Improved target detection and edge intelligent algorithm deployment method, system and equipment for remote monitoring of offshore photovoltaic power station and medium

Abstract

The application discloses an improved target detection and edge intelligent algorithm deployment method, system and device for remote monitoring of a marine photovoltaic power station and a medium, aiming at the problems of easy missed detection of small targets and complex background, an improved model based on YOLO11n is constructed, a P2 detection branch is newly added, a lightweight feature enhancement module and a shallow detail enhancement module are set, and data enhancement training facing a marine scene is adopted; an edge gateway collects video streams and operating data such as voltage, current and temperature in real time, and a sensing data preprocessing algorithm and an improved target detection method are deployed on the edge gateway, collected data is fused and verified, and after an abnormal event is confirmed, alarm information is sent to a cloud management platform and a monitoring terminal. The application improves detection accuracy, real-time performance and robustness while ensuring lightweight deployment and reducing bandwidth occupation.

Description

Technical Field

[0001] This invention relates to intelligent operation and maintenance and remote monitoring technology for offshore photovoltaic power plants, specifically to an improved target detection and edge intelligence algorithm deployment method, system, equipment and medium for remote monitoring of offshore photovoltaic power plants. Background Technology

[0002] As the scale of offshore photovoltaic (PV) power plant construction continues to expand, the operating environment of these plants exhibits characteristics such as high salt spray, strong winds and waves, high humidity, high reflectivity, drastic changes in sunlight, and unstable network links. Compared to terrestrial PV systems, offshore PV equipment is more susceptible to disturbances from complex environments, leading to minor anomalies such as surface contamination of modules, loosening of structural components, and abnormal operating parameters. Failure to detect and warn of these issues in a timely manner can result in anything from reduced power generation efficiency to serious consequences such as equipment damage, increased maintenance costs, and even safety risks.

[0003] Existing remote monitoring solutions for offshore photovoltaic systems have the following main shortcomings:

[0004] (i) Most existing computer vision detection algorithms are designed for general scenarios and are not well adapted to situations such as small target anomalies, complex backgrounds, high brightness reflections, low contrast and partial occlusion in marine photovoltaic scenarios, which easily leads to missed detections and false detections.

[0005] (ii) Existing monitoring systems mostly adopt a centralized cloud processing mode, which requires the continuous uploading of large-scale video streams and operational data to the cloud, resulting in communication bandwidth consumption and large latency, and there is a high dependence on network stability.

[0006] (iii) Existing video detection results and sensor-collected data are usually processed separately, lacking a unified data fusion and verification mechanism, resulting in low accuracy of abnormal event detection and difficulty in effectively distinguishing between real anomalies and occasional noise disturbances.

[0007] (iv) Existing edge gateway deployment solutions are mostly general-purpose solutions, lacking dedicated model optimization, hardware adaptation, quantization compression and real-time processing design for complex offshore photovoltaic conditions, making it difficult to balance accuracy, real-time performance and engineering deployment feasibility. Summary of the Invention

[0008] The purpose of this invention is to propose an improved target detection and edge intelligence algorithm deployment method, system, equipment and medium for remote monitoring of offshore photovoltaic power plants, so as to achieve low latency, high accuracy and low bandwidth occupancy monitoring of abnormal events.

[0009] The technical solution to achieve the purpose of this invention is as follows: an improved target detection method for remote monitoring of offshore photovoltaic power plants. This method employs an improved target detection model based on YOLOv11n to identify and locate abnormal targets in offshore photovoltaic images. The improved target detection model includes a newly added P2 small target detection branch, a lightweight feature enhancement module, a shallow detail enhancement module, and a data augmentation training mechanism tailored to offshore scenarios. This addresses the problems of missed and false detections caused by small target scale, complex background, and reflective interference in offshore photovoltaic scenarios. The method includes the following steps:

[0010] S1. Collect visible light images of the marine photovoltaic array, label abnormal targets in the images, and construct training, validation and test sets;

[0011] S2. Construct an improved target detection model, specifically including: based on the original YOLO11n three-scale detection head with P3, P4, and P5 as detection outputs, introduce the P2 shallow high-resolution feature from the backbone network, and further upsample the generated P3 feature map to make its spatial resolution comparable to the P2 feature output from the backbone network. Figure 1 Then, the upsampled P3 feature map is spliced ​​and fused with the P2 feature map to add a new P2 small target detection output branch; a lightweight feature enhancement module is set in the feature fusion path of the detection head; and a shallow detail enhancement module is set at the P2 feature output end or the fusion end after splicing P2 and P3 features.

[0012] S3. During model training, the images loaded from the training set are subjected to real-time data augmentation processing for marine photovoltaic scenarios to generate augmented samples.

[0013] S4. Train the improved target detection model with enhanced sample input, and update the network parameters through joint optimization of bounding box regression loss, classification loss and distribution focus loss;

[0014] S5. Evaluate the performance of the trained model based on the validation set, and fine-tune the training hyperparameters or model weights according to the evaluation results.

[0015] S6. Input the image of the marine photovoltaic system to be detected into the improved target detection model after training, and output the location, category and confidence level of the abnormal target.

[0016] Furthermore, the newly added P2 small target detection branch is constructed as follows: based on the original upsampling path from high-level features to low-level features in the detection head, the P3 feature map is further upsampled so that its spatial resolution is similar to that of the P2 features output by the backbone network. Figure 1Then, the upsampled P3 feature map is concatenated and fused with the P2 feature map to obtain a high-resolution fused feature map. The high-resolution fused feature map is input into the new detection layer to form a joint output structure of four detection scales: P2, P3, P4, and P5. The backbone network retains the original four-level feature output of YOLO11n, which are downsampled by 4 times, 8 times, 16 times, and 32 times relative to the input image, respectively, and are denoted as P2, P3, P4, and P5.

[0017] Furthermore, the lightweight feature enhancement module is set in the feature fusion path of the detection head to replace the original upsampling operation, so as to restore edge information and local texture information during the feature map resolution improvement process; the lightweight feature enhancement module includes one or more of pointwise convolution, depthwise separable convolution, and learnable upsampling units; and a weighted fusion mechanism is introduced in the multi-scale feature fusion process to set learnable fusion weights for features at different levels, so as to adaptively allocate the contribution ratio of shallow texture features and deep semantic features.

[0018] Furthermore, the shallow detail enhancement module is located at the output end of the P2 feature or at the fusion end after concatenating the P2 and P3 features, and includes a channel compression unit, a local texture extraction unit, an attention recalibration unit, and a residual connection unit. The channel compression unit uses 1×1 convolution to reduce the number of shallow feature channels; the local texture extraction unit uses 3×3 depthwise convolution or ordinary convolution to extract the edges and fine texture responses of abnormal regions; the attention recalibration unit uses lightweight channel attention to enhance the features of abnormally sensitive regions; and the residual connection unit is used to maintain the stable transmission of the original shallow features. The lightweight channel attention includes global average pooling, one-dimensional or two-dimensional convolution transformation, and a sigmoid activation function.

[0019] Furthermore, the data enhancement processing for marine photovoltaic scenarios includes one or more of the following: color perturbation, brightness perturbation, translation transformation, scale scaling, perspective transformation, image stitching enhancement, sea surface high-brightness noise simulation, and local reflection spot simulation;

[0020] During the training process, the ability to locate small targets is improved by adjusting the weight ratio of bounding box regression loss, classification loss and distribution focus loss. When the number of abnormal class samples is lower than that of normal class, class weights and / or hard example reinforcement mechanisms are applied to the abnormal class.

[0021] The performance evaluation metrics include one or more of the following: small target average precision, overall average precision, recall rate, and inference speed.

[0022] The abnormal targets include one or more of the following: surface contamination, foreign object obstruction, cracks, component damage, abnormal connection parts, and abnormal support structure;

[0023] The backbone network of the improved object detection model maintains the same lightweight scaling configuration as YOLO11n.

[0024] A method for deploying an edge intelligent algorithm for remote monitoring of offshore photovoltaic power plants includes the following steps:

[0025] A1. Deploy the improved target detection method and sensor data processing algorithm in the edge gateway;

[0026] A2. Real-time acquisition of video streams and operational data of offshore photovoltaic power stations via edge gateways, wherein the operational data includes one or more of voltage, current, and temperature data;

[0027] A3. The video anomaly results are output by improving the target detection model, and the running data anomaly signal is output by the sensor data processing algorithm; the sensor data processing algorithm is used to perform filtering, noise reduction, time alignment, threshold judgment and standardization processing;

[0028] A4. Perform fusion verification on abnormal video results and abnormal operation data signals. The fusion verification includes time correlation and logical cross verification.

[0029] A5. After confirming an abnormal event, store relevant data locally and send alarm information to the cloud platform. If the abnormal event is not confirmed, only transmit key status data or statistical summaries to the cloud.

[0030] Furthermore, the specific rules for the fusion verification are as follows: when video anomalies and operational data anomalies are detected simultaneously within a preset time window, the event is determined to be a high-confidence anomaly and an alarm is triggered; when only video anomalies are detected and no operational data anomalies are detected, the event is marked as an event to be reviewed; when only operational data anomalies are detected and no video anomalies are detected, secondary sampling confirmation is performed or the event is marked as an event to be reviewed.

[0031] A remote monitoring system for offshore photovoltaic power plants includes an edge gateway, a cloud management platform, and a monitoring terminal;

[0032] The edge gateway includes a core processing unit, a data acquisition interface, a communication module, a storage module, and an environmental protection module. It is used to acquire video streams and running data, and deploy the edge intelligent algorithm deployment method to complete real-time analysis, fusion verification, local storage, and preliminary anomaly judgment.

[0033] The cloud management platform is used to receive abnormal event information and key status data uploaded by the edge gateway, and to complete abnormal statistical analysis, historical record tracing, report generation, policy configuration, model version management and update instruction issuance.

[0034] The monitoring terminal is used to display on-site video screenshots, equipment status information, electrical parameter curves, and historical records of abnormal events, and to issue alarm notifications.

[0035] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the improved target detection method or the edge intelligence algorithm deployment method.

[0036] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the improved target detection method or the edge intelligence algorithm deployment method.

[0037] Compared with the prior art, the significant advantages of this invention are:

[0038] 1) By adding a P2 small target detection branch to the original three-scale detection structure, and combining lightweight feature enhancement, weighted multi-scale fusion and shallow detail enhancement modules, the detection capability of small targets, low contrast and complex background abnormal targets in marine photovoltaic scenarios has been improved, and the false negative rate has been reduced.

[0039] 2) To address issues such as sea surface reflection, local overexposure, shadow occlusion, low contrast, and wind and wave background disturbances in marine photovoltaic images, data augmentation and training optimization strategies tailored to marine scenarios were introduced to improve the robustness and generalization ability of the model in complex environments.

[0040] 3) Integrating video anomaly detection and sensor data processing on the edge gateway side, the accuracy of anomaly event confirmation is improved through multi-source data fusion verification, and false alarms and false negatives caused by a single data source are reduced.

[0041] 4) By adopting an edge deployment approach, video streams and sensor data are processed locally in real time. Only abnormal event information and key status data are uploaded to the cloud, which significantly reduces communication bandwidth consumption and cloud computing power pressure, and improves system real-time performance and network adaptability.

[0042] 5) The edge gateway is designed to adapt to marine environments such as salt spray resistance, waterproofing, and vibration resistance, and supports remote model updates and parameter configuration to meet the needs of long-term continuous operation and remote intelligent operation and maintenance of offshore photovoltaic power plants. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall process of the edge intelligent algorithm deployment method for remote monitoring of offshore photovoltaic power stations according to the present invention.

[0044] Figure 2 This is a schematic diagram of the improved target detection model structure built based on YOLO11n according to the present invention.

[0045] Figure 3 This is a schematic diagram of the hardware structure of the edge gateway of the present invention.

[0046] Figure 4 This is a schematic diagram of the composition of the remote monitoring system for offshore photovoltaic power plants of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] A method and system for deploying edge intelligent algorithms for remote monitoring of offshore photovoltaic power plants that integrates edge computing, computer vision anomaly detection, sensor data processing, and multi-source information fusion verification. Figure 1 This is a schematic diagram illustrating the overall process of the edge intelligent algorithm deployment method for remote monitoring of offshore photovoltaic power plants according to the present invention. Figure 2 This is a schematic diagram of the improved target detection model structure based on YOLO11n of this invention. Figure 3 This is a schematic diagram of the hardware composition structure of the edge gateway of the present invention. Figure 4 This is a schematic diagram of the composition of the remote monitoring system for offshore photovoltaic power plants of the present invention.

[0049] This invention addresses the challenges of small target scale, complex background, strong sea surface reflection, drastic lighting changes, and indistinct texture differences between abnormal and normal areas in marine photovoltaic scenarios, which can easily lead to missed detections and false detections. Based on the lightweight YOLO11n model, this invention collaboratively improves its detection scale, feature fusion method, shallow detail preservation capability, and adaptive training mechanism for marine scenarios, constructing a single-modal RGB visual detection network suitable for detecting abnormal events in marine photovoltaic systems.

[0050] like Figure 2 As shown, the improved target detection model is based on YOLO11n. The backbone network of the original YOLO11n includes convolutional downsampling layers, a C3k2 feature extraction module, a SPPF spatial pyramid pooling fast module, and a C2PSA attention module; its detection head adopts a top-down and bottom-up feature fusion approach, outputting three detection scales, namely P3 / 8, P4 / 16, and P5 / 32, corresponding to small, medium, and large-scale target detection. Since the original network does not directly output a high-resolution P2 / 4 detection layer, it suffers from insufficient utilization of shallow details for small target anomalies such as cracks, foreign object occlusion, and surface contamination in marine photovoltaic images.

[0051] To address the aforementioned issues, this invention adds a P2 small target detection branch to the original YOLO11n detection head. The four-level feature outputs of the original backbone network (P2 / 4, P3 / 8, P4 / 16, and P5 / 32) are retained. After the original detection head's upsampling path from P5 to P4 and then from P4 to P3, the P3 features are further upsampled and concatenated with the shallower P2 features in the backbone network to form a high-resolution small target enhancement feature map. This high-resolution feature map is then used as the input to the new detection layer, thereby expanding the original three-scale detection to four-scale detection, i.e., from Detect(P3,P4,P5) to Detect(P2,P3,P4,P5). By introducing the P2 detection branch, the model maintains its original ability to detect medium and large targets while enhancing its ability to identify distant, small-sized, low-contrast anomalies.

[0052] This invention also lightweightly enhances the feature fusion path in the detection head to reduce detail loss and aliasing errors during upsampling. The simple nearest-neighbor upsampling operation in the original detection head is replaced with a lightweight feature enhancement module. This lightweight feature enhancement module preferably includes one or more combinations of pointwise convolution, depthwise separable convolution, and learnable upsampling units, used to recover edge and local texture information while amplifying the feature map resolution. Furthermore, a weighted fusion mechanism is introduced during multi-scale feature fusion, setting learnable fusion weights for features at different levels, so that the model adaptively allocates the contribution ratio of high-resolution shallow texture features and deep semantic features during training. The weighted fusion can be expressed as:

[0053]

[0054]

[0055] in, Input feature maps at different scales. For size alignment and channel transformation operations, These are the learnable weight parameters corresponding to the scale features. Here, represents the normalized fusion weights, and n represents the number of feature scales involved in the fusion. This method can improve the fusion quality of anomalous features at different scales without significantly increasing model complexity.

[0056] This invention incorporates a detail enhancement module within a shallow, high-resolution feature path to improve edge representation capabilities in marine photovoltaic anomaly regions. Preferably, the detail enhancement module is positioned at the output of the P2 feature or at the fusion point after concatenating the P2 and P3 features, and includes a channel compression unit, a local texture extraction unit, an attention recalibration unit, and a residual connection unit. Specifically, firstly, shallow features are channel-compressed using a 1×1 convolution to reduce computational load; then, local edges and fine texture responses are extracted using a 3×3 depthwise convolution or ordinary convolution; subsequently, anomaly-sensitive regions are enhanced using lightweight channel attention or spatial attention; finally, the enhanced result is superimposed on the original input via residuals to obtain shallow enhanced features that combine stability and detail representation capabilities. The lightweight channel attention can be expressed as:

[0057]

[0058]

[0059] Where F represents the input feature, GAP represents global average pooling, and Conv represents a one-dimensional or two-dimensional convolution transformation. It is the Sigmoid activation function. This represents element-wise multiplication. By introducing detail enhancements at shallow layers, rather than stacking complex attention modules across the entire network, both lightweight requirements and the need for anomalous texture representation can be met.

[0060] During the model training phase, this invention employs a data augmentation strategy adapted to the marine photovoltaic scenario to enhance the model's robustness to complex environmental changes. Marine photovoltaic images are easily affected by factors such as specular reflection from the sea surface, shadow occlusion, low contrast, and background disturbances from wind and waves. Therefore, during training, color perturbation, brightness perturbation, scaling, translation, perspective changes, and mosaic enhancement are applied to the input image. Furthermore, degradation enhancements simulating high-brightness noise on the sea surface and local reflective spots can be added to improve the model's adaptability to complex visual conditions at sea.

[0061] Regarding loss function optimization, this invention still employs a collaborative weighted mechanism of bounding box regression loss, classification loss, and distribution focus loss in target detection. Based on the uneven distribution of anomaly categories in marine photovoltaic systems and the low proportion of a few anomaly samples, the weights of different loss terms are adjusted to enhance the model's learning ability for rare anomaly categories. Preferably, the proportion of bounding box regression loss and distribution focus loss in the overall loss is increased to improve the localization accuracy of small targets. When the class imbalance is significant, class weights or hard-case reinforcement mechanisms are applied to anomaly categories to alleviate the classification bias caused when the number of normal samples is much greater than the number of anomaly samples. The improved YOLO11n model can maintain a lightweight design during deployment; that is, the backbone network still uses a nano-scale scaling configuration, and performance enhancement is achieved only by adding a high-resolution detection branch and a few lightweight modules, thus balancing edge deployment requirements with detection accuracy requirements.

[0062] Based on the improved target detection model described above, this invention provides an improved target detection method for remote monitoring of offshore photovoltaic power plants, comprising the following steps:

[0063] Step S1: Acquire visible light images of the offshore photovoltaic array, label abnormal targets in the images, and construct training, validation, and test sets. The abnormal targets include one or more of the following: surface contamination, foreign object obstruction, cracks, component damage, abnormal connection points, and abnormal support structure.

[0064] Step S2: Construct an improved object detection model. This includes adding a P2 small object detection branch, and setting up a lightweight feature enhancement module and a shallow detail enhancement module.

[0065] Step S3: Perform data augmentation processing on the images in the training set for the marine photovoltaic scenario to generate enhanced samples. The data augmentation includes one or more of the following: color perturbation, brightness perturbation, contrast perturbation, translation transformation, scale scaling, perspective transformation, image stitching enhancement, blending enhancement, local highlight noise simulation, sea surface reflection simulation, shadow occlusion simulation, low-contrast degradation simulation, and background ripple perturbation simulation. These data augmentation operations can be implemented based on the interfaces provided by existing training frameworks.

[0066] Step S4: Train the improved target detection model using enhanced sample inputs, updating the network parameters through joint optimization of bounding box regression loss, classification loss, and distribution focus loss. During training, adjust the weight ratios of each loss to improve the ability to locate small targets; when the number of abnormal class samples is lower than that of normal class samples, apply class weights and / or hard example reinforcement mechanisms to the abnormal class.

[0067] Step S5: Evaluate the performance of the trained model based on the validation set, and adjust the network structure parameters and training parameters according to the evaluation results. Performance evaluation metrics include one or more of the following: small target mean precision, overall mean precision, recall, and inference speed.

[0068] Step S6: Input the image of the marine photovoltaic system to be detected into the improved target detection model after training, and output the location, category and confidence of the abnormal target to realize the identification and localization of abnormal events of marine photovoltaic systems.

[0069] After the model training is completed and it passes evaluation, it can be converted to a format suitable for edge inference engines (such as ONNX), and optionally subjected to INT8 or FP16 quantization compression to reduce model size and inference latency. The converted model is then used for subsequent edge gateway deployment.

[0070] This invention also provides a method for deploying an edge intelligent algorithm for remote monitoring of offshore photovoltaic power plants, including the following steps (see...). Figure 1 ):

[0071] Step A1: Deploy the improved target detection model described above in the edge gateway, and deploy the sensor data processing algorithm.

[0072] Step A2: Collect video streams and operational data of the offshore photovoltaic power station in real time via an edge gateway. The operational data includes one or more of the following: voltage, current, and temperature data.

[0073] Step A3: Output video anomaly results by improving the target detection model, and output runtime data anomaly signals by using a sensor data processing algorithm. The sensor data processing algorithm is used to perform filtering, denoising, time alignment, thresholding, and standardization.

[0074] Step A4: Perform fusion verification on the abnormal video results and abnormal operation data signals. The fusion verification includes time correlation and logical cross-verification.

[0075] Step A5: After confirming the abnormal event, store the relevant data locally and send alarm information to the cloud platform; if the abnormal event is not confirmed, only transmit key status data or statistical summaries to the cloud.

[0076] Furthermore, the specific rules for the fusion verification are as follows: when video anomalies and operational data anomalies are detected simultaneously within a preset time window, the event is determined to be a high-confidence anomaly and an alarm is triggered; when only video anomalies are detected and no operational data anomalies are detected, the event is marked as an event to be reviewed; when only operational data anomalies are detected and no video anomalies are detected, secondary sampling confirmation is performed or the event is marked as an event to be reviewed.

[0077] like Figure 3 As shown, the edge gateway is used to complete video stream access, sensor data acquisition, model inference, data fusion verification, local storage, and cloud communication at the offshore photovoltaic site. Its hardware modules include at least a core processing unit, a data acquisition interface, a communication module, a storage module, and an environmental protection module. The core processing unit is used to execute improved YOLO11n model inference tasks, sensor data processing tasks, and fusion verification tasks. It can be a low-power processor or system-on-a-chip with edge AI inference capabilities; in one embodiment, an edge computing platform using a Rockchip RK3568 or other equivalent AI acceleration units can be used. The data acquisition interface is used to access video and sensor data from the offshore photovoltaic power station site. The video interface is used to connect image acquisition devices such as network cameras and industrial cameras, while analog and digital interfaces are used to access data from different sensors. In one embodiment, the sensor data communication protocol can be one or more of RS485, Modbus, and Ethernet protocols. The communication module is used to realize data communication and remote management between the edge gateway and the cloud management platform, and can include one or more of 5G / 4G cellular communication units, WiFi communication units, and Ethernet communication units. The storage module is used to locally cache video clips, historical operating data, abnormal event records, and model files at the edge, and can use SD cards, SSDs, or other non-volatile storage media. The environmental protection module is used to improve the reliability of the edge gateway in marine environments, and preferably uses protective materials with salt spray resistance, water resistance, moisture resistance, vibration resistance, corrosion resistance, and wide temperature range operation capabilities.

[0078] like Figure 4 As shown, this invention also provides a remote monitoring system for offshore photovoltaic power plants, including an edge gateway, a cloud management platform, and a monitoring terminal. The edge gateway is used to collect video streams and operational data, and deploy the aforementioned edge intelligent algorithm deployment method to complete real-time analysis, fusion verification, local storage, and preliminary anomaly judgment. The cloud management platform is used to receive abnormal event information and key status data uploaded by the edge gateway, and to complete anomaly statistical analysis, historical record tracing, report generation, strategy configuration, model version management, and update command issuance. The monitoring terminal is used to display on-site video screenshots, equipment status information, electrical parameter curves, and historical records of abnormal events, and to issue alarm notifications.

[0079] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the improved target detection method or the edge intelligent algorithm deployment method.

[0080] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the improved target detection method or the edge intelligence algorithm deployment method.

[0081] In summary, this invention, by deploying an improved YOLO11n target detection model and sensor data processing algorithm at the edge, achieves integrated real-time processing and fusion verification of video streams and multiple types of operational data, thereby improving the real-time performance, accuracy, and reliability of detecting abnormal events in marine photovoltaic systems.

[0082] To verify the effectiveness of the present invention, the following experimental design was conducted.

[0083] Example 1: Construction of anomaly dataset for marine photovoltaic systems

[0084] High-definition network cameras with salt spray resistance are used to collect images of the offshore photovoltaic power station. The camera resolution is 4K or other suitable resolutions, and the frame rate can be set according to the on-site bandwidth and monitoring requirements. Simultaneously, multiple sensors, such as voltage, current, and temperature sensors, are configured to collect operational data synchronously. The data collection scenarios cover sunny, cloudy, foggy, windy and wavey weather, high-reflectivity conditions, and different time periods and shooting angles to improve the dataset's coverage of actual operating conditions. Target annotation is performed on video frames and images, with anomaly categories including one or more of the following: component surface contamination, foreign object obstruction, cracks, component damage, connection anomalies, and support structure anomalies, with particular attention paid to small target anomalies. During the data augmentation stage, enhancement operations such as brightness changes, contrast changes, blur perturbation, rotation and flipping, perspective transformation, small target cropping and magnification, local highlight noise simulation, shadow occlusion simulation, and background perturbation simulation are performed on the original samples to construct a training dataset adapted to complex offshore scenarios.

[0085] Example 2: Improved YOLO11n Model Training

[0086] The improved YOLO11n model was trained using a PyTorch-based object detection training framework. The training environment was based on a 64-bit Windows 11 operating system, with the following software: Ultralytics 8.4.21, Python 3.10.19, CUDA 12.8, CUDNN 9.18, and PyTorch 2.11.0. The computer used in the experiment had an Intel(R) Core(TM) Ultra7 265KF (3.90GHz) CPU, an NVIDIA GeForce RTX 5070 Ti GPU, and 16GB of RAM. During training, the initial learning rate and learning rate decay strategy could be preset, and end-to-end training was performed on the network after adding the P2 small object detection branch, the lightweight upsampling fusion module, and the shallow detail enhancement module. During training, bounding box regression loss, classification loss, and distribution focus loss are jointly optimized. The weights for bounding box regression loss can be set to 5.0–10.0, classification loss weights to 0.3–2.0, and distribution focus loss weights to 1.0–3.0. Class weights or hard-example enhancement mechanisms are set according to the sample quantity distribution of different anomaly categories. In the model validation phase, precision, recall, average precision, and inference speed are comprehensively evaluated. Based on the validation results, feature fusion weights, detail enhancement module parameters, and training hyperparameters are adjusted. In a specific example, the input image size can be set to 640×640, the batch size to 8, and the number of training epochs to 200. After training, the improved model achieves higher small target detection accuracy compared to the basic YOLO11n model on the marine photovoltaic small target anomaly detection task, while maintaining a lightweight model size suitable for edge deployment.

[0087] Example 3: Edge Gateway Deployment

[0088] A processor platform with edge AI inference capabilities (such as Rockchip RK3568 or other equivalent AI acceleration units) is used, running the Ubuntu embedded operating system. The trained improved YOLO11n model is deployed to the edge inference engine after model conversion (e.g., conversion to ONNX format), and can be further quantized and compressed to reduce model size and inference latency. The edge gateway accesses camera video streams via a video interface and sensor data via RS485. Video anomaly detection is performed by the improved YOLO11n model, while sensor data preprocessing is performed by the edge-side data processing program. The outputs of both are jointly judged by the fusion verification module. If an anomaly is confirmed, alarms are triggered in local storage and the cloud; if no anomaly is confirmed, only a key status summary is uploaded.

[0089] The real-time processing flow of the edge gateway is as follows: Step T1, the edge gateway acquires video streams from cameras in real time through the video interface and collects operational data such as voltage, current, and temperature in real time through the sensor interface; Step T2, the video stream is decoded and frames are extracted, and the improved YOLO11n model is input to output the abnormal target category, location, and confidence level; Step T3, the real-time acquired operational data is filtered, denoised, and thresholded to obtain the corresponding data anomaly signal; Step T4, the data fusion module performs time correlation and logical cross-validation on the video anomaly result and the data anomaly signal to confirm the authenticity of the anomaly event; Step T5, when an anomaly event is confirmed, the edge gateway stores the corresponding video segment, anomaly screenshot, and associated sensor data locally, and sends anomaly alarm information to the cloud platform through the communication module; Step T6, when an anomaly event is not confirmed, the edge gateway only uploads compressed key status data or statistical summaries to reduce communication bandwidth consumption.

[0090] Example 4: System Operation Flow

[0091] The real-time video stream captured by the camera is decoded by the edge gateway and input into the improved YOLO11n model. When the model detects abnormal targets such as component contamination, it simultaneously extracts corresponding voltage, current, and other operating parameter data from a preset time window. The data processing algorithm determines whether there are parameter mutations, continuous deviations from thresholds, or abnormal fluctuations within this time window. If both the video anomaly and the parameter anomaly meet the conditions, a high-confidence alarm message is generated. The alarm message includes the anomaly type, anomaly location, timestamp, screenshot, and video clip, and is uploaded to the cloud platform and monitoring terminal. The cloud management platform receives the abnormal event information and key status data uploaded by the edge gateway, and completes anomaly statistical analysis, historical record tracing, report generation, policy configuration, model version management, and update command issuance. The monitoring terminal is used to display on-site video screenshots, equipment status information, electrical parameter curves, and historical records of abnormal events, and issues alarm notifications.

[0092] Compared with existing technologies, this invention has at least the following beneficial effects: By adding a P2 small target detection branch to the original three-scale detection structure, and combining lightweight feature enhancement, weighted multi-scale fusion, and shallow detail enhancement modules, the detection capability of small targets, low-contrast targets, and complex background anomalies in marine photovoltaic scenes is improved, and the false negative rate is reduced; Addressing issues such as sea surface reflection, local overexposure, shadow occlusion, low contrast, and wind and wave background disturbances in marine photovoltaic images, a data augmentation and training optimization strategy for marine scenes is introduced, improving the robustness and generalization ability of the model in complex environments; integrating video anomaly detection with sensor data... Data processing is integrated at the edge gateway. Through multi-source data fusion and verification, the accuracy of abnormal event confirmation is improved, and false alarms and missed alarms caused by relying solely on a single data form for judgment are reduced. The edge deployment method enables local real-time processing of video streams and sensor data, and only abnormal event information and key status data are uploaded to the cloud, which significantly reduces communication bandwidth consumption and cloud computing power pressure, and improves system real-time performance and network adaptability. The edge gateway is designed to adapt to marine environments such as salt spray resistance, waterproofing, and vibration resistance, and supports remote model updates and parameter configuration, which can meet the needs of long-term continuous operation and remote intelligent operation and maintenance of offshore photovoltaic power plants.

[0093] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An improved target detection method for remote monitoring of offshore photovoltaic power plants, characterized in that, An improved target detection model based on YOLO11n is used to identify and locate abnormal targets in marine photovoltaic images. This improved model includes a newly added P2 small target detection branch, a lightweight feature enhancement module, a shallow detail enhancement module, and a data augmentation training mechanism tailored to marine scenarios. This addresses the issues of missed and false detections caused by small target scale, complex backgrounds, and reflective interference in marine photovoltaic scenes. The method includes the following steps: S1. Collect visible light images of the marine photovoltaic array, label abnormal targets in the images, and construct training, validation and test sets; S2. Construct an improved target detection model, specifically including: based on the original YOLO11n three-scale detection head with P3, P4, and P5 as detection outputs, introduce the P2 shallow high-resolution feature from the backbone network, further upsample the generated P3 feature map to make its spatial resolution consistent with the P2 feature map output by the backbone network, and then concatenate and fuse the upsampled P3 feature map with the P2 feature map to add a P2 small target detection output branch; set a lightweight feature enhancement module in the feature fusion path of the detection head; and set a shallow detail enhancement module at the P2 feature output end or the fusion end after concatenating P2 and P3 features. S3. During model training, the images loaded from the training set are subjected to real-time data augmentation processing for marine photovoltaic scenarios to generate augmented samples. S4. Train the improved target detection model with enhanced sample input, and update the network parameters through joint optimization of bounding box regression loss, classification loss and distribution focus loss; S5. Evaluate the performance of the trained model based on the validation set, and fine-tune the training hyperparameters or model weights according to the evaluation results. S6. Input the image of the marine photovoltaic system to be detected into the improved target detection model after training, and output the location, category and confidence level of the abnormal target.

2. The improved target detection method according to claim 1, characterized in that, The newly added P2 small target detection branch is constructed as follows: based on the original upsampling path from high-level features to low-level features in the detection head, the P3 feature map is further upsampled to make its spatial resolution consistent with the P2 feature map output by the backbone network. Then, the upsampled P3 feature map is concatenated and fused with the P2 feature map to obtain a high-resolution fused feature map. The high-resolution fused feature map is input into the new detection layer to form a joint output structure of four detection scales: P2, P3, P4, and P5. The backbone network retains the four-level feature output of the original YOLO11n backbone network, with downsampling factors of 4x, 8x, 16x, and 32x relative to the input image, respectively, and are denoted as P2, P3, P4, and P5.

3. The improved target detection method according to claim 1, characterized in that, The lightweight feature enhancement module is set in the feature fusion path of the detection head to replace the original upsampling operation, so as to restore edge information and local texture information during the feature map resolution improvement process. The lightweight feature enhancement module includes one or more of pointwise convolution, depthwise separable convolution, and learnable upsampling units. Furthermore, a weighted fusion mechanism is introduced in the multi-scale feature fusion process to set learnable fusion weights for features at different levels, so as to adaptively allocate the contribution ratio of shallow texture features and deep semantic features.

4. The improved target detection method according to claim 1, characterized in that, The shallow detail enhancement module is located at the output end of the P2 feature or at the fusion end after concatenating the P2 and P3 features. It includes a channel compression unit, a local texture extraction unit, an attention recalibration unit, and a residual connection unit. The channel compression unit uses 1×1 convolution to reduce the number of shallow feature channels. The local texture extraction unit uses 3×3 depth convolution or ordinary convolution to extract the edges and fine texture responses of abnormal regions. The attention recalibration unit uses lightweight channel attention to enhance the features of abnormally sensitive regions. The residual connection unit is used to maintain the stable transmission of the original shallow features. The lightweight channel attention includes global average pooling, one-dimensional or two-dimensional convolution transformation, and a sigmoid activation function.

5. The improved target detection method according to claim 1, characterized in that, The data enhancement processing for marine photovoltaic scenarios includes one or more of the following: color perturbation, brightness perturbation, translation transformation, scale scaling, perspective transformation, image stitching enhancement, sea surface high brightness noise simulation, and sea surface reflective spot simulation. During the training process, the ability to locate small targets is improved by adjusting the weight ratio of bounding box regression loss, classification loss and distribution focus loss. When the number of abnormal category samples is lower than that of normal category, a category weight or hard example reinforcement mechanism is applied to the abnormal category. The performance evaluation metrics include one or more of the following: small target average precision, overall average precision, recall rate, and inference speed. The abnormal targets include one or more of the following: surface contamination, foreign object obstruction, cracks, component damage, abnormal connection parts, and abnormal support structure; The backbone network of the improved object detection model maintains the same lightweight scaling configuration as YOLO11n.

6. A method for deploying an edge intelligent algorithm for remote monitoring of offshore photovoltaic power plants, characterized in that, Includes the following steps: A1. Deploy the improved target detection method according to any one of claims 1 to 5 in an edge gateway, and deploy a sensor data processing algorithm; A2. Real-time acquisition of video streams and operational data of offshore photovoltaic power stations via edge gateways, wherein the operational data includes one or more of voltage, current, and temperature data; A3. The video anomaly results are output by improving the target detection model, and the running data anomaly signal is output by the sensor data processing algorithm; the sensor data processing algorithm is used to perform filtering, noise reduction, time alignment, threshold judgment and standardization processing; A4. Perform fusion verification on abnormal video results and abnormal operation data signals. The fusion verification includes time correlation and logical cross verification. A5. After confirming an abnormal event, store relevant data locally and send alarm information to the cloud platform. If the abnormal event is not confirmed, only transmit key status data or statistical summaries to the cloud.

7. The edge intelligence algorithm deployment method according to claim 6, characterized in that, The specific rules for the fusion verification are as follows: when video anomalies and operational data anomalies are detected simultaneously within a preset time window, the event is determined to be a high-confidence anomaly and an alarm is triggered; when only video anomalies are detected and no operational data anomalies are detected, the event is marked as an event to be reviewed; when only operational data anomalies are detected and no video anomalies are detected, secondary sampling confirmation is performed or the event is marked as an event to be reviewed.

8. A remote monitoring system for offshore photovoltaic power plants, characterized in that, This includes edge gateways, cloud management platforms, and monitoring terminals; The edge gateway includes a core processing unit, a data acquisition interface, a communication module, a storage module, and an environmental protection module, used to acquire video streams and running data, and deploy the edge intelligent algorithm deployment method described in claim 6 or 7 to complete real-time analysis, fusion verification, local storage, and preliminary anomaly judgment; The cloud management platform is used to receive abnormal event information and key status data uploaded by the edge gateway, and to complete abnormal statistical analysis, historical record tracing, report generation, policy configuration, model version management and update instruction issuance. The monitoring terminal is used to display on-site video screenshots, equipment status information, electrical parameter curves, and historical records of abnormal events, and to issue alarm notifications.

9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the improved target detection method according to any one of claims 1 to 5, or the edge intelligent algorithm deployment method according to claim 6 or 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the improved target detection method according to any one of claims 1 to 5, or the edge intelligent algorithm deployment method according to claim 6 or 7.

Images