Submarine cable fault diagnosis method under poor data condition based on confrontation generation

Abstract

The invention discloses a submarine cable fault diagnosis method based on confrontation generation under a poor data condition. The method comprises the following steps: constructing an MPAGCNN model which comprises an adversarial generative network and a fault diagnosis network; obtaining a source sample data set, inputting real data and random noise in the source sample data set into an adversarial generation network for training, optimizing the model by minimizing adversarial generation loss until a generator and a discriminator reach Nash equilibrium, and obtaining a generated sample data set by using the current generator; merging the generated sample data set with the source sample data set to obtain a comprehensive sample data set; and inputting the comprehensive sample data set into a fault diagnosis network for training, optimizing the model by minimizing fault classification positioning loss until a preset performance threshold is met, and outputting the current fault diagnosis network as a final submarine cable fault diagnosis model. According to the invention, the reliability of submarine cable fault diagnosis can be improved.

Description

Technical Field

[0001] This invention relates to the field of cable fault detection technology, and in particular to a method for diagnosing submarine cable faults under poor data conditions based on adversarial generation. Background Technology

[0002] Submarine cables are the primary medium for interconnecting offshore power grids, acquiring power from offshore platforms, and transmitting clean energy from the sea. Due to the harsh and complex working environment, submarine cables are prone to failure during operation, and troubleshooting and locating these faults are challenging, while repairs are difficult and time-consuming. Submarine cable faults not only affect their own normal operation but can also lead to the disconnection and shutdown of other systems, resulting in greater economic losses and even threatening personal safety. Therefore, effective inspection of the working condition of submarine cables, and the automatic and accurate identification and location of potential fault threats, are of great significance and value in ensuring the safe and stable operation of offshore power systems.

[0003] In recent years, deep learning has made significant progress in the field of fault diagnosis. Scholars have conducted extensive research on deep learning methods for fault diagnosis of onshore cables, achieving certain performance results and publishing numerous findings. However, due to the poor conditions of submarine cable inspections, it is difficult to obtain a large number of well-balanced training samples. This is highly detrimental to the large number of balanced samples required for deep learning-based detection models, resulting in poor accuracy and generalization of the trained models. Therefore, how to achieve data augmentation for rare sample categories to eliminate severe data imbalance and effectively train the model is a significant challenge. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a submarine cable fault diagnosis method based on adversarial generation under poor data conditions, which can improve the reliability of submarine cable fault diagnosis.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A method for fault diagnosis of submarine cables under data-scarce conditions based on adversarial generation includes the following steps:

[0007] Construct an MPAGCNN model, which includes a generative adversarial network and a fault diagnosis network;

[0008] Obtain the source sample dataset, input the real data and random noise from the source sample dataset into the adversarial generative network for training, optimize the model by minimizing the adversarial generative loss until the generator and discriminator reach Nash equilibrium, and use the current generator to obtain the generated sample dataset; merge the generated sample dataset with the source sample dataset to obtain the comprehensive sample dataset;

[0009] The comprehensive sample dataset is input into the fault diagnosis network for training. The model is optimized by minimizing the fault classification and localization loss until a preset performance threshold is met. The current fault diagnosis network is then output as the final submarine cable fault diagnosis model.

[0010] The data to be predicted is input into the final submarine cable fault diagnosis model for reasoning and diagnosis to obtain fault diagnosis results.

[0011] Preferably, the adversarial generative network includes a generator and a discriminator. The generator includes a linear layer, a batch normalization layer, an activation function, five feature recovery modules 1, 2, 3, 4, and 5 with identical structures, connected in sequence, as well as a transposed convolutional layer and an activation function. The feature recovery module 1 includes a transposed convolutional layer, a batch normalization layer, and an activation function connected in sequence. The discriminator includes a convolutional layer, an activation function, five feature extraction modules A, B, C, D, and E with identical structures, connected in sequence, as well as a linear layer. The feature extraction module A includes a convolutional layer, a batch normalization layer, and an activation function connected in sequence.

[0012] Preferably, the fault diagnosis network includes a feature extractor and a fault detector. The feature extractor includes five identical feature extraction modules 1, 2, 3, 4, and 5 connected in sequence. Feature extraction module 1 includes a convolutional layer, a batch normalization layer, an activation function, and a pooling layer connected in sequence. The fault detector includes three prediction modules respectively located after feature extraction modules 1, 3, and 5. The prediction modules include a fully connected layer and a Softmax layer.

[0013] Preferably, obtaining the source sample dataset includes the following steps:

[0014] Obtain submarine cable dataset;

[0015] The fault types in the submarine cable dataset are labeled to obtain the source sample dataset.

[0016] Preferably, the adversarial generation loss is as follows:

[0017]

[0018] In the formula, λ1 and λ2 are hyperparameters with scheduling strategy, and L RpGAN The relative matching loss is expressed by the formula:

[0019]

[0020] Where f is the activation function, and D ψ It is a discriminator, G θ Here, z is the generator, and x is the noise distribution.

[0021] R1 is a penalty applied to the gradient of the discriminator on the real data, and the formula is:

[0022]

[0023] R² is a penalty applied to the gradient of the discriminator on the generated data, as shown in the formula:

[0024] .

[0025] Preferably, the fault classification and location loss L y As shown below:

[0026]

[0027] In the formula, M represents the total number of prediction modules. Let the loss weight corresponding to the m-th prediction module satisfy the following condition: , ; The fault classification and localization loss value of the m-th prediction module is defined as:

[0028]

[0029] Where K represents the number of prediction boxes output by the prediction module; This represents the intersection-union ratio (IoU) between the k-th predicted bounding box and its corresponding ground truth bounding box. This represents the category prediction probability of the k-th prediction box; This represents the true category label corresponding to the k-th predicted bounding box; and Let represent the adaptive weight coefficients of the positive and negative samples, respectively, and satisfy . ; Represents the distributed focus loss function. The weighted classification loss function that takes into account class imbalance is defined as follows:

[0030]

[0031] In the formula, N represents the number of samples and C represents the number of fault categories; This represents the predicted probability that the i-th sample belongs to the c-th class; As a category balance factor;

[0032] To characterize the spatial offset relationship between the predicted bounding box and the ground truth bounding box, a normalized position offset vector is introduced. Its definition is:

[0033]

[0034] In the formula , This represents the center coordinates of the actual fault area. , This represents the width and height of the actual fault area; , Indicates the center coordinates of the initial prediction box. , H represents the width and height of the initial predicted bounding box; H and W represent the height and width of the feature map or input image, respectively.

[0035] Adaptive weighting coefficients for positive and negative samples and , represented as:

[0036]

[0037] In the formula, , These represent the positive and negative sample indices corresponding to the current prediction box, respectively. The weight adjustment function dynamically adjusts the sample weights based on the prediction difficulty, and is defined as follows:

[0038]

[0039] Where a is the upper limit of the weight adjustment function; c is the curvature adjustment parameter, used to control the rate of weight change;

[0040] B represents the total number of predicted boxes, and n represents the sorting index of the spatial offset of the predicted boxes relative to the ground truth boxes. It is obtained by sorting the offsets of all predicted boxes in the same prediction module, and the value range is [0, B].

[0041] Based on the above technical solution, the beneficial effects of this invention are as follows: This invention discloses a method for submarine cable fault diagnosis under limited data conditions based on adversarial generation, comprising the following steps: constructing an MPAGCNN model, including an adversarial generative network and a fault diagnosis network; obtaining a source sample dataset, inputting real data and random noise from the source sample dataset into the adversarial generative network for training, optimizing the model by minimizing the adversarial generation loss until the generator and discriminator reach Nash equilibrium, and obtaining a generated sample dataset using the current generator; merging the generated sample dataset with the source sample dataset to obtain a comprehensive sample dataset; inputting the comprehensive sample dataset into the fault diagnosis network for training, optimizing the model by minimizing the fault classification and localization loss until a preset performance threshold is met, and outputting the current fault diagnosis network as the final submarine cable fault diagnosis model; inputting the data to be predicted into the final submarine cable fault diagnosis model for inference and diagnosis, and obtaining the fault diagnosis result. This invention can improve the accuracy of submarine cable fault diagnosis under limited fault data. This model introduces a multi-penalty adversarial generation mechanism on the basis of traditional convolutional neural networks, making the samples more balanced under multi-classification data. This method can effectively improve convergence, robustness, accuracy and generalization of model training under conditions of scarce data, thereby improving the reliability of submarine cable fault diagnosis. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating a method for diagnosing submarine cable faults under data-poor conditions based on adversarial generation in one embodiment.

[0043] Figure 2 This is a schematic diagram of the MPAGCNN model in one embodiment;

[0044] Figure 3 This is a schematic diagram of the fault diagnosis process in one embodiment. Detailed Implementation

[0045] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0046] like Figures 1 to 3 As shown, this embodiment provides a method for fault diagnosis of submarine cables under scarce data conditions based on adversarial generation, including the following steps:

[0047] Step 100: Construct the MPAGCNN model, which includes a generative adversarial network and a fault diagnosis network.

[0048] In this embodiment, a Multi-Penalty Adversarial Generation with Convolutional Neural Network (MPAGCNN) model is constructed. See the schematic diagram of the overall structure of this MPAGCNN model. Figure 2 The diagram uses solid and hollow arrows to represent the flow of source sample data and generated data during model training, respectively, and green and blue dashed boxes to distinguish between the fault diagnosis network and the generative adversarial network.

[0049] The Generative Adversarial Network (GAN) includes a generator and a discriminator. The generator comprises a linear layer, a batch normalization layer, an activation function, five identical feature recovery modules (Module 1, Module 2, Module 3, Module 4, and Module 5), a transposed convolutional layer, and an activation function, all connected in sequence. The feature recovery module 1 comprises a transposed convolutional layer, a batch normalization layer, and an activation function, all connected in sequence. The discriminator comprises a convolutional layer, an activation function, five identical feature extraction modules (Module A, Module B, Module C, Module D, and Module E), and a linear layer, wherein feature extraction module A comprises a convolutional layer, a batch normalization layer, and an activation function, all connected in sequence.

[0050] The fault diagnosis network includes a feature extractor and a fault detector. The feature extractor includes five identical feature extraction modules 1, 2, 3, 4, and 5 connected in sequence. Feature extraction module 1 includes a convolutional layer, a batch normalization layer, an activation function, and a pooling layer connected in sequence. The fault detector includes three prediction modules respectively located after feature extraction modules 1, 3, and 5. Each prediction module includes a fully connected layer and a Softmax layer. The three prediction modules output prediction results for the degree of fault damage, fault type, and fault location, respectively.

[0051] Step 200: Obtain the source sample dataset. Input the real data and random noise from the source sample dataset into the adversarial generative network for training. Optimize the model by minimizing the adversarial generative loss until the generator and discriminator reach Nash equilibrium. Use the current generator to obtain the generated sample dataset. Merge the generated sample dataset with the source sample dataset to obtain the comprehensive sample dataset.

[0052] In this embodiment, data samples that reflect the operation and fault types of submarine cables are collected to construct a submarine cable dataset; the fault types in the submarine cable dataset are labeled to obtain the source sample dataset.

[0053] The generator produces fake samples by inputting random noise, aiming to make the discriminator identify them as real data. The discriminator accurately distinguishes between real and fake samples by inputting real data and the generated fake data. The generator and discriminator compete against each other, aiming to achieve Nash equilibrium. A trained generator is used to synthesize data from the impoverished sample class, reducing the adverse effects of severe sample imbalance on the fault detector.

[0054] The adversarial generation loss of the adversarial generative network uses the pairing matching between generated and real samples as a benchmark reference, and introduces a weighted combination of two zero-center gradient penalties, expressed as:

[0055]

[0056] In the formula, λ1 and λ2 are hyperparameters with a scheduling strategy. In this embodiment, a cosine decaying learning rate scheduling strategy with preheating is used to control the weights of the gradient penalty term. L RpGAN To achieve the relative matching loss, the difference between the discriminator outputs of generated and real samples is used as input, instead of inputting the discriminator outputs of generated and real samples separately. The specific formula is as follows:

[0057]

[0058] Where f is the activation function, and D ψ It is a discriminator, G θ Here, z is the generator, z is the noise distribution, and x is the real data distribution.

[0059] R1 and R2 represent two forms of zero-center gradient penalties, where...

[0060] R1 is a penalty applied to the gradient of the discriminator on the real data, and the formula is:

[0061]

[0062] R² is a penalty applied to the gradient of the discriminator on the generated data, as shown in the formula:

[0063] .

[0064] Step 300: Input the comprehensive sample dataset into the fault diagnosis network for training. Optimize the model by minimizing the fault classification and localization loss until a preset performance threshold is met. Output the current fault diagnosis network as the final submarine cable fault diagnosis model.

[0065] In this embodiment, the fault classification and localization loss L of the fault diagnosis network y It is the weighted sum of the loss values ​​of all prediction modules, defined as:

[0066]

[0067] In the formula, M represents the total number of prediction modules. Let the loss weight corresponding to the m-th prediction module satisfy the following condition: , ; Let represent the fault classification and localization loss value of the m-th prediction module. For the m-th prediction module, its fault classification and localization loss is defined as:

[0068]

[0069] Where K represents the number of prediction boxes output by the prediction module; This represents the intersection-union ratio (IoU) between the k-th predicted bounding box and its corresponding ground truth bounding box. This represents the category prediction probability of the k-th prediction box; This represents the true category label corresponding to the k-th predicted bounding box; and Let represent the adaptive weight coefficients of the positive and negative samples, respectively, and satisfy . ; This represents a weighted classification loss function that considers class imbalance. This represents the distributed focus loss function, used to model the distribution of predicted bounding box positions to enhance the model's ability to constrain positioning accuracy. The distributed focus loss function uses the difference between the discrete probability distribution of the predicted bounding box and the corresponding true position distribution as the optimization objective. Its specific implementation can employ any existing distributed focus loss calculation method; this invention does not limit it.

[0070] The Used to alleviate the imbalance in the number of different categories in fault samples, defined as:

[0071]

[0072] In the formula, N represents the number of samples and C represents the number of fault categories; This represents the predicted probability that the i-th sample belongs to the c-th class; γ is the class balancing factor, used to adjust the weights of positive and negative samples; γ is the focusing parameter, used to enhance the model's ability to focus on difficult-to-classify samples.

[0073] Furthermore, to characterize the spatial offset relationship between the predicted bounding box and the ground truth bounding box, a normalized position offset vector is introduced. Its definition is:

[0074]

[0075] In the formula , This represents the center coordinates of the actual fault area. , This represents the width and height of the actual fault area; , Indicates the center coordinates of the initial prediction box. , H represents the width and height of the initial predicted bounding box; H and W represent the height and width of the feature map or input image, respectively.

[0076] Adaptive weighting coefficients for positive and negative samples and It can be represented as:

[0077]

[0078] In the formula, , These represent the positive and negative sample indices corresponding to the current prediction box, respectively. The weight adjustment function dynamically adjusts the sample weights based on the prediction difficulty, and is defined as follows:

[0079]

[0080] Where a is the upper limit of the weight adjustment function; c is the curvature adjustment parameter, used to control the rate of weight change; B represents the total number of prediction boxes. n represents the spatial offset sorting index of the prediction box relative to the ground truth box, which is obtained by sorting the offsets of all prediction boxes in the same prediction module, and the value range is [0, B].

[0081] The MPAGCNN model comprises two optimization objectives: minimizing the adversarial generation loss and minimizing the fault classification and localization loss. Specifically, minimizing the adversarial generation loss allows the discriminator to achieve optimal ability to distinguish between true and false samples while remaining unable to differentiate between samples originating from the source data and those generated by the generator, thus achieving the goal of obtaining the best generator. Combining the generated samples of different classes with the scarce source samples provides a relatively rich and balanced training sample set for the model. Supervised training is employed to minimize the fault classification and localization loss, thereby improving the model's diagnostic performance. By achieving these two optimization objectives, a diagnostic model with high accuracy and good generalization performance can be obtained even under conditions of scarce samples, providing a new solution to the problems of difficult sample acquisition and severe inter-class data imbalance faced in the field of submarine cable fault detection.

[0082] Step 400: Input the data to be predicted into the final submarine cable fault diagnosis model for reasoning and diagnosis to obtain the fault diagnosis results.

[0083] This invention can enrich the data under the condition of scarce samples by training adversarial generative networks, thereby improving the balance of the distribution between different classes of the dataset. This enables the training of the fault diagnosis network under the condition of scarce fault samples in submarine cable fault diagnosis to achieve better convergence, and improves the generalization performance and fault diagnosis accuracy of the final submarine cable fault diagnosis model.

[0084] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0085] The above are merely preferred embodiments of the present application and are not intended to limit the embodiments of the present application. For those skilled in the art, the embodiments of the present application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of the present application should be included within the protection scope of the embodiments of the present application.

Claims

1. A method for fault diagnosis of submarine cables under data-scarce conditions based on adversarial generation, characterized in that, Includes the following steps: Construct an MPAGCNN model, which includes a generative adversarial network and a fault diagnosis network; Obtain the source sample dataset, input the real data and random noise from the source sample dataset into the adversarial generative network for training, optimize the model by minimizing the adversarial generative loss until the generator and discriminator reach Nash equilibrium, and use the current generator to obtain the generated sample dataset; merge the generated sample dataset with the source sample dataset to obtain the comprehensive sample dataset; The comprehensive sample dataset is input into the fault diagnosis network for training. The model is optimized by minimizing the fault classification and localization loss until a preset performance threshold is met. The current fault diagnosis network is then output as the final submarine cable fault diagnosis model. The data to be predicted is input into the final submarine cable fault diagnosis model for reasoning and diagnosis to obtain fault diagnosis results.

2. The method for diagnosing submarine cable faults under data-scarce conditions based on adversarial generation as described in claim 1, characterized in that, The adversarial generative network includes a generator and a discriminator. The generator includes a linear layer, a batch normalization layer, an activation function, five feature recovery modules 1, 2, 3, 4, and 5 with identical structures, connected in sequence, as well as a transposed convolutional layer and an activation function. The feature recovery module 1 includes a transposed convolutional layer, a batch normalization layer, and an activation function connected in sequence. The discriminator includes a convolutional layer, an activation function, five feature extraction modules A, B, C, D, and E with identical structures, connected in sequence, as well as a linear layer. The feature extraction module A includes a convolutional layer, a batch normalization layer, and an activation function connected in sequence.

3. The method for diagnosing submarine cable faults under data-scarce conditions based on adversarial generation as described in claim 2, characterized in that, The fault diagnosis network includes a feature extractor and a fault detector. The feature extractor includes five identical feature extraction modules 1, 2, 3, 4, and 5 connected in sequence. Feature extraction module 1 includes a convolutional layer, a batch normalization layer, an activation function, and a pooling layer connected in sequence. The fault detector includes three prediction modules respectively set after feature extraction modules 1, 3, and 5. The prediction modules include a fully connected layer and a Softmax layer.

4. The method for diagnosing submarine cable faults under data-scarce conditions based on adversarial generation as described in claim 1, characterized in that, The process of obtaining the source sample dataset includes the following steps: Obtain submarine cable dataset; The fault types in the submarine cable dataset are labeled to obtain the source sample dataset.

5. The method for diagnosing submarine cable faults under data-scarce conditions based on adversarial generation as described in claim 1, characterized in that, The adversarial generation loss is as follows: ； In the formula, λ1 and λ2 are hyperparameters with scheduling strategy, and L RpGAN The relative matching loss is expressed by the formula: ； Where f is the activation function, and D ψ It is a discriminator, G θ Here, z is the generator, x is the noise distribution, and z is the real data distribution. R1 is a penalty applied to the gradient of the discriminator on the real data, and the formula is: ； R² is a penalty applied to the gradient of the discriminator on the generated data, and the formula is: 。 6. The method for diagnosing submarine cable faults under data-scarce conditions based on adversarial generation as described in claim 1, characterized in that, The fault classification and location loss L y As shown below: ； In the formula, M represents the total number of prediction modules. Let the loss weight corresponding to the m-th prediction module satisfy the following condition: , ; The fault classification and localization loss value of the m-th prediction module is defined as: ； Where K represents the number of prediction boxes output by the prediction module; This represents the intersection-union ratio (IoU) between the k-th predicted bounding box and its corresponding ground truth bounding box. This represents the category prediction probability of the k-th prediction box; This represents the true category label corresponding to the k-th predicted bounding box; and Let represent the adaptive weight coefficients of the positive and negative samples, respectively, and satisfy . ; Represents the distributed focus loss function. This represents a weighted classification loss function that considers class imbalance, used to alleviate the imbalance in the number of different classes in faulty samples, and is defined as follows: ； In the formula, N represents the number of samples, and C represents the number of fault categories; This represents the predicted probability that the i-th sample belongs to the c-th class; As a category balance factor; To characterize the spatial offset relationship between the predicted bounding box and the ground truth bounding box, a normalized position offset vector is introduced. Its definition is: ； In the formula , This represents the center coordinates of the actual fault area. , This represents the width and height of the actual fault area; , Indicates the center coordinates of the initial prediction box. , H represents the width and height of the initial predicted bounding box; H and W represent the height and width of the feature map or input image, respectively. Adaptive weighting coefficients for positive and negative samples and , is represented as: ； In the formula, , These represent the positive and negative sample indices corresponding to the current prediction box, respectively. The weight adjustment function dynamically adjusts the sample weights based on the prediction difficulty, and is defined as follows: ； Where a is the upper limit of the weight adjustment function; c is the curvature adjustment parameter, used to control the rate of weight change; B represents the total number of prediction boxes; n represents the spatial offset sorting index of the prediction boxes relative to the ground truth boxes, which is obtained by sorting the offsets of all prediction boxes in the same prediction module, and the value range is [0, B].

Images