A device appearance and environment anomaly detection method for intelligent inspection of an offshore wind farm booster station

Abstract

The present application relates to a kind of equipment appearance and environment anomaly detection method for offshore wind farm booster station intelligent inspection, first by reversible column main network RCN processing complex feature, then by Neck Multi-scale feature fusion is carried out, finally using SA Head generates final detection result: including the class of target, boundary box coordinate and confidence degree.And in order to better solve the influence of low-quality example to model, a kind of dynamic non-monotonic focusing method Wise-IoU is used in network training process.The present application can realize the accurate and efficient detection of equipment ontology and surrounding environment anomaly in offshore wind farm booster station, find the abnormal change of equipment ontology and surrounding environment, can provide direct and important basis for preventive maintenance of offshore wind farm booster station.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for offshore wind farms, and more specifically, to a method for detecting equipment appearance and environmental anomalies during intelligent inspection of offshore wind farm booster stations. Background Technology

[0002] Offshore wind farm booster stations are the core hubs for collecting and transmitting electrical energy, and their safe and stable operation is a crucial link in ensuring the power generation efficiency of the entire wind farm and the safety of the power grid. Considering the harsh offshore environment and the high cost and significant risks of manual inspections, the adoption of intelligent inspection systems based on robots, drones, or fixed sensors has become an important technical means to ensure the efficient and safe operation of booster stations. In the entire intelligent inspection system, the front-end image acquisition capability is relatively fixed; therefore, the accuracy and reliability of the back-end anomaly detection algorithm directly determines the final efficiency of the entire system. Poor detection results not only reduce inspection efficiency but may also deteriorate the operational safety of the entire booster station. Existing intelligent inspection methods for equipment appearance and environmental anomalies in offshore wind farm booster stations have poor capabilities in "segmenting" and "identifying" abnormal features, heavily relying on isolated, simple, and insufficiently robust algorithm models. This limitation directly prevents the system from conducting in-depth, interconnected, and global state assessments, severely undermining the core value of intelligent inspection systems in achieving "preventive maintenance." It may cause maintenance personnel to miss the optimal time to address early-stage faults, ultimately allowing minor issues to escalate into major accidents, posing a significant risk to the safe and stable operation of offshore wind farms. Therefore, there is an urgent need for a novel anomaly detection method that can effectively detect multiple types of anomalies, possesses contextual reasoning capabilities, and exhibits high robustness to environmental changes. Summary of the Invention

[0003] The purpose of this invention is to provide an anomaly detection method for intelligent inspection of offshore wind farm booster stations. This method can accurately and efficiently detect anomalies in the equipment itself and the surrounding environment of offshore wind farm booster stations, and discover abnormal changes in the equipment itself and the surrounding environment. This provides a direct and important basis for preventive maintenance of offshore wind farm booster stations.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A method for detecting equipment appearance and environmental anomalies in intelligent inspection of offshore wind farm booster stations is characterized by constructing a deep learning detection network, wherein the deep learning detection network sequentially includes a reversible column backbone network (RCN), a feature pyramid network (RFPN), and a detection head (SA Head) integrated with a SEAM attention module. The training method of the network is improved by introducing a dynamic non-monotonic focusing method (Wise-IoU).

[0006] This invention aims to achieve accurate and efficient detection of anomalies in the equipment and surrounding environment of offshore wind farm booster stations. It utilizes a deep learning detection network to efficiently detect abnormal changes in the equipment and its environment. The algorithm consists of three parts: Backbone, Neck, and Head. First, a Reversible Column Network (RCN) Backbone is designed to improve the efficiency of large-scale training tasks within booster stations. Second, a novel Feature Pyramid (RFPN) is designed to enhance the model's multi-scale feature fusion capability. Finally, a new SEAM Attention (SA) Head is designed, introducing SEAM Attention to improve the network's ability to perceive occluded objects and enhance robustness.

[0007] The basic principle of the deep learning network model for detecting equipment and environmental anomalies in offshore wind farm booster stations consists of the following three steps:

[0008] 1. S1: Deep learning detection network model processes data.

[0009] (1) The input image enters the Backbone, and low-level to high-level features of the image are extracted step by step through multiple convolutions. This network designs an RCN Backbone that reconstructs the Backbone structure; the main body of the RCN consists of multiple sub-networks (columns), which are connected by multi-level reversible connections. This architecture design allows the features in the RCN to be gradually unpacked as they pass through each column during forward propagation, and all their information is preserved, rather than being compressed or discarded like in other networks. Formally, forward and inverse computation follows the equation:

[0010]

[0011]

[0012] in Represents the inputs at each level of the network. Represents any nonlinear operation of the residual function. To represent a simple reversible operation, its inverse operation is... express.

[0013] In terms of parameter settings, the number of RCN channels is set to [64, 128, 256, 512] to control the computational complexity of the network. RCN enables the network to process complex features more effectively, and through reversible design, it reduces memory usage, allows for training deeper networks, and has better performance in detection tasks that require large-scale data and complex computation.

[0014] (2) The Neck, as the second part of the network structure, is responsible for further fusing features from different scales output by the Backbone and providing multi-scale feature information to the Head. A novel Feature Pyramid RFPN was designed, employing different scale feature maps and channel sizes in the Neck feature fusion. The performance of using the same and different channels was compared, as well as the accuracy advantage gained from the trade-off between Neck depth and width. By flexibly controlling the number of channels at different scales, the network can achieve higher accuracy than sharing the same channels at all scales. Ideal performance can be obtained when the depth is equal to 3 and the width is equal to [96, 192, 384].

[0015] (3) The Head is the final part of the object detection process, responsible for generating the final detection results, including the object's category, bounding box coordinates, and confidence score. A new detection head, SAHead, is designed by introducing SEAM Attention into the head to efficiently identify occluded targets, addressing potential appearance defects, foreign object occlusion, and positional anomalies that may occur in the equipment within the booster station during the detection process. The first part of SEAM is a depthwise separable convolution with residual connections, reducing the number of parameters. The outputs of convolutions at different depths are combined through pointwise (1x1) convolutions in the second part. A hierarchical fully connected network is used to fuse information from each channel, thereby strengthening the connections between channels. By learning the relationship between occluded and unoccluded targets in the previous step, the loss in the occlusion scenario is compensated. Finally, the output of SEAM is used as attention, multiplied by the original features, enabling the network to more effectively handle the problem of target occlusion and improving the target detection performance of the Head.

[0016] 2. S2: Output metrics of deep learning detection network models

[0017] The abnormal dataset collected from wind farm booster stations at the front end often contains low-quality samples. Continuously reinforcing the bounding box regression of these low-quality samples may hinder further improvement in the model's detection capabilities. To address the impact of low-quality examples on the model, Focal-EIoU was proposed, but it employs a static focusing strategy, failing to fully leverage the potential of non-monotonic focusing mechanisms. Therefore, a dynamic non-monotonic focusing method, Wise-IoU, is proposed.

[0018]

[0019] in As a special penalty term for Wise-IoU, it is used to significantly amplify anchor frames of ordinary quality. .

[0020] To more effectively avoid harmful gradients from low-quality samples, a coefficient is introduced. Wiser-IoU is obtained by multiplying Wise-IoU by a coefficient r, and the formula is as follows:

[0021]

[0022] 3. S3: Detection is performed based on the total loss function of the deep learning detection network model.

[0023] A deep learning network structure for detecting equipment and environmental anomalies at offshore wind farm booster stations is constructed using S1. The total loss function for training this deep learning network is... The training, testing, and verification of the deep learning detection network for anomalies in substation equipment and environment are carried out to obtain the trained anomaly detection network. Based on this anomaly detection network, leakage detection, rust / corrosion detection, foreign object detection, and surface damage / deformation detection can be performed on transformer bodies, switch cabinets, cables, floors, walls, fire-fighting equipment, etc.

[0024]

[0025] in, For the training of the device and environmental anomaly detection network, The original inspection images of the equipment and environment acquired by the vision system on the inspection equipment. This refers to the abnormal feature information, which includes location and category information, obtained after detecting abnormal features in the original inspection image.

[0026] In summary, the present invention can be combined with the following further technical solutions:

[0027] The main body of the reversible column backbone network consists of multiple sub-networks, and these sub-networks are connected in a multi-level reversible manner. Its forward and inverse calculations follow the following equations:

[0028] Forward calculation: ;

[0029] Reverse calculation: ;

[0030] in Represents the inputs at each level of the network. Represents any nonlinear operation of the residual function. To represent a simple reversible operation, its inverse operation is... express.

[0031] The number of channels in the reversible column backbone network is configured as [64, 128, 256, 512] to control the computational complexity of the network.

[0032] The feature pyramid network has a depth of 3 when performing feature fusion, and the number of channels in the feature layers at different scales is configured as [96, 192, 384].

[0033] The detection head integrating the SEAM attention module has a first part of the SEAM attention module, which is a depthwise separable convolution with residual connections. The outputs of convolutions at different depths are combined through pointwise convolutions in the second part, and a hierarchical fully connected network is used to fuse the information of each channel to generate attention.

[0034] The dynamic non-monotonic focusing method is the Wise-IoU method, and its loss function is: ,in As a penalty, it is used to amplify the quality of ordinary anchor frames. .

[0035] Based on the Wise-IoU method, by introducing coefficients The Wiser-IoU method has the following loss function: .

[0036] The anomaly detection method for intelligent inspection of offshore wind farm booster stations of the present invention can accurately and efficiently detect anomalies in the equipment itself and the surrounding environment of offshore wind farm booster stations, and discover abnormal changes in the equipment itself and the surrounding environment, which can provide a direct and important basis for preventive maintenance of offshore wind farm booster stations. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the deep learning intelligent detection network structure for substation equipment and environmental anomalies according to the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be described in detail below. These embodiments are not intended to limit the scope of this invention.

[0039] 1. Data preparation and network construction

[0040] Dataset Construction: Historical images of offshore wind farm booster stations (including visible light, infrared, etc.) were collected, and experts labeled various abnormal targets, such as "leakage", "corrosion", "foreign objects", "damage", etc., to form standardized training datasets, validation datasets and test datasets.

[0041] Network Model Construction: Construct a complete detection network model according to the following scheme:

[0042] The deep learning detection network consists of an invertible column backbone network (RCN), a feature pyramid network (RFPN), and a detection head (SA Head) integrating a SEAM attention module. The training method of the network is improved by introducing a dynamic non-monotonic focusing method (Wise-IoU).

[0043] 2. Network Model Training

[0044] Loss function setting: The total loss function of the deep learning detection network is a weighted average of the classification loss function and the bounding box regression loss function. In this embodiment, the bounding box regression task adopts the Wiser-IoU loss function of formula (2.2) in the invention. By optimizing this total loss function, the model can simultaneously learn accurate anomaly category judgment and location localization.

[0045] Model training: Input the prepared training dataset into the constructed network and use the gradient descent algorithm for iterative optimization to minimize the total loss function. Monitor performance on the validation set to prevent overfitting until the model converges.

[0046] 3. Application of Network Models

[0047] The trained network model is deployed in the intelligent inspection system of the offshore wind farm's booster station. During the inspection, visual sensors capture real-time images. It is then fed into the trained anomaly detection network model. The network automatically processes the image and outputs the detection results. .

[0048] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting equipment appearance and environmental anomalies during intelligent inspection of offshore wind farm booster stations, characterized in that, A deep learning detection network is constructed, wherein the deep learning detection network sequentially includes an invertible column backbone network (RCN), a feature pyramid network (RFPN), and a detection head (SA Head) integrating a SEAM attention module. The training method of the network is improved by introducing a dynamic non-monotonic focusing method (Wise-IoU).

2. The method for detecting equipment appearance and environmental anomalies for intelligent inspection of offshore wind farm booster stations according to claim 1, characterized in that, The main body of the reversible column backbone network consists of multiple sub-networks, and these sub-networks are connected in a multi-level reversible manner. Its forward and inverse calculations follow the following equations: Forward calculation: ; Reverse calculation: ; in Represents the inputs at each level of the network. Represents any nonlinear operation of the residual function. To represent a simple reversible operation, its inverse operation is... express.

3. The method for detecting equipment appearance and environmental anomalies for intelligent inspection of offshore wind farm booster stations according to claim 1, characterized in that, The number of channels in the reversible column backbone network is configured as [64, 128, 256, 512] to control the computational complexity of the network.

4. The method for detecting equipment appearance and environmental anomalies for intelligent inspection of offshore wind farm booster stations according to claim 1, characterized in that, The feature pyramid network has a depth of 3 when performing feature fusion, and the number of channels in the feature layers at different scales is configured as [96, 192, 384].

5. The method for detecting equipment appearance and environmental anomalies for intelligent inspection of offshore wind farm booster stations according to claim 1, characterized in that, The detection head integrating the SEAM attention module has a first part of the SEAM attention module, which is a depthwise separable convolution with residual connections. The outputs of convolutions at different depths are combined through pointwise convolutions in the second part, and a hierarchical fully connected network is used to fuse the information of each channel to generate attention.

6. The method for detecting equipment appearance and environmental anomalies for intelligent inspection of offshore wind farm booster stations according to claim 1, characterized in that, The dynamic non-monotonic focusing method is the Wise-IoU method, and its loss function is: ,in As a penalty, it is used to amplify the quality of ordinary anchor frames. .

7. A method for detecting equipment appearance and environmental anomalies for intelligent inspection of offshore wind farm booster stations according to claim 6, characterized in that, Based on the Wise-IoU method, by introducing coefficients The Wiser-IoU method has the following loss function: .

Images