Umbilical cable anomaly detection method

Abstract

The invention discloses an umbilical cable anomaly detection method, and the method comprises the steps: constructing an image data set containing winding, damage and bending, employing a target detection model (including WTConv convolution, a double-domain selection mechanism and a mixed expansion residual attention module) based on deep learning, and combining fine-grained distribution refinement and a global optimal positioning self-distillation algorithm to optimize a prediction result. After the model is trained, the abnormity of the umbilical cable can be detected in real time, online learning is supported, the detection precision is improved, the method is suitable for monitoring the cable state of the underwater robot, and the operation safety and efficiency are remarkably improved.

Description

Technical Field

[0001] The invention belongs to the field of underwater robots, and in particular relates to a method for detecting abnormalities of an umbilical cable for a surface power buoy. Background Art

[0002] Maintenance, inspection, and repair of hydropower dams are crucial for energy supply and ecological protection. However, during these operations, underwater robots (AUVs) often face problems such as umbilical cable drag and entanglement, which significantly limits operational efficiency and safety. To address this issue, using surface power platforms to assist AUVs in cable management is a common technique, with umbilical cable status monitoring being a crucial component. Timely and accurate monitoring of the umbilical cable's status during AUV operation can effectively prevent accidents and ensure the safety and continuity of AUV operations. Traditional methods for detecting umbilical cable anomalies rely on manual visual inspection from shore, which is far from the AUV, making visibility unclear and prone to fatigue. Intelligent umbilical cable status monitoring, deployed on a surface power platform, can effectively address this issue, enabling automated and intelligent detection and significantly improving AUV operational safety. Summary of the Invention

[0003] The object of the present invention is to provide a method for detecting anomalies of an umbilical cable for a surface powered buoy, which is used to solve the problems of low recognition accuracy in anomaly detection of an umbilical cable for a surface powered buoy and low efficiency of the neural network model training process.

[0004] In order to solve the above problems, the technical solution of the present invention is: A method for detecting an umbilical cable anomaly comprises the following steps: S1: Construct an image dataset of umbilical cable anomalies: use image annotation tools to mark the coordinates of the anomaly area, divide the dataset into training set, validation set, and test set according to the set ratio, and convert them into COCO training format; S2: Build a deep learning object detection model: The model includes a backbone network, a hybrid encoder, and a multi-scale decoding prediction head. The backbone network uses WTConv convolution to replace the existing convolution and introduces a dual domain selection mechanism (SSM+FSM); the hybrid encoder integrates a hybrid dilated residual attention module; S3: Optimize bounding boxes and confidence: Improve prediction accuracy through fine-grained localization (FGL) and global optimal localization self-distillation (GO-LSD) algorithms, combined with weighting functions, Hungarian matching, and distillation loss; S4: Model training and optimization: Use batch gradient descent to train the model, set the number of training times, batch size, and learning rate, and select the optimal weight based on the average precision value.

[0005] Furthermore, the backbone network includes a Stem module and a multi-level HG Stage. The Stem module serves as the initial preprocessing layer, separates the high- and low-frequency components of the image through two-dimensional Haar wavelet transform, and uses WTConv convolution to extract multi-band features with a step size of 2; the multi-level HG Stage is composed of multiple HG-Block cascades, and each HG-Block realizes feature fusion through channel splicing, dual-domain selection mechanism and residual connection.

[0006] Furthermore, the dual-domain selection mechanism includes a spatial selection module (SSM) and a frequency selection module (FSM); The spatial selection module (SSM) is based on the CBAM structure and generates a spatial attention map through maximum pooling and average pooling to locate the spatial position of the abnormal area. Its calculation expression is: ; Where: DConvs5,7 is 5×5 and 7×7 cascaded depth convolution, DConvs3 is 3×3 depth convolution The frequency selection module (FSM) separates low-frequency noise through mean filtering and retains high-frequency detail features. Its calculation expression is: ; Where Fs is the SSM output feature, and Mean(Fs) is the channel-level global average pooling result.

[0007] Furthermore, the hybrid dilated residual attention module includes a dilated convolution group, a residual connection and a channel attention module (CAM).

[0008] Furthermore, the decoding prediction head includes a bounding box regression head and a D-FINE head. The prediction result is optimized by the following calculation method: the probability distribution output by the D-FINE head is used to calculate the refined edge distance in combination with the weighting function W(n). The weighting function expression is: ; Where n is the index of the predicted bounding box, and a and c are hyperparameters that control the curvature of the function and the bounds.

[0009] Furthermore, fine-grained localization loss (FGL Loss) is used to optimize distribution prediction. The formula is: ; In the formula Bounding box offset weights, by real edge distance Deviation from the prediction Calculated.

[0010] Furthermore, the global optimal location self-distillation (GO-LSD) algorithm includes: applying the Hungarian matching algorithm to the prediction results of each decoder layer to identify the cross-layer bounding box matching relationship; fusing the multi-layer predictions through the decoupled distillation focal loss (DDF Loss), the formula is: ; Where: and are the weights of matching and non-matching predictions, respectively, which are dynamically adjusted by IoU and classification confidence (Conf).

[0011] Furthermore, data enhancement strategies are introduced during the dataset construction process, including random flipping, Gaussian noise addition, contrast transformation and scale scaling, to improve the generalization ability of the model.

[0012] Furthermore, the umbilical cable abnormal image dataset in step S1 is divided into: umbilical cable entanglement, umbilical cable damage, and umbilical cable bending; the umbilical cable abnormal image dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0013] Furthermore, in step S4, the number of training times is 200, the batch size is 16, and the learning rate is 0.01.

[0014] The beneficial effects of the present invention are: 1. This backbone part uses WTConv convolution to replace ordinary convolution, designs and adds a dual-domain selection mechanism as a feature aggregation method for feature fusion, which can better extract umbilical cable abnormal features.

[0015] 2. The encoder uses a hybrid dilation residual attention module to achieve dual optimization of noise suppression and detail preservation, which improves the model's detection accuracy of umbilical cable anomalies.

[0016] 3. The bounding box prediction accuracy and confidence are optimized through the fine-grained localization (FGL) and global optimal localization self-distillation (GO-LSD) algorithms, further improving the reliability of the detection results.

[0017] 4. The introduction of data enhancement strategy improves the generalization ability of the model, enabling it to better adapt to different detection environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a structural diagram of a new target detection model based on deep learning in the present invention; Figure 2 A structural model diagram of a dual-domain selection mechanism provided by the present invention; Figure 3 A structural model diagram of a hybrid expansion residual attention module provided by the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] See Figure 1 As shown, the present invention provides an umbilical cable abnormality detection method, which specifically includes the following steps: S1: Construct an umbilical cable anomaly image dataset. Screen the collected images based on whether they contain an umbilical cable anomaly, ensuring that the number of labels for various umbilical cable anomalies is roughly equal. Use image annotation tools to annotate the preliminarily screened images, primarily marking umbilical cable entanglement, umbilical cable damage, and umbilical cable bends. Split the annotated data into training, validation, and test sets in an 8:1:1 ratio, and convert them to COCO format for training.

[0021] S2: Build a deep learning target detection model, which consists of a backbone part, a hybrid encoder part, and a decoding prediction head, such as Figure 2 As shown in the figure, the backbone part uses WTConv convolution to replace the existing convolution, and designs and adds a dual-domain selection mechanism as a feature aggregation method for feature fusion; the encoder uses a hybrid expansion residual attention module to achieve dual optimization of noise suppression and detail preservation; S2.1: The backbone consists of a Stem module and multiple HG-Blocks. The Stem module serves as the initial preprocessing layer of the network, replacing standard convolution operations with a series of WTConv convolutions to extract features from the raw input data. WTConv uses a cascaded wavelet transform decomposition and performs a series of convolutions with smaller kernels. Each convolution focuses on different frequency bands of the input across increasingly larger receptive fields, placing greater emphasis on low frequencies while adding only a small number of trainable parameters. Specifically, given an image X as an input tensor, a first-level Haar wavelet transform in one spatial dimension is given by a depthwise convolution with kernels [1, 1] / 2 and [1, −1] / 2, followed by a standard downsampling operator by a factor of 2. To perform a two-dimensional Haar wavelet transform, the operations are combined in two dimensions, using four filters, resulting in a depthwise convolution with a stride of 2. W is the weight tensor of the k × k depthwise kernel, with four times the number of input channels as X. This operation not only separates the convolution between frequency components, but also allows smaller kernels to operate over a larger area of ​​the original input, increasing the receptive field of the input.

[0022] S2.2: The HG-Block concatenates all feature maps in the channel dimension through continuous convolution, then selects a dual-domain selection mechanism as a feature aggregation method for feature fusion, and finally uses a residual connection as the final output. The dual-domain selection mechanism consists of two parts: the spatial selection module (SSM) and the frequency selection module (FSM). The SSM receives feature input and uses depthwise convolution to determine the degradation locations of each channel, allowing the network to focus on important spatial regions and provide initial positions for the FSM. It has three branches, and the main path is based on CBAM. For the intermediate feature map F, it is first squeezed by maximum and average pooling, and then convolution is used to generate a general feature map with the focused degradation locations. After that, depthwise convolution is used to perform channel separation transformation. The process is represented as follows: ; Where F represents the input feature map; MaxPool represents maximum pooling; AvgPool represents average pooling; represents a cascade of depthwise convolutional layers with kernel sizes of 5 × 5 and 7 × 7; Represents a depthwise convolution with 3×3 kernels; the spatially selected features are then fed to the FSM to amplify high-frequency signals or key areas by removing low frequencies from the features, and directly utilize To assist the restoration process. First, a mean filter is applied to Fs to generate low-frequency features. Then, the complementary high-frequency features are obtained by subtracting the resulting low-frequency signal from the input. The mean filter is implemented by channel-level global average pooling to remove defocus blur. The final output feature result is expressed as: ; Where F s is the output feature of SSM, Mean (F s ) is the channel-level global average pooling result.

[0023] S2.3: The output from step S2.1 is used as the encoder input. A hybrid dilated residual attention module (HADM) is implemented, consisting of a hybrid dilated residual block and a channel attention module (CAM). This module, with dilated convolutions at dilation rates of 1 to 4 as its core, performs multi-scale analysis, smooths noise, and expands the receptive field. Residual connections are used to mitigate gradient issues. Each dilated convolution is followed by a ReLU to enhance feature representation. Global average pooling is performed at the end to aggregate information, and the CAM is integrated to focus on key areas.

[0024] S3, Optimizing Bounding Box Prediction Accuracy and Confidence: The network structure model marks these umbilical cable anomalies with rectangular prediction boxes and gives the corresponding prediction box coordinates, labels, and intersection-over-union (IOU) ratios. The steps of the umbilical cable anomaly image evaluation method are as follows: S3.1: Preliminary bounding boxes and probability distributions are predicted using the traditional bounding box regression head and the D-FINE head through the first decoder layer of the D-FINE model; S3.2: For each layer l (from layer 1 to layer L), use the weighting function and probability distribution to calculate the refined edge distance. The weighting function W(n) is defined as follows: ; Where n is the index of the predicted bounding box, and a and c are hyperparameters that control the curvature of the function and the bounds.

[0025] And the fine-grained localization (FGL) loss is used to further improve the accuracy of distribution prediction. The calculation formula of FGL is as follows.

[0026] ; ; in represents the probability distribution corresponding to the k-th prediction. is the relative offset, and the calculation formula is . represents the edge distance of the real data, Represents The adjacent container index. The weight is The cross entropy (CE) loss shows that the interpolation between bounding boxes is accurately aligned with the ground truth offset; S3.3: Apply Hungarian Matching to the predictions of each layer of the model, identifying local bounding box matches at each stage and aggregating the matching indices from all layers into a unified union. This combines the most accurate candidate predictions across layers to optimize global matching while processing unmatched predictions to improve overall stability.

[0027] S3.4: Decoupled Distillation Focal Loss (DDF) loss calculation and optimization, applying a decoupled weighting strategy to ensure that predictions with high IoU but low confidence are given appropriate weights, and finally obtain the final bounding box and confidence. DDF The formula is: ; ; ; Among them, IoU k is the intersection-over-union ratio of the k-th prediction box, is the classification confidence, KL represents the Kullback-Leibler divergence, which is used to measure the difference between the predicted distribution and the true distribution, and T is the temperature parameter used to smooth the logits. The distillation loss of the k-th matching prediction is Weighted, where and 1- are the weights of matched and unmatched predictions, respectively, which are determined by LoU and classification confidence ( ) dynamic adjustment, where and are the number of matched and unmatched predictions, respectively. For the kth unmatched prediction, the weight is (1- )* ;in Represents the classification confidence.

[0028] S4. Use the training set to train the new deep learning network structure model to obtain the trained model and weights. The training parameters are set as follows: 200 training epochs, a batch size of 16, and a learning rate of 0.01. During training, model performance is evaluated by calculating the mean average precision (mAP). The training weight with the highest mAP is selected as the weight for the final model.

[0029] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The scope of protection of the present invention should not be regarded as limited to the specific forms described in the embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for detecting anomalies in an umbilical cable, characterized in that: The steps include: S1: Construct an image dataset of umbilical cable anomalies: use image annotation tools to mark the coordinates of the anomaly area, divide the dataset into training set, validation set, and test set according to the set ratio, and convert them into COCO training format; S2: Build a deep learning object detection model: The model includes a backbone network, a hybrid encoder, and a multi-scale decoding prediction head. The backbone network uses WTConv convolution to replace the existing convolution and introduces a dual domain selection mechanism (SSM+FSM); the hybrid encoder integrates a hybrid dilated residual attention module; S3: Optimize bounding boxes and confidence: Improve prediction accuracy through fine-grained localization (FGL) and global optimal localization self-distillation (GO-LSD) algorithms, combined with weighting functions, Hungarian matching, and distillation loss; S4: Model training and optimization: Use batch gradient descent to train the model, set the number of training times, batch size, and learning rate, and select the optimal weight based on the average precision value.

2. The method for detecting anomalies of an umbilical cable according to claim 1, wherein: The backbone network includes a Stem module and a multi-level HG Stage. The Stem module serves as the initial preprocessing layer, separates the high- and low-frequency components of the image through two-dimensional Haar wavelet transform, and uses WTConv convolution to extract multi-band features with a step size of 2. The multi-level HG Stage is composed of multiple HG-Block cascades, and each HG-Block achieves feature fusion through channel splicing, dual-domain selection mechanism and residual connection.

3. The method for detecting anomalies of an umbilical cable according to claim 2, characterized in that: The dual-domain selection mechanism includes a spatial selection module (SSM) and a frequency selection module (FSM); The spatial selection module (SSM) is based on the CBAM structure and generates a spatial attention map through maximum pooling and average pooling to locate the spatial position of the abnormal area. Its calculation expression is: ； Where: DConvs5,7 is 5×5 and 7×7 cascaded depth convolution, DConvs3 is 3×3 depth convolution The frequency selection module (FSM) separates low-frequency noise through mean filtering and retains high-frequency detail features. Its calculation expression is: ； Where Fs is the SSM output feature, and Mean(Fs) is the channel-level global average pooling result.

4. The method for detecting anomalies of an umbilical cable according to claim 1, wherein: The hybrid dilated residual attention module includes a dilated convolution group, a residual connection and a channel attention module (CAM).

5. The method for detecting anomalies of an umbilical cable according to claim 1, characterized in that: The decoding prediction head includes a bounding box regression head and a D-FINE head. The prediction results are optimized by the following calculation method: the probability distribution output by the D-FINE head is used to calculate the refined edge distance in combination with the weighting function W(n). The weighting function expression is: ； Where n is the index of the predicted bounding box, and a and c are hyperparameters that control the curvature of the function and the bounds.

6. The method for detecting anomalies of an umbilical cable according to claim 5, characterized in that: Fine-grained localization loss (FGL Loss) is used to optimize distribution prediction. The formula is: ； In the formula Bounding box offset weights, by real edge distance Deviation from the prediction Calculated.

7. The method for detecting anomalies of an umbilical cable according to claim 1, characterized in that: The Global Optimal Localization Self-Distillation (GO-LSD) algorithm involves applying the Hungarian matching algorithm to the prediction results of each decoder layer to identify cross-layer bounding box matching relationships; and fusing multiple layers of predictions using the Decoupled Distillation Focal Loss (DDF Loss), as follows: ； Where: and are the weights of matching and non-matching predictions, respectively, which are dynamically adjusted by IoU and classification confidence (Conf).

8. The method for detecting anomalies of an umbilical cable according to claim 1, characterized in that: Data enhancement strategies are introduced during the dataset construction process, including random flipping, Gaussian noise addition, contrast transformation and scale scaling, to improve the generalization ability of the model.

9. The method for detecting abnormalities in an umbilical cable according to any one of claims 1 to 8, characterized in that: The umbilical cable abnormal image dataset in step S1 is divided into: umbilical cable entanglement, umbilical cable damage, and umbilical cable bending; the umbilical cable abnormal image dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:

1.

10. The method for detecting abnormalities in an umbilical cable according to any one of claims 1 to 8, characterized in that: In step S4, the number of training times is 200, the batch size is 16, and the learning rate is 0.01.

Images