 remaining pixels in the edge rows and columns. At this time, the large language model module 10 will generate in the block division instruction:

[0138] If padding with zeros, it will indicate "edge_handling: \"padding\"" in the instruction, and indicate that there are actually only 4 rows of available pixels in the valid_area field;

[0139] If splicing with an adjacent block, it will indicate "merge_with_block = BlockID(n)" in the instruction, instructing the edge device processing module 20 to first merge the remainder area with the sub-block corresponding to BlockID(n) and then perform the analysis.

[0140] In this way, the chunking instruction precisely describes the processing solution for the edge region, facilitating the zero-padding or merging operations when the edge device preprocessing unit 20 loads image chunks without manual intervention.

[0141] When the device portrait module 30 records the size of the sub-chunks that the edge device can process, the image resolution, and the "historical fault hot zone distribution", the present invention can determine whether to use overlapping chunking for the edge region based on the existence of high-fault areas. Suppose the historical record shows that there are frequent cracks or overheating at the lower edge of the image (row coordinate > 3500). The large language model module 10 can specify, in the chunking instruction, something like "overlap_size: 20" pixels, so that adjacent sub-chunks have a 20-pixel vertical overlap area to aggregate more feature information in this high-risk area.

[0142] Exemplarily, the model can write in the instruction:

[0143] {"block_id":8, "x_start":0, "y_start":3500,"width":512, "height":512,"overlap_with_block":7,"overlap_size":20}

[0144] Among them, "overlap_with_block": 7 means overlapping with the sub-chunk numbered 7 in the vertical direction, and this value corresponds to another sub-chunk block_id defined in the chunking instruction; "overlap_size": 20 is uniformly used to indicate the number of pixels in the overlapping area, meaning that there are 20 rows of pixel areas that are repeatedly scanned or detected between the two chunks. This way enables the result merging module 40 to more precisely merge the detection results during subsequent fusion, reducing the omission of crack breakpoints or hot spots caused by boundary segmentation.

[0145] In another alternative embodiment, when the large language model module 10 generates the chunking instruction, if the image size is exactly divisible by the size of the sub-chunks that can be processed , it can be directly divided into equal-sized sub-chunks, and each sub-chunk has no additional edge_handling flag. On the contrary, if occurs, the model will add a "valid_area" field to the edge sub-chunks in the output instruction to describe the available pixel range.

[0146] Among them, K represents the side length dimension of the entire image (for example, the number of pixels in width or height); P represents the side length dimension of the sub-chunks that can be processed; K mod P = 0 means that the remainder is 0 after the "modulus" operation; K mod P≠0 indicates that there is a remainder when the image side length K is divided by P.

[0147] For example:

[0148] {"block_id":10,"x_start":3584, "y_start":3584,"width":512, "height":512,"valid_area": {"width":420, "height":420},"edge_handling": "padding"}

[0149] In this way, subsequent edge devices only analyze the 420×420 effective pixel area; the remaining positions can be regarded as padding areas or ignored areas. Through this "effective area information", the system can ensure seamless processing of the last several rows and columns of pixels of high-resolution images under any resolution conditions.

[0150] As an alternative implementation, the edge device processing module 20 includes:

[0151] An image processing unit for cropping or stitching image blocks of a processable sub-block size according to the block instruction;

[0152] An image analysis unit for performing defect detection, texture analysis, and segmentation operations on each image block;

[0153] A communication unit for uploading the output data of the image analysis unit to the result merging module 40.

[0154] In an alternative implementation of the present invention, the edge device processing module 20 can be composed of or coupled by the following three major units to better complete the processing and analysis of sub-block images:

[0155] An image processing unit, mainly responsible for receiving the block instruction issued by the large language model module 10, and performing cropping or stitching operations on the original high-resolution image according to the fields such as x_start, y_start, width, height, and edge_handling in the instruction.

[0156] Cropping scenario: When the normal sub-block is in a non-overlapping or neatly blocked area, the image processing unit can directly extract the corresponding pixel block (such as 512×512) from the source image file or stream and cache it in memory;

[0157] Stitching scenario: If the block instruction indicates edge_handling="merge_with_block":X, the image processing unit merges the edge of the currently loaded edge area with the image block corresponding to BlockID=X after loading the current edge area, and then caches the merged data to the image analysis unit;

[0158] Zero-padding scenario: If edge_handling = "padding" and valid_area < width / height, the image processing unit fills the pixels outside the valid area with 0 values to ensure that the downstream analysis algorithm can receive inputs of a unified size;

[0159] Through this block-by-block loading, the image processing unit only needs to retain the data of a processable size in memory once, avoiding large images from occupying a large amount of resources and ensuring that the subsequent image analysis unit can correctly identify the zero-padded or stitched areas.

[0160] After obtaining the sub-block images cropped / stitched by the image processing unit, the image analysis unit performs operations such as defect detection, texture analysis, and segmentation:

[0161] Defect detection: It can call algorithms based on thresholding or deep learning (CNN / Transformer) to identify surface cracks, contamination, or partial discharge traces;

[0162] Texture analysis: When the sub-block is marked as a high-risk area or requires fine-grained detection, texture extraction such as Gabor filtering and GLCM can be performed;

[0163] Segmentation operation: GrabCut, U-Net, or other segmentation methods can be selected to detect hot spot areas in infrared images or segment the device surface in visible light images;

[0164] Before execution, fields such as detection_level and priority_merge are checked. If their values indicate a high fault risk, an advanced detection model is enabled and more detailed probability distribution / mask information is output.

[0165] After the analysis is completed, the image analysis unit generates "sub-block analysis results", including:

[0166] Fields such as {"block_id":..., "defect_mask":..., "confidence_scores":..., "valid_area":...}.

[0167] When the image analysis unit completes the detection or segmentation at the sub-block level, the communication unit uploads the "sub-block analysis results" to the result merging module 40, or uploads them to a remote server for further processing in a cloud-edge collaboration scenario.

[0168] It can be transmitted by means of local shared memory queue, RPC interface, HTTP request, etc.; if the system is configured with a parallel or distributed processing mode, the communication unit will classify and package the analysis results of multiple sub-blocks according to block_id, and can attach a timestamp and algorithm version information for subsequent merging or auditing. In case of network anomalies or hardware failures, the communication unit can also cache the results and wait to send them after recovery to ensure that the locally processed data is not lost.

[0169] In this way, through the division of labor and cooperation of the above three units, the edge device processing module 20 can automatically adapt to the sizes of different sub-blocks, edge processing methods, and failure risk levels, not only reducing the demand for high-end hardware, but also being flexible to switch at the image analysis algorithm level to meet the diverse fault detection requirements of power equipment.

[0170] As an optional implementation manner, the result merging module 40 further performs a maximum value or mean value processing on the output results of adjacent sub-blocks with overlapping regions according to the fusion method specified in the block instruction. When different detection results appear at the same target pixel coordinate for multiple blocks, a preset weighted merging strategy is adopted to obtain the final result.

[0171] In an optional implementation manner of the present invention, the result merging module 40 can read a more refined "fusion method" field in the block instruction to guide the merging of detection results in the overlapping region. When an N-pixel-wide overlapping region appears in adjacent sub-blocks, or detection results are obtained for the same target pixel coordinate in multiple sub-blocks, the result merging module 40 executes the following strategy:

[0172] Maximum value merging: If the fusion method field fusion_method = "max", then take the maximum value of the detection values at the same coordinate; it is applicable to scenarios where high-risk faults (such as cracks and overheating) are preferably "reported more" rather than "missed".

[0173] Mean value merging: If the fusion method field fusion_method = "mean", then take the average of the detection values at the same coordinate; it helps to smooth detection noise.

[0174] Weighted merging: When performing weighted merging, if the large language model module 10 specifies "fusion_method = weighted" in the fusion method field, then when multiple detection values appear at the same pixel coordinate, the result merging module 40 will calculate the final pixel or defect marking value through the following formula :

[0175]

[0176] where n represents the total number of sub-blocks participating in the detection output within the overlapping region of this pixel coordinate; Indicates the detection value or defect probability of the \(i\)-th sub-block at this pixel coordinate; Indicates the weight of the \(i\)-th sub-block, which can be comprehensively determined by the confidence level, historical failure risk or priority policy of the sub-block; the numerator part Represents the accumulation of the detection values of each sub-block multiplied by their weights, and the denominator part Is used to normalize this weighted result to ensure that the final output value is within a reasonable range. Through this weighted merging formula, when there are differences in the detection results of multiple sub-blocks, the result merging module 40 can generate a more realistic final detection value based on the reliability and risk factors of different blocks, thereby reducing missed detections or misjudgments caused by edge segmentation.

[0177] When significantly different detection results appear for multiple sub-blocks at the same target pixel coordinate, if the "priority_merge" field indicates that a certain sub-block has a higher priority, its output can be directly adopted and other sub-blocks can be ignored. Through this configuration, this embodiment can more flexibly balance the risks of false detection and missed detection during overlapping merging.

[0178] In another alternative embodiment of the present invention, the large language model module 10 can be deployed on a cloud server or a high-performance computing platform, with sufficient GPU / CPU resources for model inference and block decision-making; while the edge device processing module 20 is deployed in an industrial PC, an embedded system or edge hardware at the power site and collaborates with the cloud module through network communication.

[0179] After receiving the image resolution, the size of the sub-blocks that can be processed, and the fault hot zone information provided by the device portrait module 30, the large language model module 10 on the cloud side generates corresponding block instructions and sends them to the edge device processing module 20 through HTTP / HTTPS, RPC or message queue, etc. The latter loads and detects the sub-blocks of the high-resolution image and packs and uploads the analysis results back to the cloud. When the network of the edge device is unstable, the sub-block analysis results can be cached locally first and then sent to the cloud after the link is restored. This can not only make full use of the powerful computing power of the cloud but also avoid equipping overly expensive hardware on-site.

[0180] As an alternative embodiment, the large language model module 10 is deployed on a cloud server or a high-performance computing platform, and the edge device processing module 20 is deployed in the edge hardware at the power site. The two communicate through network communication for instruction issuance and return of sub-block analysis results.

[0181] In another alternative embodiment of the present application, when the large language model module 10 generates chunking instructions, it dynamically reads incremental data such as industry specifications and historical failure cases provided by the device portrait module 30. Once it detects a change in the "processable sub-block size" or "image resolution" (for example, the camera resolution is upgraded on-site, or the edge device is replaced with hardware to support larger block processing), the model will automatically recalculate the chunking strategy. For high-risk areas with frequent failures, the detection intensity is additionally increased according to the new specifications or updated historical undetected statistics.

[0182] Specifically, each time the system acquires a new image, the cloud or local area will push its resolution and failure risk record to the large language model module 10; if it is found that the edge_device_block_size is upgraded from 512×512 to 1024×1024, the model will correspondingly reduce the number of chunks and reserve an appropriate overlap for high-risk areas to reduce the impact of edge cutting on crack detection.

[0183] Through this automatic adjustment mechanism, this embodiment can ensure that the present application can still exert high efficiency and high accuracy in fault detection after device updates, sensor resolution changes, or industry requirement upgrades, without the need for manual frequent modification of chunking rules or processing procedures, and has better scalability and self-learning potential.

[0184] As an alternative embodiment, the large language model module 10 automatically adjusts the image chunking strategy based on the industry specifications and historical failure cases provided by the device portrait module 30, and dynamically generates new chunking instructions when the processable sub-block size or image size changes.

[0185] In a specific implementation, the large language model module 10 can automatically update the image chunking strategy based on the industry specifications and historical failure case information recorded in the device portrait module 30. When conditions such as the processable sub-block size or image resolution change significantly, the large language model module 10 will dynamically generate new chunking instructions to ensure high-precision detection of power equipment images after hardware or environmental changes.

[0186] Regarding the triggering conditions of this embodiment:

[0187] For example, it can be hardware upgrade / downgrade. When the device portrait module 30 records that the processable sub-block size is upgraded from 512×512 to 1024×1024, or the edge device is reduced to 256×256 due to reasons, after the large language model module 10 detects this update, it will recalculate the number of chunks and the overlap strategy for each sub-block;

[0188] For another example, when industry specifications or fault cases change, if a new item such as "insulator crack detection requires a coverage rate of ≥95%" or "a humid environment in a certain area leads to more overheating faults" is added to the database, the model can incorporate this specification / case into the Prompt, increasing the overlapping area or raising the detection priority for high-risk areas;

[0189] For another example, regarding image size changes: when the camera resolution is increased from 4096×4096 to 8192×8192, the model needs to re-judge whether block merging, zero-padding, or expansion is required.

[0190] Regarding the further automated adjustment process, the latest portrait information can be obtained. Whenever the edge device or the cloud discovers an update in the sub-block size, image resolution, or industry specification, the device portrait module 30 will record the corresponding fields (such as edge_device_block_size = 1024×1024);

[0191] For model inference, the large language model module 10 adds new hardware parameters and historical fault cases to the Prompt and then performs internal inference. For example, if the historical missed detection rate is high for a certain fault type, the block refinement degree of that area is automatically increased;

[0192] For outputting new block instructions, in the generated JSON or XML instructions, the model updates the sub-block width, height, or the "edge_handling" strategy, and sets a "high detection level" for specific fault areas. If the industry specification requires greater coverage, "overlap_size = XX" can be written in the instructions to increase detection redundancy;

[0193] For edge device execution, the edge device processing module 20 crops the loaded image blocks according to the new instructions, and the communication unit transmits the sub-block results back to the result merging module 40. The entire process does not require manual configuration of block thresholds or rules.

[0194] Through this dynamic adjustment mechanism, this alternative embodiment enables the large language model module 10 to provide the optimal or appropriate block strategy for high-resolution power equipment images in different stages or environmental changes, thereby maintaining the overall efficient fault detection ability of the system.

[0195] As an alternative embodiment, the large language model module 10 is further configured to, based on the personalized information related to the historical fault records, operating environment, and sensing data of the power equipment obtained from the device portrait module 30, after determining the method of block processing for the power equipment image, further generate an image processing method instruction for each sub-block, and send the image processing method instruction to the edge device processing module 20, where the image processing method instruction includes:

[0196] The target algorithm type is used to indicate the image processing, defect detection, and texture analysis algorithms to be adopted for a certain sub-block;

[0197] The parameter configuration is used to mark whether additional filtering, enhancement, or segmentation thresholds are required under specific spectral or temperature distributions;

[0198] The priority information is used to indicate that the edge device processing module executes a preset second detection process for sub-blocks with a failure risk greater than or equal to a preset risk value, and a preset first detection process is adopted for sub-blocks with a failure risk less than the preset risk value;

[0199] The multi-modal fusion strategy is used to respond to the mapping relationship between visible light images and infrared images recorded in the device portrait module 30, and the image processing method instruction instructs the edge device processing module 20 to perform cross-spectral or multi-channel data fusion on the corresponding sub-blocks.

[0200] Among them, the first detection process can be the corresponding conventional image-based detection process recorded in the device portrait module 30 of the power device. And the second detection process can be the corresponding special image-based detection process for special problems recorded in the device portrait module 30 of the target power device.

[0201] Among them, the large language model module 10 is not only responsible for determining the block division strategy based on the power device image size and the size of the processable sub-blocks obtained from the device portrait module 30, but can further combine the historical fault records, operating environment, and sensing data of the power device (collectively referred to as "personalized information" hereinafter). After completing the method of dividing the image into blocks, it generates more refined image processing method instructions for each sub-block and sends the instructions to the edge device processing module 20, so as to implement customized image processing, detection, and analysis processes for sub-blocks with different risk levels or different spectral requirements.

[0202] Power devices usually face various types of faults: surface cracks, partial discharge traces, hot spot overheating, corrosion, or dirt accumulation, etc. A single image processing algorithm often has difficulty taking into account all scenarios, especially when the device parts, environmental conditions, or historical fault probabilities corresponding to each sub-block are different.

[0203] Therefore, this optional implementation proposes that after the block division scheme is determined, the large language model module 10 customizes image processing method instructions for each sub-block based on the historical fault records, operating environment (such as temperature, humidity, salt spray), and multi-modal sensing data (such as the correspondence between infrared / visible light) recorded in the device portrait, and sends them to the edge device processing module 20 for local execution.

[0204] Among them, the personalized information may include:

[0205] Historical fault records: refer to the past fault cases, maintenance records, fault occurrence frequency, and time distribution stored in the device portrait;

[0206] Operating environment: refers to the temperature, humidity, whether it is a strong salt fog or dusty environment, etc. of the environment where the device is located;

[0207] Sensing data: If the device supports the coexistence of infrared thermal images and visible light images, the device portrait may contain its registration (alignment) information or spectral mapping relationship.

[0208] These information enable the large language model module 10 to accurately evaluate the fault risk or algorithm requirements of a certain sub-block, so as to specify a more suitable algorithm category, parameters, and fusion strategy in the instruction.

[0209] In a specific implementation, the large language model module 10 first determines to divide the entire high-resolution image into several sub-blocks according to basic parameters such as the image size and the size of the processable sub-block; if there is a non-integer part at the edge, it will also mark the method of padding with zeros or splicing. After determining the sub-blocks, the model queries the device portrait module 30 again for fields such as "historical fault records, operating environment, sensing data of power equipment". According to the query results, the model will identify, for example:

[0210] Cracks frequently occurred in the corresponding area of a certain sub-block in the past (high fault risk); the current temperature of the device exceeds the standard or it is in a humid and hot environment; the image has two channels of infrared and visible light for fusion.

[0211] Furthermore, combining the fault risk level, spectral requirements, and available computing power of the hardware of the sub-block, the model outputs an image processing method instruction for each sub-block, which is used to indicate which image processing or detection algorithm should be adopted for the sub-block.

[0212] Exemplarily, when the spectrum or temperature distribution is special, it is necessary to configure the filter kernel size, enhancement parameters, or segmentation threshold in the instruction. Through the fields in the instruction, the edge device can automatically load the corresponding model weights or preprocessing parameters in the image analysis unit.

[0213] When the visible light and infrared mapping relationship is recorded in the device portrait, and the current sub-block is in a position that requires cross-spectrum alignment, the instruction can be marked as:

[0214] "fusion": "ir_visible",

[0215] "fusion_params": { "offset_x": 5, "offset_y": 2}

[0216] It is indicated that the edge device should perform detection after the image processing unit registers visible light and infrared data to improve the accuracy of defect recognition. The large language model module issues the above instructions to the edge device processing module 20 together with or in two separate times with the "block coordinate instruction".

[0217] After the edge device parses the instruction, it calls the specified algorithm process according to the field content in the image analysis unit; more computing power is invested in sub-blocks with high priority or those requiring infrared fusion.

[0218] After the processing is completed, the edge device communication unit sends the detection results of each sub-block back to the result merging module 40 or the cloud;

[0219] The large language model module 10 and the device portrait module 30 can update the historical faults or algorithm effects again for subsequent further fine-tuning and iteration.

[0220] In an alternative embodiment of the present invention, in order to determine whether a certain sub-block of the power equipment image has a high fault risk, the large language model module 10 will synthesize various information from the device portrait module 30 and real-time monitoring data to form a "fault risk value" (which can also be called "fault risk score" or "risk_score"). This score can be jointly determined by the following factors:

[0221] Firstly, it is the historical fault record. If the device portrait records that a certain area of the device (such as x, y coordinates or physical parts) has been detected with cracks or overheating multiple times in the past six months, the large language model module 10 will assign a higher basic risk score to this area in the Prompt or internal logic.

[0222] For example: If the cumulative number of historical cracks is more than 5 times, the risk_score can be initially set at 0.7 (70%); if there have never been cracks, it can be only 0.2.

[0223] Secondly, it is the operating environment or real-time sensing data. If the device is in a high-temperature, high-humidity or salt spray corrosion environment, the probability of faults (such as insulation aging, corrosion, etc.) often increases;

[0224] If the real-time sensing (temperature, vibration, etc.) shows that the current working condition is close to the rated limit of the device, it may also increase the failure rate.

[0225] The large language model module 10 can input something like: "Current temperature = 45°C, humidity = 80%, the device is under high load" in the Prompt, thereby increasing the risk_score.

[0226] Industry standards can also be recorded in the device portrait. For example, in high-altitude and low-temperature scenarios, oil-immersed transformers are prone to cracking;

[0227] The large language model increases the risk score for the corresponding area by one level according to these specifications, such as increasing from 0.5 to 0.7.

[0228] In another alternative implementation of the present invention, the large language model module 10 can also automatically select between different image detection processes recorded in the device profile module 30, and generate image processing method instructions for each sub-block based on personalized information such as the historical fault records of the power equipment, the operating environment, and the sensing data. In this implementation, the device profile module 30 stores at least two sets of detection processes in advance: the first detection process can be the conventional image-based detection process of the power equipment, which is used for ordinary or low-risk scenarios; the second detection process is a high-precision detection process specifically designed for special fault problems, which is used for more in-depth analysis when the fault risk of a certain area is relatively high.

[0229] In a specific implementation, after the large language model module 10 preliminarily determines the block division strategy according to the image size and the size of the processable sub-blocks, it will further evaluate the corresponding fault risk value of the power equipment. This risk value can be obtained by the comprehensive analysis of the large language model module 10 based on the historical fault statistics recorded in the device profile (such as cracks or hot spots have occurred many times in this area in the past year), the operating environment (such as high temperature, high humidity, salt spray corrosion, etc.), and the sensing data (such as infrared temperature distribution). Once the evaluation result shows that the risk value of a certain sub-block is greater than or equal to the preset threshold, the large language model module 10 designates this sub-block as the "second detection process" in the image processing method instructions, otherwise the "first detection process" is used by default.

[0230] In the device profile module 30, the first detection process often records the image detection means performed by the power equipment under normal circumstances, which may include relatively simple algorithms such as threshold-based segmentation, Canny edge detection, or lightweight convolutional neural networks. It has a fast execution speed and low resource occupancy, and is suitable for low-risk situations. The second detection process can be stored in the device profile module 30 with an identifier specifically for special fault problems, which often means a deep learning multi-layer model or a multi-channel fusion algorithm. It requires higher computing power and execution time, but can provide more accurate fault location in high-risk scenarios.

[0231] Exemplarily, the relevant records in the device profile can indicate the first detection process with "process_id='std_detect_v1'", the second detection process with "process_id='special_detect_v2'", and explain their respective algorithm compositions and applicable scopes.

[0232] When the large language model module 10 detects that the risk score of a certain area exceeds the threshold (e.g., 0.7), it will specify "target algorithm type ='special_detect_v2'" and mark higher priority and stricter processing parameters in the process of generating the image processing method instruction; if the evaluation result is only a low risk, "target algorithm type ='std_detect_v1'" will be written in the instruction. At the same time, if the visible light and infrared image registration information of the power equipment is also recorded in the device portrait module 30, the large language model module 10 will also attach the "multi-modal fusion strategy" to the instruction, requiring the edge device processing module 20 to perform cross-spectrum or multi-channel data alignment when analyzing this sub-block. For some scenarios with specific spectral or temperature distributions, details such as the filtering method, threshold size, or model version can be further marked in the "parameter configuration".

[0233] When the edge device processing module 20 receives the "image processing method instruction" for each sub-block, it will read the algorithm type, detection process priority, and other configuration information in the instruction field, and load the corresponding script or model according to the specific implementation method of the "first detection process" or "second detection process" recorded in the device portrait. If the instruction shows a high-priority situation, that is, perform advanced operations such as in-depth detection or cross-spectrum fusion; if it is in a normal situation, a lightweight process will be adopted to save computing power and time.

[0234] This mode significantly improves the balance between detection accuracy and efficiency: for high-risk sub-blocks, higher-precision algorithms can be used to avoid missed detections; for low-risk sub-blocks, fast detection methods can be adopted to improve the overall processing speed, so as to achieve refined and personalized analysis of power equipment images. Since these detection processes and their applicable scopes are all recorded in the device portrait module 30 in advance, the automatic reasoning process of the large language model module 10 can make full use of the device portrait information without additional manual intervention, and maintain dynamic adaptability to the system environment and fault tendency during real-time or periodic analysis.

[0235] Based on the same inventive concept, an embodiment of the present disclosure also provides a power equipment monitoring method corresponding to the power equipment monitoring system. Since the principle of solving problems by the method in the embodiment of the present disclosure is similar to that of the above-mentioned power equipment monitoring system in the embodiment of the present disclosure, the implementation of the method can refer to the implementation of the method, and the repeated parts will not be described again.

[0236] Refer to Figure 2 As shown, it is a flowchart of the power equipment monitoring method provided by the embodiment of the present application, including steps S101 to S104, where:

[0237] S101: Use a large language model to perform a chunking strategy analysis on a power equipment image with a target resolution based on the image size of the power equipment image obtained from the device portrait and the size of the sub-blocks that can be processed by the edge device, and obtain a chunking instruction; wherein, the large language model includes: a large model trained and / or fine-tuned based on power equipment scenario data;

[0238] S102: Use the edge device to perform chunking processing on the power equipment image based on the chunking instruction, and only analyze the image blocks in the area defined by the size of the sub-blocks that can be processed each time, to obtain a sub-block analysis result;

[0239] S103: Use the device portrait to store the image meta-information, historical fault records, and the size information of the sub-blocks that can be processed of the power equipment, and provide the image size and the size of the sub-blocks that can be processed to the large language model;

[0240] S104: Based on the sub-block analysis result and the address mapping corresponding to the sub-block in the chunking instruction, perform result stitching and fusion to obtain a fault identification and / or feature extraction result for the complete power equipment image with the target resolution;

[0241] Among them, the large language model determines the way of chunking processing of the power equipment image by the edge device according to the image size and the size of the sub-blocks that can be processed provided by the device portrait module 30, maps the analysis result of the sub-blocks after chunking back to the original image coordinates, and finally records the overall detection result in the device portrait module 30.

[0242] Those skilled in the art can understand that in the above method of the specific implementation manner, the writing order of each step does not mean a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined according to its function and possible internal logic. It should be understood that determining B according to A does not mean determining B only according to A, but also B can be determined according to A and / or other information.

[0243] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present invention can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0244] In the description of this specification, the description referring to terms such as "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in a suitable manner in any one or more embodiments or examples.

Claims

1. The power equipment monitoring system is characterized by: include: The large language model module performs a block strategy analysis on the target resolution power equipment image based on the image size of the power equipment image obtained from the equipment portrait and the size of the sub-block that can be processed by the edge device to obtain a block instruction; wherein the large language model module includes: a large model trained and / or fine-tuned based on the power equipment scene data; The edge device processing module performs block processing on the power equipment image based on the block instruction, and analyzes only the image blocks in the area limited by the processable sub-block size each time to obtain sub-block analysis results; An equipment portrait module, used to store image meta-information, historical fault records and processable sub-block size information of power equipment, and provide the image size and processable sub-block size to the large language model module; A result merging module, based on the sub-block analysis results and the address mapping corresponding to the sub-blocks in the block instructions, performs result splicing and fusion to obtain fault identification and / or feature extraction results for the power equipment image with full target resolution; Among them, the large language model module determines the way in which the edge device processing module performs block processing on the power equipment image according to the image size and processable sub-block size provided by the device portrait module, and the result merging module maps the sub-block analysis results after segmentation back to the original image coordinates, and finally records the overall detection results in the device portrait module.

2. The power equipment monitoring system according to claim 1, characterized in that: The large language model module calculates the number of blocks according to the image size and the processable sub-block size, and fills or splices the edge parts that cannot be divided evenly, and generates corresponding block instructions.

3. The power equipment monitoring system according to claim 1, characterized in that: The device portrait module is used to store at least one of the processable sub-block size of the edge device, the historical fault hot zone distribution, the image resolution and the multi-modal sensor data; The large language model module determines whether overlapping blocks are used in the edge area based on the processable sub-block size and the historical fault hot zone distribution.

4. The power equipment monitoring system according to claim 2, characterized in that: The large language model module generates the block instruction including: In response to the image size being divided by a processable sub-block size integer, directly dividing the entire power equipment image into a plurality of image blocks of the same size; In response to the image size not being divisible by an integer of the processable sub-block size, the edge portion that cannot be divided by an integer is padded with zeros or concatenated with an adjacent block, and valid area information of the edge sub-block is recorded in the block division instruction.

5. The power equipment monitoring system according to claim 1, characterized in that: The edge device processing module includes: An image processing unit, used for cutting or splicing image blocks of processable sub-block sizes according to the block division instruction; An image analysis unit, for performing defect detection, texture analysis, and segmentation operations on each image block; A communication unit is used to upload the output data of the image analysis unit to the result merging module.

6. The power equipment monitoring system according to claim 1, characterized in that: The result merging module takes the maximum or average value of the output results of adjacent sub-blocks with overlapping areas according to the fusion method specified in the block instruction. When multiple sub-blocks have different detection results at the same target pixel coordinate, a preset weighted merging strategy is used to obtain the final result.

7. The power equipment monitoring system according to claim 1, characterized in that: The large language model module is deployed on a cloud server or computing platform, and the edge device processing module is deployed in the edge hardware at the power site. The two communicate over the network to issue instructions and transmit back sub-block analysis results.

8. The power equipment monitoring system according to any one of claims 1 to 7, characterized in that: The large language model module automatically adjusts the image segmentation strategy based on industry specifications and historical failure cases provided by the device portrait module, and dynamically generates new segmentation instructions when the processable sub-block size or image size changes.

9. The power equipment monitoring system according to claim 1, characterized in that: The large language model module is further used to generate an image processing method instruction for each sub-block after determining the block processing method of the power equipment image based on the personalized information related to the historical fault records, operating environment and sensor data of the power equipment obtained from the equipment portrait module, and send the image processing method instruction to the edge device processing module, wherein the image processing method instruction includes: Target algorithm type, which indicates the image processing, defect detection, and texture analysis algorithms used for each sub-block; Parameter configuration, used to indicate whether additional filtering, enhancement or segmentation threshold is performed under the target spectrum or temperature distribution; Priority information, used to instruct the edge device processing module to execute a preset second detection process for a sub-block whose failure risk is greater than or equal to a preset risk value, and to use a preset first detection process for a sub-block whose failure risk is less than the preset risk value; A multimodal fusion strategy is used to respond to the mapping relationship between the visible light image and the infrared image recorded in the device portrait module, and the image processing method instruction instructs the edge device processing module to perform cross-spectral or multi-channel data fusion on the corresponding sub-blocks.

10. A method for monitoring electric power equipment, implemented based on the electric power equipment monitoring system according to any one of claims 1 to 9, characterized in that: include: Using a large language model, based on the image size of the power equipment image obtained from the equipment portrait and the size of the sub-block that can be processed by the edge device, a block strategy analysis is performed on the target resolution power equipment image to obtain a block instruction; wherein the large language model includes: a large model trained and / or fine-tuned based on the power equipment scene data; Using the edge device, based on the block instruction, the power equipment image is processed in blocks, and each time only the image blocks in the area limited by the processable sub-block size are analyzed to obtain sub-block analysis results; Using the equipment portrait, image meta information, historical fault records and processable sub-block size information of the power equipment are stored, and the image size and processable sub-block size are provided to the large language model; Based on the sub-block analysis results and the address mapping corresponding to the sub-blocks in the block instructions, the results are spliced ​​and merged to obtain fault identification and / or feature extraction results for the power equipment image with a complete target resolution; Among them, the large language model determines the way in which the edge device performs block processing on the power equipment image according to the image size and processable sub-block size provided by the device portrait, maps the sub-block analysis results after segmentation back to the original image coordinates, and records the overall detection results.

Images