Optical crossbox abnormality processing method and device, electronic equipment, medium and program product

Abstract

This application provides a method, apparatus, electronic device, medium, and program product for handling optical distribution box anomalies, applied in the field of communication network operation and maintenance technology. The method includes: acquiring initial image data collected for the optical distribution box; performing anomaly identification on the initial image data using an anomaly identification model to obtain an anomaly identification result; wherein, the anomaly identification model includes a baffle missing identification unit, a white pigtail identification unit, and a grounding wire anomaly identification unit; in response to the anomaly identification result indicating the existence of an anomaly, generating task information containing an anomaly handling instruction; when the task status indicated by the task information changes to processing completed, reacquiring verification image data collected for the optical distribution box, and performing anomaly identification on the verification image data using the anomaly identification model to verify whether the anomaly indicated by the anomaly identification result has been eliminated.

Description

Technical Field

[0001] This application relates to the field of communication network operation and maintenance technology, specifically to a method, device, electronic equipment, medium, and program product for handling optical distribution box anomalies. Background Technology

[0002] Currently, in the operation and maintenance of communication network access points, optical distribution boxes, as key access devices, rely mainly on manual methods or general AI recognition technology for daily inspection and hazard remediation. However, existing technologies have the following problems: First, general AI recognition models are not specifically optimized for the fine-grained hazards unique to optical distribution boxes (such as missing baffles, improper placement of pigtails, abnormal grounding wires, etc.), resulting in insufficient recognition accuracy under complex lighting and multi-angle scenarios, easily leading to missed detections or misjudgments. Second, existing inspection processes lack real-time edge-side inference capabilities, and the latency of image upload to the cloud for processing is high, making it difficult to meet the needs for immediate on-site feedback. In addition, the hazard remediation process lacks deep integration with AI recognition, and the remediation results still require manual verification, failing to form an automated data closed loop from hazard identification to remediation verification, resulting in low operation and maintenance management efficiency and insufficient standardization. Summary of the Invention

[0003] This application provides a method, apparatus, device, medium, and program product for handling optical distribution box anomalies, which can solve the problem that existing optical distribution box anomaly handling methods cannot be used in scenarios where user object-level detailed data cannot be accessed.

[0004] In a first aspect, embodiments of this application provide a method for handling optical distribution box anomalies, the method comprising:

[0005] Acquire initial image data for the optical distribution box;

[0006] The initial image data is subjected to anomaly identification using an anomaly identification model to obtain anomaly identification results; wherein, the anomaly identification model includes a baffle missing identification unit, a white pigtail identification unit, and a grounding wire anomaly identification unit;

[0007] In response to the anomaly identification result indicating the presence of an anomaly, task information containing anomaly handling instructions is generated;

[0008] When the task status indicated by the task information changes to processing completed, the verification image data collected for the optical distribution box is reacquired, and the verification image data is subjected to anomaly identification through the anomaly identification model to verify whether the anomaly indicated by the anomaly identification result has been eliminated.

[0009] Secondly, embodiments of this application also provide an optical distribution box anomaly handling device, comprising:

[0010] The acquisition module is used to acquire the initial image data collected from the optical distribution box;

[0011] The identification module is used to identify anomalies in the initial image data through an anomaly identification model to obtain an anomaly identification result; wherein, the anomaly identification model includes a baffle missing identification unit, a white pigtail identification unit, and a grounding wire anomaly identification unit;

[0012] The generation module is used to generate task information containing anomaly handling instructions in response to the anomaly identification result indicating the existence of an anomaly;

[0013] The verification module is used to reacquire the verification image data collected for the optical distribution box when the task status indicated by the task information changes to processing completion, and to perform anomaly identification on the verification image data through the anomaly identification model to verify whether the anomaly indicated by the anomaly identification result has been eliminated.

[0014] Thirdly, embodiments of this application also provide an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the optical distribution box anomaly handling method as described in the first aspect.

[0015] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the optical distribution box anomaly handling method as described in the first aspect.

[0016] Fifthly, a computer program product is provided, including computer instructions that, when executed by a processor, implement the steps of the optical distribution box exception handling method as described in the first aspect.

[0017] In this embodiment, the traditional single static detection is transformed into a dynamic data closed-loop process of "initial inspection - task generation - state-triggered re-inspection". This approach not only ensures the efficiency of edge inference, but more importantly, it establishes an automated verification standard that does not rely on subjective human judgment, significantly improving the standardization level and management efficiency of communication infrastructure operation and maintenance. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1This is one of the flowcharts of the optical distribution box anomaly handling method provided in the embodiments of this application;

[0020] Figure 2 This is a schematic diagram of the community optical distribution box inspection and hidden danger rectification visualization system provided in the embodiments of this application;

[0021] Figure 3 yes Figure 2 A schematic diagram of the optical distribution box inspection positioning and data acquisition module;

[0022] Figure 4 yes Figure 2 A schematic diagram of the data accumulation and visualization management module in the system;

[0023] Figure 5 This is the second flowchart of the optical distribution box anomaly handling method provided in the embodiments of this application;

[0024] Figure 6 This is a basic framework diagram of the YOLOv5 recognition model provided in the embodiments of this application;

[0025] Figure 7 This is the third flowchart of the optical distribution box anomaly handling method provided in the embodiments of this application;

[0026] Figure 8 This is a structural diagram of an optical distribution box anomaly handling device provided in an embodiment of this application;

[0027] Figure 9 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] like Figure 1 As shown and Figure 7 This embodiment provides a method for handling optical distribution box anomalies. This method achieves accurate identification and rectification verification of potential problems in optical distribution boxes by constructing a data-driven automated closed-loop process.

[0030] The following is combined with, for example Figures 2 to 4 The system architecture shown illustrates the method for handling optical distribution box anomalies provided in this application embodiment. This embodiment's method relies on an AI-based visualization system for inspecting and addressing potential hazards in community optical distribution boxes. Figure 2As shown, the system includes a fiber optic distribution box inspection and positioning data acquisition module, an AI-powered hazard identification module, a work order interaction and rectification management module, and a data accumulation and visualization management module. The fiber optic distribution box inspection and positioning data acquisition module is used to achieve precise positioning and standardized image acquisition of the fiber optic distribution boxes, providing high-quality data input for AI identification. The AI-powered hazard identification module uses an optimized YOLOv5 target detection model as its core and designs specific identification schemes for typical hazards in fiber optic distribution boxes. The work order interaction and rectification management module is used to link AI identification results with work orders, promoting a closed-loop hazard rectification process. The data accumulation and visualization management module stores inspection and rectification data and supports operation and maintenance decisions through visualization.

[0031] The aforementioned optical distribution box inspection, positioning, and data acquisition module is used to achieve precise positioning and standardized image acquisition of the optical distribution box, providing high-quality data input for AI recognition. It includes the following sub-units:

[0032] Optical Distribution Box Information Matching Subunit: It interfaces with the Home Installation and Maintenance-Outdoor Support Module of the Operation and Maintenance Platform. By inputting the community name and box number, it automatically retrieves the basic information of the target optical distribution box (such as model, grid of origin, and historical inspection records) to ensure accurate matching of the inspection target.

[0033] Inspection location recording sub-unit: Through the positioning function of the mobile terminal, the latitude and longitude coordinates of the inspection personnel's on-site check-in are automatically collected and linked to the inspection work order of the corresponding optical distribution box to realize the traceability of the inspection location;

[0034] Standardized Image Acquisition Subunit: Provides image acquisition guidelines, clearly requiring the acquisition of six types of standardized images of the optical distribution box (A / B side with door closed, A / B side with door open, grounded close-up, and overall scene inside the box), and standardizes constraints on image resolution and shooting angle to avoid affecting AI recognition accuracy due to image quality. After acquisition, the images are uploaded to the system backend in real time via mobile terminal.

[0035] The AI-powered hazard identification module, based on the optimized YOLOv5 target detection model, is designed with specific identification solutions for typical hazards in optical distribution boxes (missing baffles, unauthorized placement of white fiber optic cables, and abnormal grounding wires). It achieves high-precision and high-efficiency automatic hazard identification and includes the following sub-units:

[0036] Model training and optimization subunit:

[0037] Dataset construction sub-unit: Collect images of optical distribution boxes under different scenarios (lighting, weather, box type), label them with three types of hidden danger tags: "missing baffle", "white pigtail", and "abnormal grounding wire", to form a dedicated dataset of hidden dangers for optical distribution boxes;

[0038] Data augmentation subunit: Preprocesses the dataset, including image rotation, horizontal / vertical flipping, lighting intensity adjustment, and noise addition (Gaussian noise, salt-and-pepper noise), to increase the diversity of the dataset and avoid model overfitting;

[0039] Model structure adjustment subunit: Based on the YOLOv5s architecture, the network structure is adjusted to address the potential hazards of optical distribution boxes. A "hazard feature enhancement branch" is added to the Backbone layer, which extracts fine-grained features such as baffle edges, fiber optic lines, and yellow-green grounding wires through 3×3 convolutional kernels. The feature fusion strategy is optimized in the Neck layer, which adopts a dual-path fusion of shallow and deep features to improve the recognition ability of small targets (such as thin fiber optics).

[0040] Loss function optimization subunit: FocalLoss is used to replace the traditional cross-entropy loss function to solve the class imbalance problem of "more normal samples and fewer hidden danger samples" in the hidden danger samples. By adjusting the focusing parameter (γ=2), the model's attention to hidden danger samples is enhanced.

[0041] Multi-hazard identification sub-unit:

[0042] The model training and optimization subunit includes a dataset construction subunit, a data augmentation subunit, a model structure adjustment subunit, and a loss function optimization subunit. The dataset construction subunit collects images of light distribution boxes from multiple scenes and labels them with three categories of potential hazards. The data augmentation subunit expands the dataset through rotation, flipping, illumination adjustment, and noise addition. The model structure adjustment subunit, based on the YOLOv5s architecture, adds a hazard feature enhancement branch and a cross-layer feature fusion path. The loss function optimization subunit uses FocalLoss to address the class imbalance problem; its mathematical expression is:

[0043]

[0044] in: This represents the probability that the model predicts a sample to be of the positive class, if the sample's true class is positive. If the true class of the sample is negative, , To predict the original probability value of a sample belonging to a certain category for the model; This is a balancing factor used to adjust the proportion of positive and negative samples contributing to the loss. Different settings are applied to achieve this. The value controls the degree of attention the model pays to the minority class. When the sample classes are imbalanced, a larger value is assigned to samples of the rare class. value; To focus parameters, This is used to adjust the contribution of easily classified and difficult-to-classify samples to the loss. The larger the value, the greater the weight given to reducing the loss of easily classified samples, and the more the model focuses on difficult-to-classify samples; by adjusting... and These two parameters can effectively alleviate the problem of model performance degradation caused by class imbalance during training;

[0045] The multi-hazard identification subunit includes a baffle missing identification subunit, a white pigtail identification subunit, and a grounding wire abnormality identification subunit;

[0046] Baffle Missing Recognition Subunit: Obtains the baffle outline through edge detection algorithm and utilizes key point localization model. Perform keypoint coordinate prediction; let the set of detected keypoints be... Based on a pre-trained complete baffle keypoint distribution model, the Euclidean distance error between the current keypoint distribution and the standard distribution is calculated. :

[0047]

[0048] when hour, The missing baffle is determined based on the set key point error threshold.

[0049] White pigtail identification subunit: Based on the HSV color space, sets the hue of the white pigtail. saturation Brightness Threshold range , , For each pixel in the image If its HSV value satisfy:

[0050]

[0051] The pixel is then marked as a candidate white region; subsequent morphological opening operations are then performed. Closing operations ,in For the input image, Using structural elements, noise is removed and holes are filled to ultimately select the white pigtail area;

[0052] Grounding wire anomaly identification subunit: fusion of color features Morphological characteristics Spatial relationship characteristics Perform grounding wire status determination; define comprehensive evaluation function:

[0053]

[0054] in, , , The weight coefficients are determined through optimization using the training set; when hour, The grounding wire is determined to be abnormal based on the set abnormal threshold.

[0055] Model Inference Acceleration Subunit: The TensorRT acceleration engine is used to optimize the trained model, including model quantization (quantizing 32-bit floating-point weights to 16 bits to reduce computation), layer fusion (merging adjacent convolutional and activation layers to reduce data transmission), and kernel automatic tuning (adapting to mobile terminal hardware architecture), to achieve efficient inference on edge devices (such as inspection mobile phones and tablets) and meet the on-site "real-time recognition" requirements.

[0056] The work order interaction and rectification control module is used to link AI recognition results with work orders, promoting a closed loop for hazard rectification, and includes the following sub-units:

[0057] Work order generation and association sub-unit: Receives AI recognition results. If no hidden danger is identified, the inspection work order is automatically marked as "compliant" and archived. If a hidden danger is identified, the hidden danger type, optical junction box information, and recognition confidence level are automatically extracted to generate a "hidden danger rectification work order" and associate it with the original inspection personnel's mobile terminal account.

[0058] Work order dispatch and reminder sub-unit: Through the message push function of the operation and maintenance platform, send rectification work order reminders to the inspection personnel, specifying the rectification requirements (such as "complete the baffle replacement within 3 days") and the rectification location (the location of the hidden danger based on AI recognition and marking);

[0059] The rectification result review sub-unit receives the rectified images submitted by the inspection personnel, automatically calls the AI ​​hazard intelligent identification module for secondary identification, and if the review result is "hazard has been eliminated", the work order is marked as "closed loop" and archived; if the review result is "hazard still exists", the work order is returned to the inspection personnel, triggering a secondary rectification reminder.

[0060] The data accumulation and visualization management module is used to store inspection and rectification data and support operation and maintenance decisions through visualization. It includes the following sub-units:

[0061] Optical Distribution Box Data Asset Sub-unit: Construct a basic database for optical distribution boxes, storing full data such as the location, model, inspection records (including images), hazard types, and rectification records of optical distribution boxes, forming structured data assets;

[0062] Multi-dimensional report sub-units: Supports automatic statistics of indicators such as inspection completion rate, number of hidden dangers found, and rectification completion rate by grid, unit, and time (daily / weekly / monthly) dimensions, and generates standardized reports;

[0063] GIS map visualization sub-unit: The location of the optical distribution box is linked to the status of the hidden dangers on the GIS map. The distribution of hidden dangers is displayed intuitively through different color markings (such as green "normal", yellow "to be rectified", and red "overdue and not rectified"), which supports managers to quickly locate high-risk areas.

[0064] Please see Figure 1 The optical distribution box anomaly handling method includes the following steps:

[0065] Step S100: Acquire initial image data for the optical distribution box.

[0066] Specifically, before image acquisition, the system first performs precise matching and location recording of the inspection targets. After arriving at the optical distribution box site with a mobile terminal, the inspection personnel log into the designated function module of the operation and maintenance platform through the terminal, enter the community name and optical distribution box number, and the system's optical distribution box information matching subunit automatically retrieves the basic information of the target optical distribution box, including model, grid affiliation, and historical inspection records, ensuring accurate matching between the inspection target and the system record and avoiding false inspections due to confusion caused by box numbering. At the same time, the inspection location recording subunit automatically collects the latitude and longitude coordinates of the inspection personnel's on-site check-in through the mobile terminal's positioning function, and associates these coordinates with the corresponding optical distribution box's inspection work order, realizing the traceability of the inspection location and providing spatial data support for subsequent GIS map visualization and operation and maintenance auditing.

[0067] After confirming the inspection targets and recording their locations, the system generates an electronic inspection work order, which explicitly requires the collection of six types of standardized images. The initial image data collection is not a random snapshot, but follows a preset standardized collection specification to ensure input consistency for subsequent algorithm processing. In this embodiment, the standardized collection specification requires the collection of six specific views for the same optical distribution box, including an A-side image with the door closed, a B-side image with the door closed, an A-side image with the door open, a B-side image with the door open, a close-up image of the grounding terminal, and an overall scene image inside the box. Of the A-side and B-side images, one can be the front and the other the back. This multi-dimensional view combination can cover key detection dimensions such as the external integrity of the optical distribution box, the standardization of internal wiring, and grounding safety, effectively avoiding blind spots caused by a single shooting angle. It should be understood that although this embodiment lists six types of views, in other embodiments, depending on the detection target, it can be adjusted to four or eight types of views, as long as it can provide a complete representation of the key parts of the target facility. After data acquisition, the mobile terminal encrypts and uploads the images to the system backend. The system performs standardized preprocessing on the uploaded images, including adjusting the image size to a preset resolution (e.g., 640×640 pixels) and normalizing the pixel values ​​to eliminate imaging differences between different terminal devices. This standardized data acquisition and preprocessing provides a high-quality, structured input foundation for subsequent automated analysis.

[0068] Step S200: The initial image data is subjected to anomaly recognition by an anomaly recognition model to obtain anomaly recognition results; wherein, the anomaly recognition model includes a baffle missing recognition unit, a white pigtail recognition unit, and a grounding wire anomaly recognition unit.

[0069] Specifically, the anomaly detection model is not a single, general-purpose detector, but a composite model architecture integrating multiple specialized processing units. During anomaly detection, the system can input initial image data into these three specialized units for parallel processing, or process it serially according to a preset priority strategy. The baffle missing unit focuses on extracting and matching geometric features to determine the integrity of the enclosure structure; the white pigtail identification unit focuses on analyzing color space and texture features to detect illegally laid cables; and the grounding wire anomaly detection unit integrates spatial relationships and connection status features to assess the compliance of the grounding system. This modular model design allows each unit to be deeply optimized for specific types of hazards, avoiding the problem of mutual accuracy interference in multi-task scenarios of general-purpose models, and significantly improving the accuracy of identifying fine-grained hazards.

[0070] Step S300: In response to the anomaly identification result indicating the existence of an anomaly, task information containing anomaly handling instructions is generated.

[0071] Specifically, the system executes branch decision logic based on the anomaly identification results. If the anomaly identification result indicates that no hidden danger has been identified, the system automatically marks the current inspection work order as "compliant" and archives the work order along with the identification result to the data accumulation module, forming the operation and maintenance history of the optical distribution box. Compliance confirmation can be completed without manual intervention. Here, not identifying a hidden danger means not identifying an anomaly. If the anomaly identification result indicates that an anomaly exists, task information is generated. This task information is a structured data carrier connecting the identification and handling stages. Its essence is a set of machine-readable instructions, rather than traditional manual paper documents or verbal notifications. The data structure of this task information includes at least an anomaly type identifier, anomaly location annotation data (such as bounding box coordinates or key point sets), and processing time limit parameters. For example, when a missing baffle is identified, the system automatically generates a JSON format data packet containing a "missing baffle" type label, pixel coordinates of the missing part, and a "72-hour" time limit field. The "72-hour" time limit field indicates that the handling of this anomaly must be completed within 72 hours. Meanwhile, the system sends rectification work order reminders to inspection personnel via message push services (such as SMS and APP push), clearly specifying rectification requirements (such as "complete the baffle installation within 3 days") and the location information of potential hazards based on anomaly identification markings, ensuring that inspection personnel can receive the information in a timely manner and accurately locate the rectification areas. This structured task information generation and automatic dispatch mechanism ensures the accurate transmission and traceability of anomaly handling instructions, providing standardized data objects for subsequent automated status transitions.

[0072] Step S400: When the task status indicated by the task information changes to processing completed, the verification image data collected for the optical distribution box is reacquired, and the verification image data is subjected to anomaly identification through an anomaly identification model to verify whether the anomaly indicated by the anomaly identification result has been eliminated.

[0073] Specifically, after receiving a rectification work order, inspection personnel arrive at the optical distribution box site with the necessary tools (such as spare baffles, cable ties, grounding wire accessories, etc.). Based on the hazard location markings included in the work order, they perform targeted rectification operations, such as replacing missing baffles, tidying up improperly laid fiber optic cables, or reconnecting loose grounding wires. After rectification, the inspection personnel submit a task completion instruction via a mobile terminal and, according to the standardized specifications in step S100, re-collect the six types of standardized images of the optical distribution box as verification image data, and submit them to the system backend.

[0074] The system internally maintains a state machine for task information, monitoring changes in the task status field in real time. Only when a clear status change signal is received—that is, when the task status changes from "pending" or "processing" to "processing completed"—will the image acquisition process for verification be automatically triggered. This triggering mechanism is entirely based on objective changes in data status, rather than relying on subjectively initiated verification requests. Once triggered, the system will call the acquisition terminal to reacquire the verification image data according to the standardized specifications in step S100, and then call the anomaly identification model again for secondary inference. If the secondary inference result shows that the anomaly has been eliminated, the task status will be updated to "closed loop"; if the anomaly still exists, new task information will be automatically generated or the original task status will be rolled back. This state-triggered automated verification mechanism fundamentally eliminates false or incomplete rectification, ensuring the authenticity of maintenance data and the rigor of closed-loop management, achieving a fully unmanned data closed loop from hazard discovery to rectification verification.

[0075] Through steps S100 to S400, this embodiment transforms the traditional single static detection into a dynamic data closed-loop process of "initial inspection - task generation - state-triggered re-inspection". This approach not only ensures the efficiency of edge inference but, more importantly, establishes an automated verification standard that does not rely on subjective human judgment, significantly improving the standardization level and management efficiency of communication infrastructure operation and maintenance.

[0076] Furthermore, such as Figure 5 As shown, after the task status is updated to closed loop, the system automatically aggregates all process data to the data accumulation and visualization management module for storage and visualization. Specifically, the optical distribution box data asset sub-unit constructs and maintains the optical distribution box basic database. This database uses the unique optical distribution box number as the primary key and stores all operation and maintenance data for each optical distribution box, including but not limited to: the location information of the optical distribution box (latitude and longitude coordinates, cell and grid), equipment model, six types of standardized images collected during each inspection and their corresponding anomaly identification results, records of the types and severity of each hidden danger, the complete flow trajectory of rectification work orders (including generation time, dispatch time, rectification completion time, and review results), and comparison images before and after rectification. This data is organized in time sequence to form a complete operation and maintenance file for each optical distribution box, providing a traceable data foundation for subsequent fault trend analysis, equipment life assessment, and operation and maintenance quality evaluation.

[0077] The multi-dimensional reporting sub-unit, based on the full data stored in the optical distribution box data asset sub-unit, supports automatic statistics of key operation and maintenance indicators by multiple dimensions such as grid, unit, and time (day / week / month). These indicators include at least: inspection completion rate (the proportion of completed optical distribution boxes to the total planned inspections), number of hidden dangers discovered (categorized by hidden danger type), rectification completion rate (the proportion of closed-loop work orders to the total number of generated work orders), and average rectification cycle (the average time from work order generation to closure). The system automatically generates standardized reports according to a preset cycle (e.g., daily) and pushes these reports to administrator accounts via message push service, enabling administrators to promptly grasp the operation and maintenance status of each area.

[0078] The GIS map visualization sub-unit links the geographical location information of optical distribution boxes with their real-time operation and maintenance status to the GIS map interface, intuitively displaying the distribution of potential hazards through differentiated color markings. For example, optical distribution boxes currently in a "normal" state are marked in green, those "pending rectification" in yellow, and those "overdue for rectification" in red. Managers can easily view the distribution of potential hazards for all optical distribution boxes within their jurisdiction through the GIS map, quickly locate high-risk areas and boxes that have overdue for rectification, and thus formulate targeted operation and maintenance resource allocation plans. Furthermore, the GIS map also supports clicking on individual optical distribution box icons to expand their detailed operation and maintenance files, including historical inspection images, hazard records, and rectification timelines, facilitating case tracing and analysis by managers.

[0079] Optionally, the step of performing anomaly detection on the initial image data using an anomaly detection model to obtain an anomaly detection result includes:

[0080] The baffle missing identification unit in the anomaly identification model performs baffle missing identification on the initial image data based on the key point distribution to obtain a first identification result.

[0081] The white pigtail recognition unit in the anomaly recognition model performs white pigtail recognition on the initial image data based on color space threshold and morphological operations to obtain a second recognition result.

[0082] The grounding wire anomaly identification unit in the anomaly identification model performs grounding wire anomaly identification on the initial image data based on multi-dimensional feature fusion to obtain a third identification result.

[0083] The anomaly identification results include the first identification result, the second identification result, and the third identification result.

[0084] In this implementation, a differentiated feature extraction strategy is adopted for different types of hazards, rather than relying on a single general feature. This enables the model to adapt to different detection needs such as geometric shape, color texture and spatial relationship, and significantly improves the comprehensive recognition accuracy of multiple types of fine-grained hazards in complex scenes.

[0085] Optionally, the step of performing baffle missing recognition on the initial image data based on key point distribution through the baffle missing recognition unit in the anomaly recognition model to obtain a first recognition result includes:

[0086] The current set of key points of the optical distribution box is obtained through the baffle missing identification unit in the anomaly identification model. Based on a preset standard key point distribution model corresponding to the optical distribution box, the Euclidean distance error between the current set of key points and the standard key point distribution model is calculated. ;

[0087] Based on the Euclidean distance error, the baffle missing identification unit in the anomaly identification model... The first identification result, wherein, when the Euclidean distance error When the error exceeds a preset key point error threshold, the first identification result indicates that the optical distribution box has a missing baffle; when the Euclidean distance error... When the error is less than or equal to the key point error threshold, the first identification result indicates that the optical distribution box does not have a missing baffle.

[0088] The white pigtail recognition unit in the anomaly recognition model performs white pigtail recognition on the initial image data based on color space thresholds and morphological operations to obtain a second recognition result, including:

[0089] The white pigtail recognition unit in the anomaly recognition model sets the hue based on the HSV color space. saturation and brightness The threshold range is used to filter out pixels located within the threshold range as candidate regions, and morphological opening and closing operations are performed on the candidate regions to obtain the target texture region; the white pigtail recognition unit in the anomaly recognition model is used to perform white pigtail recognition on the target texture region to obtain the second recognition result;

[0090] The grounding wire anomaly identification unit in the anomaly identification model performs grounding wire anomaly identification on the initial image data based on multi-dimensional feature fusion to obtain a third identification result, including:

[0091] The color features of the optical distribution box are extracted using the grounding wire anomaly identification unit in the anomaly identification model. Morphological characteristics Spatial relationship characteristics The grounding wire anomaly identification unit in the anomaly identification model is used to identify anomalies based on a comprehensive evaluation function. Calculate the evaluation score; using the grounding wire anomaly identification unit in the anomaly identification model, generate the third identification result based on the evaluation score, wherein when the evaluation score is less than a preset anomaly threshold, the third identification result indicates that the optical distribution box has a grounding wire anomaly; when the evaluation score is greater than or equal to the anomaly threshold, the third identification result indicates that the optical distribution box does not have a grounding wire anomaly. , and These are the preset weighting coefficients.

[0092] In this embodiment, the internal processing logic of the anomaly recognition model mentioned in the above embodiments is further refined. Specifically, the baffle missing recognition unit in the anomaly recognition model performs baffle missing recognition on the initial image data based on the key point distribution; the white pigtail recognition unit in the anomaly recognition model performs white pigtail recognition on the initial image data based on color space threshold and morphological operations; and the grounding wire anomaly recognition unit in the anomaly recognition model performs grounding wire anomaly recognition on the initial image data based on multi-dimensional feature fusion.

[0093] Specifically, this differentiated identification strategy is specifically designed to address the varying physical characteristics of potential hazards within optical distribution boxes. It aims to solve the problem that general models struggle to accommodate diverse heterogeneous features such as geometric structure, color texture, and spatial relationships. For different types of hazards, the system invokes different algorithm branches for processing, and finally logically aggregates the outputs of each branch to obtain a comprehensive and accurate target anomaly status.

[0094] For the geometrically related potential hazard of missing baffles, the core of the assessment lies in the quantitative verification of structural integrity. One implementation method is to obtain the current set of key points of the optical distribution box. Based on a preset standard keypoint distribution model, the Euclidean distance error between the current set of keypoints and the standard keypoint distribution model is calculated. The formula for calculating D is as follows:

[0095]

[0096] when hour, The missing baffle is determined by setting a critical point error threshold.

[0097] In the specific implementation process, the current key points are set. The acquisition can be achieved by performing skeleton extraction or corner detection algorithms on the baffle region in the initial image data, where... The number of key points is represented by the four vertices and the center point of the baffle. Indicates the number of times in the current image The actual pixel coordinates of each key point, and This represents the theoretical coordinates of the key point under standard defect-free conditions. These theoretical coordinates can be obtained in advance through statistical averaging of a large number of qualified samples. Euclidean distance error. This characterizes the overall deviation of the current baffle attitude from the standard attitude. Keypoint error threshold. The setting has a clear physical meaning, representing the maximum permissible mechanical tolerance during the installation of the optical distribution box baffle. For example, considering thermal expansion and contraction and installation gaps, this threshold is typically set to the pixel value corresponding to 3% to 5% of the baffle's diagonal length. When the threshold is exceeded, it indicates that the deformation or displacement of the baffle has exceeded the normal assembly range, thus objectively determining that it is missing or improperly installed, effectively avoiding false alarms caused by slight tilting of the shooting angle.

[0098] Regarding the potential safety hazard of illegally laid white fiber optic cables (a color-texture issue), the difficulty in determining its suitability lies in the color stability under complex lighting conditions and the cable's continuity. One implementation method involves setting a hue based on the HSV color space. saturation and brightness The threshold range is used to filter out pixels that meet the threshold range as candidate regions, and morphological opening and closing operations are performed on the candidate regions to obtain the target texture region.

[0099] In some embodiments of this application, the white pigtail identification subunit: sets the hue of the white pigtail based on the HSV color space. saturation Brightness Threshold range , , For each pixel in the image If its HSV value satisfy:

[0100]

[0101] The pixel is then marked as a candidate white region; subsequent morphological opening operations are then performed. Closing operations ,in For the input image, Using structural elements, noise is removed and holes are filled, ultimately selecting the white pigtail area.

[0102] The reason for choosing the HSV color space instead of the traditional RGB space is that the HSV model decouples color information from luminance information, where hue... and saturation The component exhibits strong robustness to changes in light intensity, adapting to various environments such as dim lighting or direct sunlight within the optical distribution box. Regarding specific parameter settings, for white fiber optic pigtails, the hue... The threshold range is usually set to the full hue range or a specific low saturation range, saturation Set to a lower value (e.g., 0-30), brightness A higher value (e.g., 200-255) is set to accurately locate the white highlight area. However, candidate areas obtained solely through color thresholding often contain noise or fiber breaks caused by shadows. Therefore, this embodiment introduces a morphological processing mechanism: first, an opening operation (erosion followed by dilation) is performed to remove isolated noise caused by reflective points or dust using small-sized structuring elements; then, a closing operation (dilation followed by erosion) is performed to fill the tiny holes caused by occlusion or shadows on the fiber and reconnect broken line segments into complete connected regions. After this series of processes, the final target texture area can accurately reflect the true direction and coverage of the white fiber, providing a reliable basis for subsequent judgment on whether it violates functional area regulations.

[0103] For spatially related hazards such as grounding wire anomalies, a single feature is insufficient because it involves a comprehensive assessment of color, shape, and relative position. One approach is to extract the color characteristics of the optical distribution box. Morphological characteristics Spatial relationship characteristics Through comprehensive evaluation function Calculate the evaluation score when hour, The grounding wire is determined to be abnormal based on the set abnormal threshold.

[0104] In this evaluation system, color characteristics The clarity and purity of the yellow-green bicolor marking used to characterize the grounding wire can be quantified by color histogram matching degree or the proportion of a specific color; morphological characteristics Used to describe the curvature, aspect ratio, and end crimping shape of the grounding wire, reflecting the standardization of its physical connection; spatial relationship characteristics. This defines the relative distance and angle between the grounding terminal and the enclosure grounding point to ensure correct connection positioning. Weighting coefficient , and The scores are not arbitrarily assigned, but automatically optimized through multivariate regression analysis or machine learning training on historical labeled datasets to ensure that the contribution of each dimension's features to the final judgment aligns with actual operational experience. (Comprehensive evaluation score) Essentially, it is a weighted confidence index based on multi-dimensional features; a higher score indicates that the grounding wire condition is closer to the standard specification. When When the value falls below a preset anomaly threshold, it means that at least one or more dimensions of the features have deviated significantly from the normal range, and the system determines that there is an anomaly in spatial relationships. This multi-feature fusion scoring mechanism significantly improves the accuracy and generalization ability of identifying complex anomalies such as loose grounding wires, incorrect connections, and detached labels compared to a single threshold judgment.

[0105] Optionally, the step of performing anomaly detection on the initial image data using an anomaly detection model to obtain an anomaly detection result includes:

[0106] The anomaly detection model is used to detect anomalies in the initial image data and output the initial detection result.

[0107] The initial identification result is subjected to confidence threshold filtering and nonmaximum suppression processing to obtain the anomaly identification result.

[0108] To meet the need for real-time on-site feedback, some embodiments of this application can perform extreme lightweight adaptation on the trained model. Specifically, model quantization converts the original 32-bit floating-point (FP32) weights into 16-bit integers (INT16) or even 8-bit integers (INT8), compressing the model size by more than 50% with almost no loss of accuracy (accuracy loss is usually controlled within 1%), and significantly reducing memory bandwidth requirements. Layer fusion technology combines adjacent convolutional layers (Conv), batch normalization layers (BN), and activation function layers (ReLU) into a single operator, reducing the number of GPU memory read / write operations on intermediate feature maps and kernel startup overhead. Kernel auto-tuning searches for the optimal operator implementation strategy for the specific hardware architecture of edge devices (such as mobile GPUs or NPUs). After the above comprehensive optimization, the single-frame inference time of the model on mainstream inspection terminals can be reduced from more than 200ms before optimization to less than 50ms, and memory usage is reduced by about 60%, thus truly realizing real-time hazard perception in offline mode.

[0109] like Figure 6As shown, the network architecture of the anomaly detection model itself supports lightweight design. This model is based on an improved YOLOv5s architecture and consists of three main parts: Backbone, Neck, and Output. The Backbone is composed of a Focus module, multiple C3 modules, and an SPP module cascaded together to extract multi-scale features. The Neck uses a structure combining a Feature Pyramid Network (FPN) and a Path Aggregation Network (PANet), achieving the fusion of deep and shallow features through Concat and upsampling operations, thus enhancing the detection capability of small targets. Building upon this, this invention adds a hazard feature enhancement branch after the C3 module of the Backbone and adds cross-layer connections in the Neck layer. While these structural improvements enhance feature representation capabilities, they also increase computational cost. Therefore, the aforementioned layer fusion and quantization optimizations are crucial for balancing accuracy and speed, ensuring that the improved model can still run smoothly on edge devices.

[0110] In the post-processing stage of inference output, confidence threshold filtering and non-maximum suppression are performed on the initial identification results to obtain preliminary abnormal features. This step is the last line of defense to ensure the real-time performance and accuracy of the edge device. Confidence threshold filtering (e.g., setting the threshold to 0.25) can quickly eliminate a large number of background false detections and low-quality prediction boxes, reducing the subsequent computational load; non-maximum suppression (NMS) removes redundant detection boxes for the same potential target by calculating the intersection-over-union ratio (IoU, e.g., setting the threshold to 0.45), retaining only the best prediction. Under the constraint of limited computing power of edge devices, this efficient post-processing mechanism avoids sending massive amounts of raw prediction results back to the cloud or performing complex clustering operations, ensuring that the end-to-end latency from image acquisition to result output meets the interactive experience requirements of on-site inspection. It should be understood that although this embodiment uses YOLOv5s as an example, the above lightweight deployment and training optimization strategies are also applicable to other target detection networks, as long as they can achieve real-time inference on edge devices and reach the preset accuracy indicators.

[0111] Optionally, the process of obtaining the anomaly recognition model includes:

[0112] Construct an optical distribution box anomaly dataset, which includes samples of optical distribution box images labeled with missing baffles, samples of optical distribution box images labeled with white pigtails, and samples of optical distribution box images labeled with grounding wire anomaly tags.

[0113] Data augmentation processing is performed on the optical crossbox anomaly dataset to obtain the target dataset. The data augmentation processing includes at least one of image rotation, flipping, illumination intensity adjustment, and noise addition.

[0114] The initial anomaly detection model is trained using the Focal Loss loss function based on the target dataset to obtain the trained detection model.

[0115] The trained recognition model is subjected to lightweight optimization processing to obtain the anomaly recognition model, wherein the lightweight optimization processing includes at least one of model quantization, layer fusion and kernel auto-tuning.

[0116] In this embodiment, the complete training and deployment process of the anomaly detection model is described in detail.

[0117] During the dataset construction phase, images of optical distribution boxes from different communities and environments were collected, covering normal conditions and three types of potential hazard states: missing baffles, non-compliant pigtails, and abnormal grounding wires. Annotation tools were used to label the hazard areas in the images, generating annotation files in VOC or COCO format. The labeled dataset was then divided into training, validation, and test sets according to a preset ratio (e.g., 7:2:1). When constructing the optical distribution box anomaly dataset, in addition to collecting clear images under normal lighting conditions, extreme scenario samples such as low light, direct strong light, lens contamination, and component obstruction were specifically included to cover the long-tail distribution in real-world operation and maintenance.

[0118] During the data augmentation phase, data augmentation processing is performed on the training set, including at least one of the following: image rotation (e.g., random rotation from 0° to 90°), horizontal / vertical flipping (probability 0.5), illumination intensity adjustment (amplitude ±15% to ±30% of the baseline brightness), and noise addition (Gaussian noise variance 0.01-0.03, salt-and-pepper noise). This targeted augmentation strategy forces the model to learn the essential structural features of the hidden dangers rather than the background texture, effectively preventing overfitting on specific cells or specific types of enclosures.

[0119] During the model structure adjustment phase, based on the YOLOv5s open-source architecture, a hazard feature enhancement branch was added after the C3 module in the Backbone layer. This branch uses 3×3 convolution and ReLU activation functions to extract fine-grained hazard features. In the feature fusion stage of the Neck layer, cross-layer connections between shallow and deep features were added to improve the ability to identify small targets. When initializing the model weights, pre-trained YOLOv5s weights were used as the basis, and the weights of the newly added branches were randomly initialized only.

[0120] During model training, Focal Loss is used instead of the traditional cross-entropy loss function to address the extreme imbalance between positive and negative samples. In the optical distribution box inspection scenario, the number of normal samples far exceeds the number of potential hazard samples, and most normal areas are easily classified. This can easily lead to model training being dominated by a large number of simple samples, neglecting minor, difficult-to-classify hazards. The mathematical expression for Focal Loss is:

[0121]

[0122] in, This represents the probability that the model predicts a sample to be of the positive class, if the sample's true class is positive. If the true class of the sample is negative, , To predict the original probability value of a sample belonging to a certain category for the model; This is a balancing factor used to adjust the proportion of positive and negative samples contributing to the loss. Different settings are applied to achieve this. The value controls the degree of attention the model pays to the minority class. When the sample classes are imbalanced, a larger value is assigned to samples of the rare class. value; To focus parameters, This is used to adjust the contribution of easily classified and difficult-to-classify samples to the loss. A larger value means a greater weight is given to reducing the loss on easily classified samples, and the model pays more attention to difficult-to-classify samples. This can be achieved by adjusting... and These two parameters can effectively alleviate the problem of model performance degradation caused by class imbalance during training.

[0123] In terms of specific training parameter settings, the batch size was set to 16, the number of epochs to 100, and the initial learning rate (lr) to 0.001, employing a cosine annealing learning rate scheduling strategy. Focal Loss was used to calculate the model loss, and the model accuracy (e.g., mAP@0.5) was monitored in real-time using the validation set. Training was stopped and the optimal model weights were saved when the validation set accuracy showed no improvement for 10 consecutive epochs. In this embodiment, γ was preferably set to 2.0, and α was adaptively set according to the proportion of samples from each category. Experimental data showed that compared to traditional loss functions, introducing Focal Loss improved the average accuracy (mAP@0.5) for detecting small targets such as white pigtails by approximately 5%-8%, significantly improving the false negative rate.

[0124] In the lightweight optimization phase, lightweight optimization processing is performed on the trained anomaly detection model, including at least one of the following: model quantization (quantizing 32-bit floating-point weights to 16-bit), layer fusion (merging adjacent convolutional layers and activation layers), and automatic kernel tuning (adapting to mobile terminal hardware architecture). The optimal model is then optimized using the TensorRT acceleration engine to generate a .engine model file adapted for mobile terminals.

[0125] During the deployment phase, the optimized model is deployed to the mobile terminals of inspection personnel (such as mobile apps), and image acquisition, uploading, and result display functions are integrated to achieve efficient inference on edge devices and meet the real-time recognition needs on site.

[0126] Optionally, the generation of task information containing exception handling instructions includes:

[0127] Extract the anomaly type, anomaly location annotation data, and processing time limit requirements corresponding to the anomaly identification results, generate the task information, and associate the task information with the target processing terminal;

[0128] When the task status indicated by the task information changes to processing completion, the verification image data acquired for the optical distribution box is reacquired, and the verification image data is subjected to anomaly identification using the anomaly identification model to verify whether the anomaly indicated by the anomaly identification result has been eliminated, including:

[0129] When the task status indicated by the task information changes to processing completed, in response to receiving the task completion instruction from the target processing terminal, the image re-acquisition process is triggered to acquire the review image data;

[0130] The anomaly detection model is used to perform anomaly detection on the verification image data to obtain the verification result;

[0131] If the verification and identification result indicates that the anomaly has not been eliminated, secondary task information containing the anomaly location annotation data is generated, and the task status is updated to pending processing until the verification and identification result indicates that the anomaly has been eliminated.

[0132] Specifically, the system extracts the anomaly type, anomaly location annotation data, and processing time limit requirements corresponding to the target anomaly state, generates task information, and associates this task information with the target processing terminal. The task information is not an unstructured text description or a manually filled-out form, but a standardized machine-readable data object. In this embodiment, the data object is encapsulated in JSON format and includes at least the following key fields: anomaly type identifier (e.g., "baffle_missing"), anomaly location annotation data (e.g., normalized bounding box coordinates [x, y, w, h] or a set of key points), processing time limit requirements (e.g., Unix timestamp or relative duration), and a unique device identifier for the target processing terminal. This structured design allows the task information to be directly parsed and scheduled by downstream systems without relying on natural language understanding or manual transcription. For example, when the anomaly recognition model outputs a baffle missing result, the system automatically extracts the coordinates of the predicted bounding box with the highest confidence as the anomaly location annotation data, matches the corresponding processing time limit according to the preset operation and maintenance level strategy (e.g., 24 hours for emergency hazards, 72 hours for general hazards), and finally assembles it into a complete task data package. The process of linking task information to the target processing terminal essentially involves establishing a mapping relationship between task IDs and terminal device IDs in the database, and sending rectification work order reminders to inspection personnel through message push services (SMS, APP push), clarifying rectification requirements and the location information of hidden dangers based on AI identification and annotation, so as to ensure that inspection personnel can receive the information in a timely manner and accurately locate the rectification parts.

[0133] In one implementation, in response to receiving a task completion instruction from the target processing terminal, an image re-acquisition process is triggered to acquire verification image data.

[0134] Specifically, this is the key triggering node for achieving automated closed-loop verification. The system maintains a state machine throughout the task lifecycle, constantly monitoring state change events from the target processing terminal. The task completion instruction is a specific digital signal or API call request, not a verbal report or paper signature. When the state machine captures this instruction, it automatically verifies its validity (e.g., signature verification, time window check). Upon successful verification, the image re-acquisition process is immediately triggered. This process sends standardized acquisition configuration parameters to the target processing terminal, requiring it to re-capture images according to the six view specifications described in the above embodiment, thereby obtaining the verification image data. This event-driven triggering mechanism completely eliminates the arbitrariness of manual subjective initiation of verification, ensuring that the verification action is only executed after the system confirms the objective state of "processing completed," effectively preventing missed inspections or premature archiving due to human negligence.

[0135] As one implementation method, an anomaly detection model is used to identify anomalies in the verification image data to obtain the verification result.

[0136] Specifically, the reasoning logic in the review stage remains consistent with that in the initial inspection stage to ensure uniformity of verification standards. The system inputs the review image data into the same anomaly recognition model and outputs a review result including anomaly type, location, and confidence level. If the review result shows no anomaly or the anomaly confidence level is below a preset threshold, the review recognition result is determined to be "anomaly eliminated"; otherwise, it is determined to be "anomaly not eliminated." It is worth noting that during the review reasoning process, the system can also use the anomaly location annotation data recorded during the initial inspection as prior knowledge to focus on or perform ROI cropping reasoning on the corresponding areas in the review image. This not only improves review efficiency but also effectively eliminates the influence of newly added background interference, enhancing the targeting of the verification.

[0137] As one implementation method, if the verification and identification results indicate that the anomaly has not been eliminated, secondary task information containing the anomaly location annotation data is generated, and the task status is updated to pending processing until the verification and identification results indicate that the anomaly has been eliminated, at which point the task status is updated to closed loop.

[0138] Specifically, when the review fails, the system does not simply reject the task but automatically generates a new secondary task information. This secondary task information, in addition to inheriting the original task's exception type and time limit, also carries the original exception location annotation data and the specific reason code for this review failure. This design has significant technical implications: on the one hand, the original location data provides the personnel with a precise comparison benchmark, enabling them to quickly locate the specific points where previous rectification was inadequate, avoiding blind investigation; on the other hand, the system rolls the task status back from "processed" to "pending processing" and records a "review failure" event in the database, forming a complete audit trail. Only when the subsequent review identification result clearly indicates that the exception has been eliminated will the state machine transition the task status to the final state "closed loop." The entire state transition process (pending processing → processing → processing completed → reviewing → closed loop / rollback to pending processing) is automatically controlled by preset program logic, and each state change is accompanied by data consistency verification. This robust data state machine mechanism fundamentally differs from traditional manual verbal confirmation or subjective acceptance, ensuring the authenticity and integrity of operational data. It also provides a reliable data source for subsequent big data analysis and operational quality assessment. It should be understood that although this embodiment uses JSON format and a state machine as examples, other implementations can also use data formats such as XML and Protobuf, or employ technologies such as workflow engines and blockchain smart contracts to achieve the same data flow and state management functions, as long as they can achieve automated triggering and closed-loop verification based on data state.

[0139] Please see Figure 5To facilitate understanding, this application further explains and illustrates the optical distribution box anomaly handling method with some embodiments:

[0140] Step 1: System Initialization and AI Model Training and Deployment

[0141] 1.1 Dataset Construction and Preprocessing:

[0142] Images of optical distribution boxes were collected from different communities and environments, covering three types of potential hazard states: "normal state" and "missing baffle, illegal pigtail, and abnormal grounding wire".

[0143] The LabelImg tool is used to annotate the hazard areas in the image, generating annotation files in VOC or COCO format;

[0144] The labeled dataset is divided into training, validation, and test sets in a 7:2:1 ratio, and the training set is expanded using data augmentation sub-units.

[0145] 1.2 AI Model Structure Adjustment and Initialization:

[0146] Based on the YOLOv5s open-source architecture, a "hazard feature enhancement branch" is added after the C3 module in the Backbone layer. The branch uses 3×3 convolution and ReLU activation function to extract fine-grained hazard features.

[0147] In the feature fusion stage of the Neck layer, a cross-layer connection of "shallow features (C3 layer output) - deep features (SPP layer output)" is added to improve the ability to recognize small targets;

[0148] Initialize the model weights using pre-trained YOLOv5s weights as a base, and only randomly initialize the weights for newly added branches.

[0149] 1.3 AI Model Training and Loss Optimization:

[0150] Set the training parameters: batch size (batchsize=16), number of iterations (epochs=100), and initial learning rate (lr=0.001, using cosine annealing learning rate scheduling).

[0151] The model loss is calculated using FocalLoss, and the model accuracy is monitored in real time using the validation set (e.g., mAP@0.5). When the validation set accuracy does not improve for 10 consecutive epochs, training is stopped and the optimal model weights are saved.

[0152] 1.4 Model Inference Acceleration Optimization and Deployment:

[0153] TensorRT is used to optimize the optimal model and generate an .engine model file adapted for mobile devices.

[0154] The optimized model is deployed to the mobile terminals of inspection personnel (such as mobile apps), and image acquisition, uploading, and result display functions are integrated.

[0155] Step 2: Inspection and positioning of optical distribution boxes and generation of work orders

[0156] 2.1 Inspection personnel arrive at the optical distribution box site with a mobile terminal and log in to the "Home Customer Installation and Maintenance - Outdoor Support" module of the operation and maintenance platform through the terminal;

[0157] 2.2 Enter the community name and optical distribution box number. The optical distribution box information matching sub-unit will automatically retrieve the basic information of the target optical distribution box and confirm the inspection target.

[0158] 2.3 The inspection location recording subunit automatically collects the latitude and longitude coordinates of the site and associates them with the inspection task;

[0159] 2.4 The system generates electronic inspection work orders, which clearly require the collection of six types of standardized images (door closed A / B side, door open A / B side, grounding close-up, and overall view inside the box). Inspection personnel complete the image collection according to the instructions.

[0160] Step 3: Uploading Inspection Images and AI-Powered Hazard Identification

[0161] 3.1 The inspection personnel submit the collected images, and the mobile terminal encrypts and uploads the images to the system backend;

[0162] 3.2 The image preprocessing unit of the AI ​​hazard intelligent identification module performs standardization processing on the uploaded image (such as resizing to 640×640 pixels and normalizing pixel values).

[0163] 3.3 The model inference acceleration subunit loads the optimized AI model, performs inference on the preprocessed image, and outputs the hazard identification results, including hazard type, location coordinates, and confidence level.

[0164] 3.4 The recognition result filtering unit performs post-processing on the inference results: low-confidence prediction boxes are filtered using a confidence threshold (e.g., 0.25), duplicate prediction boxes are removed by non-maximum suppression (IOU threshold = 0.45), and the final recognition result is output.

[0165] Step 4: Hazard Work Order Issuance and On-site Rectification

[0166] 4.1 Work order generation and association sub-unit receiving final identification results:

[0167] If no hidden dangers are identified, the inspection work order will be automatically marked as "compliant" and synchronized to the data accumulation and visualization management module for archiving;

[0168] If a hazard is identified, a "hazard rectification work order" will be automatically generated. The work order includes the type of hazard (such as "missing baffle"), a map showing the location of the hazard, and the rectification deadline.

[0169] 4.2 The work order dispatch and reminder sub-unit sends rectification work order reminders to inspection personnel via SMS and APP push to ensure that inspection personnel receive them in a timely manner.

[0170] 4.3 After receiving the reminder, the inspection personnel will bring the necessary tools for rectification (such as spare baffles, cable ties, and grounding wire accessories) to the optical distribution box site and complete the targeted rectification (such as installing baffles, tidying up the pigtails, and reconnecting the grounding wire) according to the location of the hidden dangers marked in the work order.

[0171] Step 5: Verification of rectification results and archiving of work orders

[0172] 5.1 After the rectification is completed, the inspection personnel shall re-collect the six standardized images of the optical distribution box as required in step 2.4 and submit them to the system backend;

[0173] 5.2 The rectification result review subunit automatically calls the AI ​​hazard intelligent identification module to perform secondary recognition on the rectified images;

[0174] 5.3 If the secondary identification result is "no hidden danger", the system will mark the work order as "closed loop" and synchronously update the rectification record in the optical distribution box data asset sub-unit;

[0175] 5.4 If the secondary identification result is "the hidden danger still exists", the system will return the work order to the inspection personnel and attach a "the hidden danger has not been eliminated" annotation to prompt the inspection personnel to rectify the problem again until the hidden danger is eliminated.

[0176] Step 6: Data Accumulation and Visualized Management

[0177] The optical distribution box data asset sub-unit of the data accumulation and visualization management module automatically stores all data from this inspection, including images, recognition results, and work order records, forming the operation and maintenance file of the optical distribution box;

[0178] 6.2 The multi-dimensional report sub-unit calculates the inspection completion rate, number of hidden dangers found, and rectification completion rate of each grid according to a preset period (such as daily), generates reports, and pushes them to the management personnel's accounts;

[0179] The 6.3 GIS map visualization sub-unit links the location of the optical distribution box with its current status (normal / pending rectification / closed loop) to the GIS map. Managers can intuitively view the distribution of potential hazards through the map and formulate targeted operation and maintenance resource allocation plans.

[0180] In some embodiments of this application, taking the inspection and rectification of a missing baffle in a certain community's optical distribution box as an example, the method for handling optical distribution box anomalies is further explained and described, specifically including the following steps:

[0181] 1. System Initialization and Model Deployment

[0182] A total of 5,000 images of optical distribution boxes in this community and surrounding communities were collected, including 3,000 images in "normal condition", 1,000 images with "missing baffle", 600 images with "white pigtails", and 400 images with "abnormal grounding wire".

[0183] Images are labeled using LabelImg, generating COCO format labeled files, which are then divided into a training set (3500 images), a validation set (1000 images), and a test set (500 images) in a 7:2:1 ratio.

[0184] Data augmentation of the training set: random rotation (0°-90°), horizontal flip (probability 0.5), brightness adjustment (±15%), addition of Gaussian noise (variance 0.01);

[0185] Adjust the YOLOv5s model: Add a 3×3 convolutional branch (64 output channels) after the C3-128 layer in the Backbone, and add a cross-layer connection between C3-64 and SPP-512 in the FPN layer of the Neck.

[0186] The model was trained using FocalLoss (γ=2), with a batch size of 16 and 100 epochs. The final validation set mAP@0.5 reached 92%, and the optimal weights were saved.

[0187] The model was quantized to 16 bits using TensorRT, generating an .engine file adapted for mobile apps, which was then deployed to the Android mobile apps used by inspection personnel.

[0188] 2. On-site inspection and image acquisition

[0189] The inspection personnel arrived at the optical distribution box in the target community with their mobile phones, opened the APP, logged into the operation and maintenance platform, entered "XX Community" and "GJ001", and retrieved the basic information of the optical distribution box (model: 24-core optical distribution box, grid belonging to: XX Street).

[0190] The APP automatically collects the on-site latitude and longitude (31°XX′N, 120°XX′E) and generates an electronic inspection work order.

[0191] Following the work order instructions, the inspection personnel took photos in sequence: front view with door closed (A side), side view with door closed (B side), left side of the box with door open (A side inside), right side of the box with door open (B side inside), close-up view of grounding (grounding terminal), and overall view of the box (top view), for a total of 6 images.

[0192] 3. AI-powered hazard identification and work order generation

[0193] The inspection personnel submit images, and the APP encrypts and uploads the images to the system backend.

[0194] The background preprocesses the image: resizes it to 640×640 pixels and normalizes the pixel values ​​to [0,1];

[0195] Load the TensorRT optimized model and perform inference on the image (inference time on mobile device is 1.5 seconds). Output result: "Baffle missing" (location: lower edge of the door, confidence level 0.85).

[0196] The system filters out low-confidence results (confidence 0.85 > 0.25), removes duplicate bounding boxes, and generates the final recognition result;

[0197] Based on the identification results, the work order generation subunit generates a "Baffle missing rectification work order", links it to the inspection personnel's account, and pushes a reminder to the APP: "The optical distribution box GJ001 in XX community has a missing baffle, which needs to be rectified within 3 days".

[0198] 4. On-site rectification and result verification

[0199] After receiving the reminder, the inspection personnel arrived at the scene with spare baffles and screwdrivers, removed the remaining part of the old baffle on the optical distribution box door, installed the new baffle and fixed it in place;

[0200] After rectification, take 6 new photos as required by the work order (focusing on the installation area of ​​the baffle) and submit them to the backend.

[0201] The rectification result review subunit calls the AI ​​model to reason about the new image and outputs "no hidden dangers";

[0202] The system marks the work order as "closed loop" and updates the optical distribution box data assets: rectification record (date: 202X, installation of lower baffle), and archives the rectified images.

[0203] 5. Data visualization and control

[0204] The data asset subunit of the optical distribution box stores all the data from this inspection, forming the operation and maintenance file of the "GJ001 optical distribution box";

[0205] Daily statistics for multi-dimensional report sub-units: XX Street grid inspection completion rate 100%, number of hidden dangers discovered 1 (missing barrier), rectification completion rate 100%;

[0206] The GIS map visualization sub-unit marks "GJ001 Optical Distribution Box" in green (normal status) on the map, and the administrator can view the historical inspection and rectification records of the optical distribution box through the backend.

[0207] See Figure 8 , Figure 8 This is a structural diagram of an optical distribution box anomaly processing device 800 provided in an embodiment of this application. The optical distribution box anomaly processing device 800 includes:

[0208] The acquisition module 801 is used to acquire the initial image data collected from the optical distribution box;

[0209] The identification module 802 is used to identify anomalies in the initial image data through an anomaly identification model to obtain an anomaly identification result; wherein, the anomaly identification model includes a baffle missing identification unit, a white pigtail identification unit, and a grounding wire anomaly identification unit;

[0210] The generation module 803 is used to generate task information containing anomaly handling instructions in response to the anomaly identification result indicating the existence of an anomaly.

[0211] The verification module 804 is used to reacquire the verification image data collected for the optical distribution box when the task status indicated by the task information changes to processing completion, and to perform anomaly identification on the verification image data through the anomaly identification model to verify whether the anomaly indicated by the anomaly identification result has been eliminated.

[0212] Optionally, the recognition module 802 is specifically used to perform baffle missing recognition on the initial image data based on the key point distribution through the baffle missing recognition unit in the anomaly recognition model to obtain a first recognition result;

[0213] The recognition module 802 is specifically used to perform white fiber recognition on the initial image data based on color space threshold and morphological operations through the white fiber recognition unit in the anomaly recognition model to obtain a second recognition result.

[0214] The identification module 802 is specifically used to perform ground wire anomaly identification on the initial image data based on multi-dimensional feature fusion through the ground wire anomaly identification unit in the anomaly identification model to obtain a third identification result.

[0215] The anomaly identification results include the first identification result, the second identification result, and the third identification result.

[0216] Optionally, the identification module 802 includes:

[0217] The acquisition submodule is used to acquire the current set of key points of the optical distribution box through the baffle missing identification unit in the anomaly identification model. Based on a preset standard key point distribution model corresponding to the optical distribution box, the Euclidean distance error between the current set of key points and the standard key point distribution model is calculated. ;

[0218] The generation submodule is used to identify the missing baffle in the anomaly identification model based on the Euclidean distance error. The first identification result, wherein, when the Euclidean distance error When the error exceeds a preset key point error threshold, the first identification result indicates that the optical distribution box has a missing baffle; when the Euclidean distance error... When the error is less than or equal to the key point error threshold, the first identification result indicates that the optical distribution box does not have a missing baffle.

[0219] The running submodule is used to set the hue based on the HSV color space using the white pigtail identification unit in the anomaly identification model. saturation and brightness The threshold range is used to filter out pixels located within the threshold range as candidate regions, and morphological opening and closing operations are performed on the candidate regions to obtain the target texture region; the white pigtail recognition unit in the anomaly recognition model is used to perform white pigtail recognition on the target texture region to obtain the second recognition result;

[0220] The generation submodule is further configured to extract the color features of the optical distribution box through the grounding wire anomaly identification unit in the anomaly identification model. Morphological characteristics Spatial relationship characteristics The grounding wire anomaly identification unit in the anomaly identification model is used to identify anomalies based on a comprehensive evaluation function. Calculate the evaluation score; using the grounding wire anomaly identification unit in the anomaly identification model, generate the third identification result based on the evaluation score, wherein when the evaluation score is less than a preset anomaly threshold, the third identification result indicates that the optical distribution box has a grounding wire anomaly; when the evaluation score is greater than or equal to the anomaly threshold, the third identification result indicates that the optical distribution box does not have a grounding wire anomaly. , and These are the preset weighting coefficients.

[0221] Optionally, the identification module 802 includes:

[0222] The inference submodule is used to perform anomaly identification on the initial image data through the anomaly identification model and output the initial identification result;

[0223] The initial identification result is subjected to confidence threshold filtering and nonmaximum suppression processing to obtain the anomaly identification result.

[0224] Optionally, the device further includes:

[0225] A construction module is used to construct an optical distribution box anomaly dataset, which includes samples of optical distribution box images labeled with missing baffles, samples of optical distribution box images labeled with white pigtails, and samples of optical distribution box images labeled with grounding wire anomaly tags.

[0226] An enhancement module is used to perform data enhancement processing on the optical cross-box anomalous dataset to obtain a target dataset. The data enhancement processing includes at least one of image rotation, flipping, illumination intensity adjustment, and noise addition.

[0227] The training module is used to train the initial anomaly recognition model based on the target dataset using the Focal Loss loss function, so as to obtain the trained recognition model.

[0228] An optimization module is used to perform lightweight optimization processing on the trained recognition model to obtain the anomaly recognition model, wherein the lightweight optimization processing includes at least one of model quantization, layer fusion, and kernel auto-tuning.

[0229] Optionally, the generation module 803 includes:

[0230] The extraction submodule is used to extract the anomaly type, anomaly location annotation data and processing time limit requirements corresponding to the anomaly identification result, generate the task information, and associate the task information with the target processing terminal;

[0231] The verification module 804 includes:

[0232] The acquisition submodule is used to, in response to receiving the task completion instruction from the target processing terminal, trigger the image re-acquisition process and acquire the verification image data when the task status indicated by the task information changes to processing completion;

[0233] The identification submodule is used to perform anomaly identification on the verification image data through the anomaly identification model to obtain the verification identification result;

[0234] The generation submodule is used to generate secondary task information containing the abnormal location annotation data if the abnormality indicated by the review and identification result has not been eliminated, and update the task status to pending processing until the abnormality indicated by the review and identification result has been eliminated.

[0235] The optical distribution box anomaly handling device 800 of this application embodiment can implement all the steps of the above-described optical distribution box anomaly handling method and achieve the same beneficial effects. To avoid repetition, it will not be described again.

[0236] This application also provides an electronic device. Since the principle by which the electronic device solves the problem is similar to the optical distribution box anomaly handling method in this application, the implementation of this electronic device can refer to the implementation of the above-described optical distribution box anomaly handling method; repeated details will not be elaborated further. Figure 9 As shown, the electronic device according to an embodiment of this application includes: a processor 900, configured to read a program from a memory 920 and execute the following processes:

[0237] Acquire initial image data for the optical distribution box;

[0238] The initial image data is subjected to anomaly identification using an anomaly identification model to obtain anomaly identification results; wherein, the anomaly identification model includes a baffle missing identification unit, a white pigtail identification unit, and a grounding wire anomaly identification unit;

[0239] In response to the anomaly identification result indicating the presence of an anomaly, task information containing anomaly handling instructions is generated;

[0240] When the task status indicated by the task information changes to processing completed, the verification image data collected for the optical distribution box is reacquired, and the verification image data is subjected to anomaly identification through the anomaly identification model to verify whether the anomaly indicated by the anomaly identification result has been eliminated.

[0241] Among them, Figure 9 In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 900) and memory (memory 920). The bus architecture can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides the interface. Processor 900 is responsible for managing the bus architecture and general processing, and memory 920 can store data used by processor 900 during operation.

[0242] Optionally, the processor 900 is configured to read the program from the memory 920 and execute the following processes:

[0243] The baffle missing identification unit in the anomaly identification model performs baffle missing identification on the initial image data based on the key point distribution to obtain a first identification result.

[0244] The white pigtail recognition unit in the anomaly recognition model performs white pigtail recognition on the initial image data based on color space threshold and morphological operations to obtain a second recognition result.

[0245] The grounding wire anomaly identification unit in the anomaly identification model performs grounding wire anomaly identification on the initial image data based on multi-dimensional feature fusion to obtain a third identification result.

[0246] The anomaly identification results include the first identification result, the second identification result, and the third identification result.

[0247] Optionally, the processor 900 is configured to read the program from the memory 920 and execute the following processes:

[0248] The current set of key points of the optical distribution box is obtained through the baffle missing identification unit in the anomaly identification model. Based on a preset standard key point distribution model corresponding to the optical distribution box, the Euclidean distance error between the current set of key points and the standard key point distribution model is calculated. ;

[0249] Based on the Euclidean distance error, the baffle missing identification unit in the anomaly identification model... The first identification result, wherein, when the Euclidean distance error When the error exceeds a preset key point error threshold, the first identification result indicates that the optical distribution box has a missing baffle; when the Euclidean distance error... When the error is less than or equal to the key point error threshold, the first identification result indicates that the optical distribution box does not have a missing baffle.

[0250] The white pigtail recognition unit in the anomaly recognition model sets the hue based on the HSV color space. saturation and brightness The threshold range is used to filter out pixels located within the threshold range as candidate regions, and morphological opening and closing operations are performed on the candidate regions to obtain the target texture region; the white pigtail recognition unit in the anomaly recognition model is used to perform white pigtail recognition on the target texture region to obtain the second recognition result;

[0251] The color features of the optical distribution box are extracted using the grounding wire anomaly identification unit in the anomaly identification model. Morphological characteristics Spatial relationship characteristics The grounding wire anomaly identification unit in the anomaly identification model is used to identify anomalies based on a comprehensive evaluation function. Calculate the evaluation score; using the grounding wire anomaly identification unit in the anomaly identification model, generate the third identification result based on the evaluation score, wherein when the evaluation score is less than a preset anomaly threshold, the third identification result indicates that the optical distribution box has a grounding wire anomaly; when the evaluation score is greater than or equal to the anomaly threshold, the third identification result indicates that the optical distribution box does not have a grounding wire anomaly. , and These are the preset weighting coefficients.

[0252] Optionally, the processor 900 is configured to read the program from the memory 920 and execute the following processes:

[0253] The anomaly detection model is used to detect anomalies in the initial image data and output the initial detection result.

[0254] The initial identification result is subjected to confidence threshold filtering and nonmaximum suppression processing to obtain the anomaly identification result.

[0255] Optionally, the processor 900 is configured to read the program from the memory 920 and execute the following processes:

[0256] Construct an optical distribution box anomaly dataset, which includes samples of optical distribution box images labeled with missing baffles, samples of optical distribution box images labeled with white pigtails, and samples of optical distribution box images labeled with grounding wire anomaly tags.

[0257] Data augmentation processing is performed on the optical crossbox anomaly dataset to obtain the target dataset. The data augmentation processing includes at least one of image rotation, flipping, illumination intensity adjustment, and noise addition.

[0258] The initial anomaly detection model is trained using the Focal Loss loss function based on the target dataset to obtain the trained detection model.

[0259] The trained recognition model is subjected to lightweight optimization processing to obtain the anomaly recognition model, wherein the lightweight optimization processing includes at least one of model quantization, layer fusion and kernel auto-tuning.

[0260] Optionally, the processor 900 is configured to read the program from the memory 920 and execute the following processes:

[0261] Extract the anomaly type, anomaly location annotation data, and processing time limit requirements corresponding to the anomaly identification results, generate the task information, and associate the task information with the target processing terminal;

[0262] When the task status indicated by the task information changes to processing completed, in response to receiving the task completion instruction from the target processing terminal, the image re-acquisition process is triggered to acquire the review image data;

[0263] The anomaly detection model is used to perform anomaly detection on the verification image data to obtain the verification result;

[0264] If the verification and identification result indicates that the anomaly has not been eliminated, secondary task information containing the anomaly location annotation data is generated, and the task status is updated to pending processing until the verification and identification result indicates that the anomaly has been eliminated.

[0265] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, this computer program implements the various processes of the above-described optical distribution box anomaly handling method embodiments and achieves the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0266] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the above-described... Figure 1 The various processes of the optical distribution box anomaly handling method embodiment shown are all able to achieve the same technical effect, and will not be described again here to avoid repetition.

[0267] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0268] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0269] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for handling optical distribution box anomalies, characterized in that, include: Acquire initial image data for the optical distribution box; The initial image data is subjected to anomaly identification using an anomaly identification model to obtain anomaly identification results; wherein, the anomaly identification model includes a baffle missing identification unit, a white pigtail identification unit, and a grounding wire anomaly identification unit; In response to the anomaly identification result indicating the presence of an anomaly, task information containing anomaly handling instructions is generated; When the task status indicated by the task information changes to processing completed, the verification image data collected for the optical distribution box is reacquired, and the verification image data is subjected to anomaly identification through the anomaly identification model to verify whether the anomaly indicated by the anomaly identification result has been eliminated.

2. The method according to claim 1, characterized in that, The step of performing anomaly detection on the initial image data using an anomaly detection model to obtain anomaly detection results includes: The baffle missing identification unit in the anomaly identification model performs baffle missing identification on the initial image data based on the key point distribution to obtain a first identification result. The white pigtail recognition unit in the anomaly recognition model performs white pigtail recognition on the initial image data based on color space threshold and morphological operations to obtain a second recognition result. The grounding wire anomaly identification unit in the anomaly identification model performs grounding wire anomaly identification on the initial image data based on multi-dimensional feature fusion to obtain a third identification result. The anomaly identification results include the first identification result, the second identification result, and the third identification result.

3. The method according to claim 2, characterized in that, The step of using the baffle missing identification unit in the anomaly identification model to perform baffle missing identification on the initial image data based on the key point distribution to obtain a first identification result includes: The current set of key points of the optical distribution box is obtained through the baffle missing identification unit in the anomaly identification model. Based on a preset standard key point distribution model corresponding to the optical distribution box, the Euclidean distance error between the current set of key points and the standard key point distribution model is calculated. ; Based on the Euclidean distance error, the baffle missing identification unit in the anomaly identification model... The first identification result, wherein, when the Euclidean distance error When the error exceeds a preset key point error threshold, the first identification result indicates that the optical distribution box has a missing baffle; when the Euclidean distance error... When the error is less than or equal to the key point error threshold, the first identification result indicates that the optical distribution box does not have a missing baffle. The white pigtail recognition unit in the anomaly recognition model performs white pigtail recognition on the initial image data based on color space thresholds and morphological operations to obtain a second recognition result, including: The white pigtail recognition unit in the anomaly recognition model sets the hue based on the HSV color space. saturation and brightness The threshold range is used to filter out pixels located within the threshold range as candidate regions, and morphological opening and closing operations are performed on the candidate regions to obtain the target texture region; the white pigtail recognition unit in the anomaly recognition model is used to perform white pigtail recognition on the target texture region to obtain the second recognition result; The grounding wire anomaly identification unit in the anomaly identification model performs grounding wire anomaly identification on the initial image data based on multi-dimensional feature fusion to obtain a third identification result, including: The color features of the optical distribution box are extracted using the grounding wire anomaly identification unit in the anomaly identification model. Morphological characteristics Spatial relationship characteristics The grounding wire anomaly identification unit in the anomaly identification model is used to identify anomalies based on a comprehensive evaluation function. Calculate the evaluation score; using the grounding wire anomaly identification unit in the anomaly identification model, generate the third identification result based on the evaluation score, wherein when the evaluation score is less than a preset anomaly threshold, the third identification result indicates that the optical distribution box has a grounding wire anomaly; when the evaluation score is greater than or equal to the anomaly threshold, the third identification result indicates that the optical distribution box does not have a grounding wire anomaly. , and These are the preset weighting coefficients.

4. The method according to claim 1, characterized in that, The step of performing anomaly detection on the initial image data using an anomaly detection model to obtain anomaly detection results includes: The anomaly detection model is used to detect anomalies in the initial image data and output the initial detection result. The initial identification result is subjected to confidence threshold filtering and nonmaximum suppression processing to obtain the anomaly identification result.

5. The method according to claim 4, characterized in that, The process of acquiring the anomaly detection model includes: Construct an optical distribution box anomaly dataset, which includes samples of optical distribution box images labeled with missing baffles, samples of optical distribution box images labeled with white pigtails, and samples of optical distribution box images labeled with grounding wire anomaly tags. Data augmentation processing is performed on the optical crossbox anomaly dataset to obtain the target dataset. The data augmentation processing includes at least one of image rotation, flipping, illumination intensity adjustment, and noise addition. The initial anomaly detection model is trained using the Focal Loss loss function based on the target dataset to obtain the trained detection model. The trained recognition model is subjected to lightweight optimization processing to obtain the anomaly recognition model, wherein the lightweight optimization processing includes at least one of model quantization, layer fusion and kernel auto-tuning.

6. The method according to claim 1, characterized in that, The generation of task information containing exception handling instructions includes: Extract the anomaly type, anomaly location annotation data, and processing time limit requirements corresponding to the anomaly identification results, generate the task information, and associate the task information with the target processing terminal; When the task status indicated by the task information changes to processing completion, the verification image data acquired for the optical distribution box is reacquired, and the verification image data is subjected to anomaly identification using the anomaly identification model to verify whether the anomaly indicated by the anomaly identification result has been eliminated, including: When the task status indicated by the task information changes to processing completed, in response to receiving the task completion instruction from the target processing terminal, the image re-acquisition process is triggered to acquire the review image data; The anomaly detection model is used to perform anomaly detection on the verification image data to obtain the verification result; If the verification and identification result indicates that the anomaly has not been eliminated, secondary task information containing the anomaly location annotation data is generated, and the task status is updated to pending processing until the verification and identification result indicates that the anomaly has been eliminated.

7. An optical distribution box anomaly handling device, characterized in that, include: The acquisition module is used to acquire the initial image data collected from the optical distribution box; The identification module is used to identify anomalies in the initial image data through an anomaly identification model to obtain an anomaly identification result; wherein, the anomaly identification model includes a baffle missing identification unit, a white pigtail identification unit, and a grounding wire anomaly identification unit; The generation module is used to generate task information containing anomaly handling instructions in response to the anomaly identification result indicating the existence of an anomaly; The verification module is used to reacquire the verification image data collected for the optical distribution box when the task status indicated by the task information changes to processing completion, and to perform anomaly identification on the verification image data through the anomaly identification model to verify whether the anomaly indicated by the anomaly identification result has been eliminated.

8. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the optical distribution box anomaly handling method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the optical distribution box anomaly handling method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, It includes computer instructions, which, when executed by a processor, implement the steps of the optical distribution box anomaly handling method as described in any one of claims 1 to 6.

Images