Marine ranch power supply system line fault diagnosis method based on MS-2DResNet and ICBAM

Abstract

The invention relates to a marine ranch power supply system line fault diagnosis method based on MS-2DResNet and ICBAM, and belongs to the field of power supply system line fault diagnosis. The fault diagnosis method is mainly divided into two stages, namely, multi-scale feature extraction and attention mechanism feature fusion. Firstly, a preprocessed fault data sample is converted into two-dimensional data to be input into a fault diagnosis model; in the first stage, local and global feature extraction is performed on the converted fault data by using multi-scale convolution, so that the robustness of the model for fault data features of different sizes and scales is improved. And then, entering an attention feature fusion stage, and carrying out weighted fusion on stacked data from different scale channels by utilizing the improved CBAM, so that the model pays more attention to useful fault features and neglects useless features, and further, the fault diagnosis accuracy of the model is improved. The method has the characteristics of high diagnosis accuracy and good anti-noise performance.

Description

Technical Field

[0001] The invention belongs to the field of power supply system line fault diagnosis, and specifically relates to a marine ranch power supply system line fault diagnosis method based on MS-2DResNet and ICBAM. Background Art

[0002] As an emerging way of utilizing marine resources, marine ranching is gradually becoming an important means to solve the problems of global marine resource depletion and ecological environment deterioration. The marine ranch power supply system is an important infrastructure to ensure the normal operation of the ranch. The stability and reliability of its operation directly affect the production efficiency and economic benefits of the ranch. However, due to the particularity of the marine environment, such as salt spray corrosion, tidal changes, seawater erosion, and extreme weather, the reliability and stability of the power supply line have been severely challenged. These environmental factors may not only cause physical damage to the line, but also cause a series of electrical faults, such as short circuit, disconnection, and grounding fault. In addition, the aging of the line itself, design defects, and construction quality problems may also cause faults. Once these faults occur, they will directly affect the normal operation of the marine ranch, and may even cause serious economic losses and environmental problems. Therefore, line fault diagnosis has become a key link in ensuring the stable operation of the marine ranch power supply system.

[0003] Effective diagnosis of line faults in the power supply system of the ocean ranch can not only improve the efficiency of troubleshooting and reduce power outage time, but also provide data support and decision-making basis for the daily maintenance and preventive maintenance of the system. At present, the power supply line fault diagnosis methods are mainly divided into traditional methods and data-driven methods. Traditional fault diagnosis methods mainly include methods based on mathematical models and fault diagnosis methods based on traditional machine learning. Among them, the methods based on mathematical models require a clear understanding of the system dynamic characteristics and operation mechanism of the diagnosis object to establish an accurate mathematical model. However, when faced with a complex power supply system structure, it is difficult to establish an accurate mathematical model, and different fault diagnosis objects require different mathematical models. Traditional machine learning fault diagnosis methods include support vector machines (SVM), decision trees, random forests, K nearest neighbors (KNN), naive Bayes, and artificial neural networks (ANN). Although traditional machine learning has been widely used in fault diagnosis applications, its network structure is shallow and cannot automatically optimize the feature extraction process. It relies on human experience and signal processing technology, and its feature extraction ability is weak. The extracted features may not fully characterize the complexity of fault signals. To solve the above problems, researchers began to use some artificial feature extraction techniques to extract more comprehensive fault features, which mainly include fast Fourier transform, wavelet transform, empirical mode decomposition, etc. After feature extraction, machine learning algorithms are used to perform fault classification. Although the combination of the two further improves the accuracy of fault diagnosis tasks, the artificial feature extraction process relies on the experience of researchers, and the feature extraction process may ignore some subtle fault features and be easily disturbed by noise. When processing fault signals containing noise, noise processing technology is needed to alleviate the influence of noise in the feature extraction process. Therefore, it is crucial to design an end-to-end fault diagnosis method that can directly extract fault features from the original fault signal and deal with the influence of noise.

[0004] In recent years, due to the powerful feature extraction ability of convolutional neural networks, more and more scholars have applied them to fault diagnosis. These methods can achieve correct identification and classification of complex faults by using a large amount of fault data for training in a data-driven manner. At the same time, deep learning models can also automatically extract fault features without manual intervention, greatly improving the efficiency and accuracy of fault diagnosis. For example, Li Yan et al. used the 1DCNN and DS multi-information fusion method to detect the DC bus series arc fault in the photovoltaic system. The author used the powerful feature extraction ability of 1DCNN to perform preliminary feature extraction on the DC current and voltage data, and then fused the extracted fault feature data using DS theory, and finally used certain decision rules to make fault judgments. Huang Wendong et al. proposed a method for distribution network line fault diagnosis using an improved graph convolutional neural network. The author first used BPA software to build an IEEE39-node distribution network simulation model to simulate the collection of fault data. Although the final diagnosis result of the proposed method is good, it is not very generalizable because it needs to build a corresponding graph neural network according to the specific distribution network. In addition, in view of the fact that traditional deep learning methods cannot be well used for multi-signal power system fault diagnosis, the method proposed by PangY et al. divides the fault diagnosis process into three stages, namely, the multidimensional image construction (MIB) stage, the feature decoupling mapping (FDM) stage, and the system fault state classification (SSC) stage. In these three stages, the important information of multiple signals used for fault diagnosis is integrated together to achieve fast and accurate fault identification. Can Ding et al. took advantage of the advantages of CNN for image classification problems and proposed a convolutional neural network (CNN) based on Fourier transform (FFT) and Graham angle field (GAF) for DC transmission line fault identification. The method first uses FFT to perform time-frequency domain analysis on the obtained fault data to detect periodic changes and abnormal frequency changes in the signal, and then uses GAF to further convert the spectrum data obtained by FFT processing into Graham angle matrix, and finally uses the advantage of CNN in identifying image data to identify faults. Similarly, Yabo Cui et al. used the Gramsummation angle field algorithm to convert the collected one-dimensional fault signal of the inverter into a two-dimensional image, and then used the advantages of the CNN variant AlexNet algorithm in image feature extraction to classify the converted fault image data. Huang Y-J et al. proposed a CA-MCNN-based fault diagnosis model by combining the channel attention mechanism with a multi-scale convolutional neural network. This method extracts multi-scale information of the input signal by using the maximum pooling layer and the average pooling layer to increase the feature extraction capability of the convolutional neural network, and can give the model the ability to focus on and prioritize the most informative features through the attention mechanism.Si et al. proposed a fault diagnosis model based on attention mechanism and LSTM network to solve the problem of inverter fault diagnosis under variable working conditions. The model uses two time series signals, current and motor speed, as input data, uses LSTM network for feature extraction, and then uses attention mechanism to suppress useless features, further improving the accuracy of inverter fault diagnosis.

[0005] Through the above analysis, it can be confirmed that the existing research results have made significant progress in the field of fault diagnosis. However, with the increasing complexity of the power supply network topology, the fault types of power supply lines have become more diverse, especially the emergence of problems such as simultaneous faults of multiple lines. The feature differences between some faults are small, which puts higher requirements on the feature extraction ability of fault diagnosis methods. At present, many methods still use single-scale convolution to extract features of fault data. However, since the features of different fault types may have subtle differences, single-scale convolution has limitations in capturing complex fault features and has poor robustness to fault data of different sizes and scales.

[0006] In addition, some studies rely on artificial feature extraction technology to extract fault features and construct feature maps, and then use convolutional neural networks for classification. This method may lead to the loss of some important fault features, and does not fully consider the impact of noise on fault feature extraction. Invalid features are extracted under noise interference, which ultimately affects the accuracy of fault identification, especially in noisy environments. This shows that existing methods are still insufficient in dealing with complex fault types and coping with noise effects, and feature extraction and diagnosis strategies still need to be further optimized. Summary of the invention

[0007] The purpose of the present invention is to provide a line fault diagnosis method for a marine ranch power supply system based on MS-2DResNet and ICBAM in view of the shortcomings of the prior art. The fault diagnosis method is mainly divided into two stages during fault diagnosis, namely, multi-scale residual feature extraction and attention feature fusion stage. The method has the characteristics of high diagnostic accuracy and good noise resistance.

[0008] To achieve the above purpose, the technical solution of the present invention is: a method for diagnosing line faults in a marine ranch power supply system based on MS-2DResNet and ICBAM, comprising a multi-scale feature extraction stage and an attention mechanism feature fusion stage; wherein,

[0009] In the multi-scale feature extraction stage, multi-scale convolution is used to extract local and global features from the two-dimensionally transformed fault data;

[0010] In the attention feature fusion stage, after extracting features in the multi-scale feature extraction stage, the improved CBAM is used to perform weighted fusion on the stacked data from different scale channels to finally complete the fault classification.

[0011] In one embodiment of the present invention, before the fault data is input into the multi-scale feature extraction stage, it needs to be pre-processed and then converted into two-dimensional data.

[0012] In one embodiment of the present invention, the fault data is line fault data obtained by building a power supply system simulation model using MATLAB / Simulink. Considering the actual working environment, the fault data is preprocessed by introducing Gaussian white noise into the fault data, and expanding the fault data by using a random slice sampling method, and performing a splicing operation on the three-phase AC current signal of the fault data to obtain the preprocessed fault data, and dividing it into a training set, a validation set, and a test set according to a predetermined ratio.

[0013] In one embodiment of the present invention, the specific implementation method of expanding fault data using the random slice sampling method is as follows:

[0014] Assume that the length of the simulated fault sample for each fault type is L, where L 1 is the sampling interval of training samples, L 2 is the sampling interval of the test set and the validation set, the length of each fault sample is l, and the sampling starting point is S i (i∈[0,L 1 -l]), and each time S i The value is different. By setting the total number N of samples of each fault type, the fault data can be expanded through random slice sampling.

[0015] In one embodiment of the present invention, in the multi-scale feature extraction stage, a multi-scale two-dimensional residual network feature extraction module is designed by combining multi-scale convolution and residual network, multi-scale convolution is integrated into the residual network, and multi-scale convolution is used to extract features of different scales from the input two-dimensionally converted fault data, so as to simultaneously perceive local details and global features in the data and extract rich feature representations from the fault data.

[0016] In one embodiment of the present invention, the improvement method of CBAM is as follows:

[0017] CBAM consists of two independent sub-modules, namely the channel attention module CAM and the spatial attention module SAM, which perform channel and spatial attention respectively;

[0018] CAM increases the model's attention to important channel features and weakens the impact on unimportant features by assigning different weights to each channel. Specifically, CAM performs global pooling on the input features, using global average pooling and global maximum pooling respectively, and then calculates the weight of each channel through a shared fully connected layer. After being processed by the Sigmoid activation function, these weights are applied to the features of the corresponding channels to form enhanced channel features. SAM aims to focus on important position information in the spatial dimension of the feature map, which complements CAM. SAM generates a spatial attention weight map by combining the average and maximum pooling information of the channel.

[0019] Since the fault data needs to be fused with attention features after the multi-scale feature extraction stage, a multi-stage attention mechanism is introduced so that the channel attention module of CBAM can allocate weights more accurately to more effectively extract and utilize important features. In addition, the introduction of residual connection in CBAM enables the model to retain the direct path of the original input features during the feature extraction process, so that the features processed by the attention module are fused with the original features to prevent excessive loss of feature information. In this way, an improved CBAM is obtained.

[0020] In one embodiment of the present invention, the output formula of the CAM is abbreviated as follows:

[0021] M c =σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0022] Where F represents the input feature map, AvgPool and MaxPool represent global average pooling and global maximum pooling operations respectively, MLP represents multi-layer perceptron, and σ represents the sigmoid activation function;

[0023] The SAM expression is as follows:

[0024] M s (F) = σ(f 7×7 ([AvgPool(F); MaxPool(F)]))

[0025] In the formula, f 7×7 Represents a 7×7 convolution operation, and [AvgPool(F); MaxPool(F)] means concatenating the global average pooling and global maximum pooling results along the channel axis.

[0026] In one embodiment of the present invention, the method is specifically implemented in the following steps:

[0027] S1. Use MATLAB / Simulink simulation software to build a power supply system network model and collect the original three-phase current signal under fault conditions, i.e., original fault data;

[0028] S2. Data preprocessing: The collected original fault data is sliced ​​randomly to increase the data samples, and the obtained fault samples are divided into training set, validation set and test set according to the predetermined ratio;

[0029] S3, constructing a fault diagnosis model based on a multi-scale two-dimensional residual network model and an improved CBAM, and initializing the fault diagnosis model parameters;

[0030] S4, input the training set into the fault diagnosis model for training, calculate the loss error between the predicted value and the true label through forward propagation, and then use BP back propagation to update the network weights and biases; input the validation set data to evaluate the prediction error of the fault diagnosis model, and when the validation error reaches the preset value or reaches the set maximum number of iterations, the training ends and the trained fault diagnosis model is saved;

[0031] S5. Load the trained fault diagnosis model, input the test set for testing, and output the test accuracy, t-SNE visualization graph, and confusion matrix.

[0032] The present invention also provides a marine ranch power supply system line fault diagnosis system based on MS-2DResNet and ICBAM, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0033] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] (1) In order to alleviate the gradient vanishing and gradient exploding problems that are prone to occur in the training process of traditional deep neural networks, the present invention uses the ResNet network to replace the traditional CNN structure. At the same time, parallel multi-scale convolution is introduced into the network to replace the single-scale convolution operation. In this way, the model can use convolution kernels of different sizes to capture a variety of fault feature signals under fault modes or signal changes in different time scales and spatial ranges, thereby improving the accuracy and robustness of fault feature extraction.

[0036] (2) The present invention introduces and improves the CBAM attention mechanism, changing the original channel attention mechanism to a multi-stage channel attention mechanism. Through the average weighted method of multi-stage channel attention, appropriate weights are assigned to the channels of each stage, so that the model can focus more on key channel features while effectively ignoring unimportant features, thereby improving the efficiency of feature extraction and the accuracy of the model.

[0037] (3) To avoid feature attenuation of multi-scale fault features when passing through the attention module, the present invention introduces a residual connection mechanism in CBAM. Through the residual connection, the weighted fault features are fused with the original features, thereby maintaining the stability of information flow and preventing the gradient vanishing problem. Especially in deep networks, the method of the present invention can significantly improve the stability of training and the performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is the power supply system topology diagram and simulation model.

[0039] Figure 2 This is the original signal waveform of the line fault.

[0040] Figure 3 This is the CNN structure diagram.

[0041] Figure 4 It is the MS-2DResNet and ICBAM network model of the present invention.

[0042] Figure 5 Schematic diagram of the data preprocessing process of the present invention.

[0043] Figure 6 This is a fault diagnosis flow chart of the present invention.

[0044] Figure 7 Schematic diagram of noise signal.

[0045] Figure 8 It is the training accuracy and loss function curve under the ideal fault data set.

[0046] Fig. 9 Visualize images for troubleshooting.

[0047] Fig.10 To compare the algorithm training accuracy.

[0048] Fig.11 It is the F1-score evaluation result.

[0049] Fig.12 Confusion matrix for fault diagnosis test of comparison algorithm.

[0050] Fig.13 is the F1 score under different noise conditions. DETAILED DESCRIPTION

[0051] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0052] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0053] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0054] The present invention provides a method for diagnosing line faults in a marine ranch power supply system based on MS-2DResNet and ICBAM, including a multi-scale feature extraction stage and an attention mechanism feature fusion stage; wherein,

[0055] In the multi-scale feature extraction stage, multi-scale convolution is used to extract local and global features from the two-dimensionally transformed fault data;

[0056] In the attention feature fusion stage, after extracting features in the multi-scale feature extraction stage, the improved CBAM is used to perform weighted fusion on the stacked data from different scale channels to finally complete the fault classification.

[0057] The following is the specific implementation process of the present invention.

[0058] 1. Power supply line fault analysis and fault data acquisition

[0059] The power supply lines of the ocean ranch are an important part of its operation, especially for offshore facilities located far from land. Due to its unique geographical environment and operating conditions, these power supply lines may face a variety of faults that affect their stability and efficiency. Common types of faults include line short circuits and circuit breaks. Short circuit faults may cause a sudden drop in the voltage of the power supply system, which may cause shutdown or damage to sensitive equipment. In extreme cases, it may also cause the overall instability of the power supply system, causing widespread power supply interruptions.

[0060] For an AC power supply system, its line faults can be divided into two categories: symmetrical faults and asymmetrical faults, as shown in Table 1. These two types of faults have significant differences in their impact on the power supply lines of the marine ranch. Among them, an asymmetrical fault refers to a short circuit in only one or two phases in a three-phase system, which may or may not involve the ground. This is the most common type of fault, such as single-phase grounding, two-phase short circuit, two-phase ground short circuit, etc. Asymmetrical faults will cause voltage imbalance, introduce negative sequence and zero sequence currents, and affect the performance of motors and other equipment in the marine ranch breeding platform. A symmetrical fault refers to a situation in which the three phases are short-circuited at the same time in a three-phase system, and the magnitude and phase of the fault current of each phase to the ground or to the neutral point are the same. This type of fault is relatively rare in practice, but it is usually the most serious type of fault. This fault will cause a sharp drop in the voltage of the power supply system, which may affect the normal operation of electrical equipment, and the fault point will flow through a current as high as the short-circuit current capacity, which may cause overheating or even damage to cables and equipment. Therefore, a safe and reliable fault diagnosis method is extremely important for identifying line fault problems in the marine ranch power supply system.

[0061] Table 1 Line fault analysis

[0062]

[0063] Since the present invention mainly studies how to perform fault diagnosis when a power supply line fails, according to the above fault analysis, the present invention also considers the situations where the power supply system has symmetrical and asymmetrical faults as well as single-line and multi-line faults. In order to obtain line short-circuit fault data, the present invention builds a 33-node power supply network topology structure, such as Figure 1 As shown on the left, Figure 1 On the right is the Matlab / Simulink simulation model of the power supply system. The three-phase fault module (Three-Phase-Fault) is used to simulate the fault condition of the line. The load and line data of the simulation model refer to the intelligent recovery decision of ship power grid fault using essential multi-objective evolutionary algorithm proposed by Jiang Yanjun et al. and the ship power system fault reconstruction strategy considering reliability constraints proposed by Han Sining et al.

[0064] The present invention simulates the single line 7, the double lines 7 and 12, and the triple lines 12, 15 and 25 when they fail, involving a total of 31 types of failures. Combined with the normal operation state, a total of 32 types of data are simulated. The current waveform obtained by simulation is as follows Figure 2 As shown in the figure. The simulation frequency of the built simulation model is 50Hz, and the sampling frequency of the signal acquisition module is f=10000Hz, so 200 data points can be collected in each signal cycle. The simulation time of each fault type is set to 0.5 seconds, and the total amount of fault data collected is 5000.

[0065] 2. Multi-scale 2D residual network algorithm

[0066] 2.1 Convolutional Neural Networks

[0067] Convolutional neural networks (CNNs) are mainly used to process data with a grid structure. The core idea is to use local connections and weight sharing convolution operations to capture local features of input data while reducing the number of parameters, thereby improving the training efficiency and performance of the model. Figure 3 As shown in the figure. The CNN model structure uses a loss function that measures the difference between the model prediction and the true label as the optimization target of model training. The convolution layer (Filters) slides on the input data and performs a dot product operation to extract the features of the input data. Subsequently, the activation function is generally used to introduce nonlinearity to improve the expression ability of the model. The convolution operation and activation function can be expressed as shown in the following equations (1) and (2):

[0068]

[0069]

[0070] In the formula, A l-1 and Represent the output results of the l-1th and lth network layers respectively; * is the convolution operator, is the convolution kernel, representing a weight matrix; Represents the i-th bias on the l-th network layer.

[0071] After the input data is feature extracted, the pooling layer can be used to extract features again in the local area of ​​the feature map, thereby reducing the size of the feature map while retaining important features, thereby speeding up the training of the model. As the last layer of the CNN network model, the fully connected layer integrates the extracted features into a vector of fixed length, and finally implements classification prediction through fully connected layers such as Softmax. When the Softmax layer is selected as the fully connected layer, the output of the model is shown in formula (3):

[0072] Output(y)=softmax(W·X+B) (3)

[0073] 2.2 Multi-scale Convolutional 2D Residual Network

[0074] For deep learning models, the depth of convolutional neural networks is very important. In theory, the deeper the network, the more dimensional features can be extracted. However, researchers have found through experiments that deep networks are not only difficult to train, but also inefficient and difficult to achieve performance breakthroughs. When the layers are too deep, even the network effect will become worse. The reason is that there are two problems with deep convolutional neural networks: gradient disappearance or explosion and network degradation. In order to solve these problems, the present invention uses a residual network instead of 1DCNN. The so-called residual network is to improve the efficiency of information propagation by adding direct edges to nonlinear convolutional layers. Assume that in a deep network, the present invention expects a nonlinear unit (one or more convolutional layers) f(x; θ) to approximate an objective function f(x). Then the objective function can be split into two parts: the identity function x and the residual function h9x)-x, as shown in formula (4).

[0075] h(x)=x+(h(x)-x) (4)

[0076] Therefore, the original optimization problem can be transformed into: let the nonlinear unit f(x; θ) approximate the residual function h9x)-x, and use f(x; θ)+x to approximate h(x). A typical residual unit consists of multiple cascaded convolutional layers and a cross-layer direct connection, and then the output is obtained after ReLU activation.

[0077] On the other hand, when extracting features from fault data, the traditional convolution layer only uses a single convolution kernel size for each convolution operation. When faced with data with multi-scale features, the model may be insensitive to certain features and it is difficult to capture the global features of the fault data. Multi-scale convolution can make up for the shortcomings of traditional CNN in multi-scale feature extraction.

[0078] After the above analysis, the present invention combines multi-scale convolution and residual network to design a multi-scale two-dimensional residual network feature extraction module, integrates multi-scale convolution into the residual network, and uses multi-scale convolution to extract features of different scales on the input fault data, so as to better perceive the local details and global features in the data at the same time, and extract rich feature representations from the fault data.

[0079] 2.3 Attention Feature Fusion

[0080] After the fault data is extracted by the multi-scale two-dimensional residual network, since different scales of convolution operations will obtain features of different scales, it is necessary to perform feature fusion to adapt to the subsequent residual connection of the deep learning network model and facilitate the network model to fully pay attention to the global and local features of the fault data. As an attention module, CBAM (Convolutional Block Attention Module) can be applied to the feature fusion operation after the multi-scale convolution operation. CBAM contains two independent sub-modules, namely the Channel Attention Module (CAM) and the Spatial Attention Module (Spartial Attention Module, SAM), which perform channel and spatial attention respectively. Among them, CAM enhances the model's attention to important channel features and weakens the influence on unimportant features by assigning different weights to each channel, thereby improving the model's feature extraction ability and performance. Specifically, CAM performs a global pooling operation on the input features, using global average pooling and global maximum pooling, and then calculates the weight of each channel through a shared fully connected layer. After being processed by the Sigmoid activation function, these weights are applied to the features of the corresponding channels to form enhanced channel features. The output formula of CAM can be simplified as follows:

[0081] M c =σ(MLP(AvgPool(F))+MLP(MaxPool(F))) (5)

[0082] Where F represents the input feature map, AvgPool and MaxPool represent global average pooling and maximum pooling operations respectively, MLP represents multi-layer perceptron, and σ represents the sigmoid activation function.

[0083] Unlike CAM, SAM aims to focus on important position information in the spatial dimension of the feature map, which complements CAM. It generates a spatial attention weight map by combining the average and maximum pooling information of the channel, and its expression is as follows:

[0084] M s (F) = σ(f 7×7 ([AvgPool(F);MaxPool(F)])) (6)

[0085] In the formula, f 7×7 Represents a 7×7 convolution operation, and [AvgPool(F); MaxPool(F)] means concatenating the average pooling and maximum pooling results along the channel axis.

[0086] Considering that the fault data needs to be fused with attention features after multi-scale convolution operation, the present invention introduces a multi-stage attention mechanism so that the channel attention module of CBAM can allocate weights more accurately, thereby more effectively extracting and utilizing important features. In addition, the introduction of residual connection in CBAM enables the model to retain the direct path of the original input features during the feature extraction process, so that the features processed by the attention module are fused with the original features, thereby preventing excessive loss of feature information. Then, after the improved CBAM is applied to the multi-scale convolution operation of the proposed model, the channel and spatial attention of CBAM are used to automatically emphasize the important features extracted by the convolution operation and ignore the unimportant features, so that the model can utilize key features more efficiently.

[0087] 3. Fault diagnosis method based on multi-scale two-dimensional residual network and improved CBAM

[0088] When dealing with power supply system line fault problems, fault data may have multi-scale features and contain a lot of noise, while traditional CNNs are not good at fusing features of different scales and are less robust when dealing with noisy data. To address the above problems, this paper proposes a method for power supply line fault diagnosis based on a multi-scale two-dimensional residual network and an improved CBAM (MS-2DResNet and ICBAM). The model structure is as follows Figure 4 As shown in the figure, the line fault data obtained by simulation is first preprocessed and divided into training set, validation set and test set, and then sent to the multi-scale two-dimensional residual network for data reconstruction and conversion into two-dimensional data form. Subsequently, fault features of different scales are extracted through multi-scale parallel convolution, the features of each channel are stacked and integrated, and weighted fusion is performed through the improved CBAM module to finally complete the fault classification.

[0089] 3.1 Data Preprocessing

[0090] Since the present invention uses MATLAB / Simulink to build a power supply system simulation model to obtain line fault data, the collected fault data curve is relatively smooth. However, in an actual working environment, due to noise interference, the collected fault data usually contains noise of different decibels, resulting in data fluctuations of different amplitudes. In order to make the simulation data closer to the actual situation, the present invention introduces Gaussian white noise into the generated fault data, and simply simulates the noisy fault signal to better simulate the situations that may be encountered in the actual power supply system and enhance the robustness of the model. The calculation formula for the signal-to-noise ratio is:

[0091]

[0092] Among them, P signal Represents the original signal power, P noiseRepresents the noise signal power; the smaller the SNR value, the more noise it contains.

[0093] On the other hand, the power supply system is generally designed with protection devices (such as circuit breakers and relays) that can respond quickly and isolate faults. When a power supply system fails, the protection device will act quickly, usually disconnecting the fault circuit within a few milliseconds to a few seconds. This results in very limited samples of available fault data collected, which is difficult to meet the training requirements of deep learning models, and may cause the risk of overfitting the model and reduce generalization ability. Therefore, it is necessary to appropriately expand the fault data to enhance the performance of the model.

[0094] Based on this analysis, the present invention introduces a random slice sampling method to expand the data. Random slice sampling is to randomly select sampling starting points in a certain length of fault samples to generate a certain length of fault samples, and ensure that the sampling starting points are not repeated and are separated by at least 5 data points. Figure 5 The detailed random slice sampling process is shown. Assume that the simulation sample length of each fault type is L, where L 1 is the sampling interval of training samples, L 2 is the sampling interval of the test set and the validation set. The length of each fault sample is l, and the sampling starting point is S i (i∈[0,L 1 -l]), and each time S i By setting the total number of samples of each fault type N, the fault data can be expanded through random slice sampling.

[0095] In addition, since the collected fault data is a three-phase AC current signal, in order to meet the feature extraction requirements of the convolutional neural network, the present invention performs a splicing operation on the three-phase current signal. For example, the tail of the A-phase current is connected to the head of the B-phase current to generate a one-dimensional fault sample with a length of 3l. Table 2 shows the detailed process of the data enhancement algorithm. In order to suppress the occurrence of model overfitting, in the interval L 2 Random slice sampling is performed within the dataset to generate validation and test sets. The validation set is used to evaluate model performance during training to select the best model structure and hyperparameters.

[0096] Table 2 Random slice sampling algorithm

[0097]

[0098] 3.2 Fault diagnosis process based on MS-2DResNet and I-CBAM

[0099] After the above analysis, the power supply system fault diagnosis flow chart based on MS-2DResNet proposed by the present invention is as follows: Figure 6 The specific steps are as follows:

[0100] 1). Use MATLAB / Simulink simulation software to build a power supply system network model and collect the original three-phase current signals under fault conditions.

[0101] 2) Data preprocessing: The collected original fault data is sliced ​​randomly using overlapping operations to increase the data samples, and the obtained fault samples are divided into training sets, validation sets, and test sets according to a certain ratio.

[0102] 3). Initialize the multi-scale two-dimensional residual network model parameters.

[0103] 4) Input the fault sample into the network model for training, calculate the loss error between the predicted value and the true label through forward propagation, and then use BP back propagation to update the network weight and bias. Input the validation set data to evaluate the prediction error of the network model. When the validation error reaches the preset value or reaches the set maximum number of iterations, the training ends and the fault diagnosis model is saved.

[0104] 5). Load the trained network model, input the test set to test and output the test accuracy, t-SNE visualization graph and confusion matrix.

[0105] 4. Experiment and result analysis

[0106] 4.1 Dataset Description

[0107] When a power supply system line fails, in order to minimize the impact of the failure, it is crucial to quickly and accurately locate the cause of the failure and the location of the failure for timely repairing the power supply system and reducing power outage losses. This requires that fault diagnosis can be completed in the shortest possible time so that measures can be taken quickly to avoid equipment damage or system shutdown. In order to ensure that fault diagnosis is implemented within a shorter signal cycle, the present invention adopts the aforementioned data preprocessing method to perform random slicing operations on each fault type to enhance the fault sample. Each sample consists of a half-cycle current signal, and the sliced ​​three-phase current signal is converted into one-dimensional data containing 900 fault data points by splicing the head and tail. In the data sample construction process, each fault type contains 500 training samples, 160 verification samples and 40 test samples.

[0108] In addition, in the actual signal acquisition environment, due to the harsh conditions at sea, the fault data is usually mixed with noise of different decibels. In order to simulate the real acquisition environment and verify the anti-noise performance of the proposed algorithm, the present invention adds different decibels of Gaussian white noise (0-25dB) to the training data set of the fault samples. Figure 7The signal samples after adding 10dB and 20dB noise are shown. It can be seen that the interference of noise makes the collected signal waveform extremely rough, which poses a greater challenge to the anti-interference ability of the network model in the process of fault feature extraction.

[0109] 4.2 Analysis of diagnostic results

[0110] In the face of power supply system line fault problems, the present invention uses the above data set to verify the effectiveness of the proposed fault diagnosis model. The simulations involved in the fault diagnosis process and the operating conditions of the MS-2DResNet and ICBAM models are all CPU i711700, 16GB memory, and a GTX3060 graphics card. The proposed algorithm model is implemented in the Keras environment, and the code is written on the Pycharm platform. After the network model hyperparameters of the proposed MS-2DResNet model are initialized according to Table 3, the training accuracy and loss function curves obtained after 50 rounds of training steps are as shown in the figure. Figure 8 As shown in the figure, the loss function value decreases rapidly during the model training process, while the training accuracy increases rapidly, and the number of iterations tends to be stable at around 20 times, indicating that the proposed algorithm has learned relatively accurate network parameters. When the training iteration reaches 50 times, the model stops training and saves the optimal network parameters.

[0111] Table 3 MS-2DResNet hyperparameter settings

[0112]

[0113]

[0114] In order to verify the training results of the model and more intuitively observe the effectiveness of the proposed algorithm in fault feature extraction, the present invention loaded the trained network model parameters and performed a fault classification test on the test set, and finally obtained a fault classification accuracy of 99.69%. In order to further evaluate the diagnostic performance of the model, in addition to calculating the diagnostic accuracy (ACC) of the model, the present invention also introduced the F1-score evaluation method to measure the comprehensive performance of the multi-classification model. The calculation formula of F1-score is as follows:

[0115]

[0116] Among them, P (Precision) is the precision rate, which indicates the proportion of samples predicted by the model as positive that are actually positive. R (Recall) is the recall rate, which indicates the proportion of samples correctly predicted by the model as positive among the samples actually positive. score It is the harmonic mean of P and R, which is used to measure the relationship between the two.

[0117] The evaluation results of the MS-2DResNet model obtained by the F1-score evaluation method are shown in Table 4. It can be seen from Table 4 that the diagnostic model proposed in the present invention has achieved high values ​​in all three indicators, with the F1-score value reaching 99.69% and the accuracy also reaching 99.69%. The results show the effectiveness of the MS-2DresNet fault diagnosis model.

[0118] Table 4 Diagnostic performance evaluation indicators

[0119]

[0120] After testing the MS-2DResNet and ICBAM fault diagnosis models, the confusion matrix and the images generated using the T-SNE visualization method are shown in Figure 2. Fig. 9 As shown. Fig. 9 It can be observed in (a) that, except for the misdiagnosis of fault types numbered 22 and 24, the overall fault diagnosis effect is good. At the same time, by comparing the visualization images of T-SNE before and after fault diagnosis, it can be found that after being processed by the MS-2DResNet model, each fault type has basically achieved complete separation. Combined with the above multiple evaluation indicators, the effectiveness of the MS-2DResNet model proposed in the present invention in the power supply line fault diagnosis task is verified.

[0121] 4.3 Comparative Analysis

[0122] Although the method proposed in the present invention has achieved good performance in power supply system line fault diagnosis, in order to further verify the effectiveness of the algorithm proposed in the present invention, the present invention compares three traditional fault diagnosis methods, 1DCNN, LSTM, and DNN, and a modern convolutional network ResNet, as well as the convolutional neural network proposed by Zhang S et al. In order to ensure the fairness of the comparison, the present invention uses the proposed MS-2DResNet and ICBAM network model complexity as the baseline, and adjusts the convolutional layer of the comparison algorithm to make each algorithm achieve the optimal performance while having as similar calculation parameters or similar number of hidden layers as possible. At the same time, keep each algorithm with the same number of iterations and batch size. Based on the above considerations, the relevant parameters of the comparison algorithm are set as follows: (1) 1DCNN: 5 convolutional layers, 3 pooling layers, the convolution step size is all 1, the model structure is Conv1D(32,3*1)-MaxPool(2*1)-Conv1D(64,3*1)-Conv1D(128,3*1)-MaxPool(2*1)-Conv1D(256,3*1)-MaxPool(2*1)-Conv1D(512,3*1)-FC(64)-FC(32), and other hyperparameters are consistent with the proposed network model. (2) LSTM: Considering that the excessive number of parameters in the LSTM training process will significantly reduce the training efficiency of the algorithm, the present invention adjusts its structure and sets it to a 5-layer network structure. It contains 3 hidden layers and 2 fully connected layers (FC layers). The number of neurons in the three hidden layers is 64, 128 and 256 respectively, and the number of neurons in the two fully connected layers is 256 and 32 respectively. The activation function is ReLU function. (3): DNN: Different from the above algorithms, the network structure of DNN consists of 4 fully connected layers, 1 Dropout layer and 1 output layer. The other hyperparameters are consistent with other algorithms. (4): The WSCNN-GMP model structure mentioned in the literature

[25] is Dropout(0.1)-BN-Conv(32,128*1)-BN-(Conv(32,3*1)-BN)×2-GMP (global maximum pooling). (5): The ResNet network model structure is similar to the MS-2DResNet network structure. The difference is that the ResNet network model inputs one-dimensional data and does not use multi-scale convolution and feature fusion mechanism.

[0123] After the training of each comparison algorithm is completed, the test set is input for fault classification test. The results show that the test accuracy of the proposed algorithm reaches 99.69%, while the robustness of the comparison algorithm is weak, and the test accuracy fluctuates slightly in each test. In order to ensure the objectivity of the comparison of each algorithm, the present invention records the average value of each comparison algorithm after 10 tests for evaluation. The training accuracy curve of each model and the test accuracy and F1-score evaluation results are shown in Figure 2. Fig.10 and 11 As shown in the figure, it can be seen that the algorithm proposed in the present invention performs best in both training accuracy and test results, indicating that the model can effectively extract fault features and maintain high stability and performance.

[0124] Fig.12 The confusion matrix results of several algorithms in fault diagnosis are shown. Among them, the accuracy of each test of MS-2DResNet and ICBAM fault diagnosis models remains consistent, and the comparison algorithm has large fluctuations in test results, so the confusion matrix with the best performance in 10 tests is selected for display. As can be seen from the figure, in the test set with 40 fault samples in each category, the diagnosis results of the proposed algorithm only have a small number of misdiagnoses in the categories with fault labels 22 and 24, while the rest of the fault categories are accurately distinguished. In contrast, in addition to misdiagnosis in categories 22 and 24, the other comparison methods also have varying degrees of misdiagnosis in other fault categories.

[0125] 4.4 Comparative analysis of noise immunity performance

[0126] In the above experiments, the data used were all noise-free fault samples. The algorithm proposed in the present invention and the five comparative algorithms all showed good fault diagnosis effects. However, in the actual fault diagnosis environment, due to the influence of the harsh marine environment, the collected fault data samples usually carry different degrees of noise. In order to further verify the anti-noise performance of the algorithm of the present invention, the experiment added different decibels (0-25dB) of Gaussian white noise to the fault samples in the training set for anti-noise analysis. Subsequently, the original fault data samples were used for testing, and the obtained F1-Score results are shown as follows: Fig.13 As shown, the fault classification accuracy of each algorithm is shown in Table 5. It can be seen from the chart that under the influence of 0dB Gaussian white noise, the fault classification accuracy and F1_score evaluation value of the MS-2DResNet and ICBAM fault diagnosis algorithms proposed in the present invention are slightly lower than those of the ResNet network model, but show better fault classification effects under the interference of the other four types of noise. Especially when the Gaussian white noise exceeds 20dB, the noise has almost no effect on the fault classification performance of the algorithm proposed in the present invention. It can be seen that the algorithm proposed in the present invention has good noise resistance and has strong practical application potential.

[0127] Table 5 Diagnostic performance of the comparison methods under different noise conditions

[0128]

[0129] 5. Conclusion

[0130] The present invention takes the power supply system line fault as the research object, obtains fault data based on Simulink simulation, and considers the influence of noise on the diagnostic signal, and proposes a power supply line fault diagnosis method based on MS-2DResNet and ICBAM. The method preprocesses the collected three-phase current fault signal into a one-dimensional signal and then automatically converts it into a two-dimensional data input model for model training. Multi-scale fault features are extracted through three parallel convolutional layers of different scales, and then the feature channels of different scales are stacked and input into the improved CBAM module for weighted fusion, thereby improving the fault diagnosis accuracy of the model. The experimental results show that in the fault data set without noise interference, the fault recognition accuracy of the algorithm proposed in the present invention can reach 99.69%. Even when 0-25dB Gaussian white noise is added to the training set, compared with the other five algorithms, the MS-2DResNet and ICBAM fault diagnosis models still maintain a higher fault recognition accuracy and show higher fault recognition stability. Although this study has achieved good diagnostic results, there are not only line faults in the actual power supply line, but also more complex sea noise. Future work will focus on solving the problem of concurrent failures caused by line faults and improving the robustness of the model in complex noise environments to cope with more complex working conditions.

[0131] The present invention also provides a marine ranch power supply system line fault diagnosis system based on MS-2DResNet and ICBAM, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0132] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0133] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0134] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0135] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0136] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0137] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A method for diagnosing line faults in a marine ranch power supply system based on MS-2DResNet and ICBAM, characterized in that: It includes the multi-scale feature extraction stage and the attention mechanism feature fusion stage; in, In the multi-scale feature extraction stage, multi-scale convolution is used to extract local and global features from the two-dimensionally transformed fault data; In the attention feature fusion stage, after extracting features in the multi-scale feature extraction stage, the improved CBAM is used to perform weighted fusion on the stacked data from different scale channels to finally complete the fault classification.

2. According to claim 1, the method for diagnosing line faults in the marine ranch power supply system based on MS-2DResNet and ICBAM is characterized in that: Before the fault data is input into the multi-scale feature extraction stage, it needs to be preprocessed and then converted into two-dimensional data.

3. The method for diagnosing line faults in a marine ranch power supply system based on MS-2DResNet and ICBAM according to claim 2 is characterized in that: The fault data is the line fault data obtained by building a power supply system simulation model using MATLAB / Simulink. Considering the actual working environment, the fault data is preprocessed by introducing Gaussian white noise into the fault data, and expanding the fault data by using the random slice sampling method, and splicing the three-phase AC current signal of the fault data to obtain the preprocessed fault data, which is then divided into training set, validation set and test set according to a predetermined ratio.

4. The method for diagnosing line faults in a marine ranch power supply system based on MS-2DResNet and ICBAM according to claim 3 is characterized in that: The specific implementation method of using random slice sampling method to expand fault data is as follows: Assume that the length of the simulated fault sample of each fault type is L, where L1 is the sampling interval of the training sample, L2 is the sampling interval of the test set and the validation set, the length of each fault sample is l, and the sampling starting point is S i (i∈[0, L1-l]), and each time S is obtained i The value is different. By setting the total number N of samples of each fault type, the fault data can be expanded through random slice sampling.

5. The method for diagnosing line faults in a marine ranch power supply system based on MS-2DResNet and ICBAM according to claim 1 is characterized in that: In the multi-scale feature extraction stage, a multi-scale two-dimensional residual network feature extraction module is designed by combining multi-scale convolution and residual network. Multi-scale convolution is integrated into the residual network, and multi-scale convolution is used to extract features of different scales from the input two-dimensional converted fault data, so as to simultaneously perceive the local details and global features in the data and extract rich feature representations from the fault data.

6. The method for diagnosing line faults in a marine ranch power supply system based on MS-2DResNet and ICBAM according to claim 1 is characterized in that: The improvements of CBAM are as follows: CBAM consists of two independent sub-modules, namely the channel attention module CAM and the spatial attention module SAM, which perform channel and spatial attention respectively; CAM increases the model's attention to important channel features and weakens the impact on unimportant features by assigning different weights to each channel. Specifically, CAM performs global pooling on the input features, using global average pooling and global maximum pooling respectively, and then calculates the weight of each channel through a shared fully connected layer. After being processed by the Sigmoid activation function, these weights are applied to the features of the corresponding channels to form enhanced channel features. SAM aims to focus on important position information in the spatial dimension of the feature map, which complements CAM. SAM generates a spatial attention weight map by combining the average and maximum pooling information of the channel. Since the fault data needs to be fused with attention features after the multi-scale feature extraction stage, a multi-stage attention mechanism is introduced so that the channel attention module of CBAM can allocate weights more accurately to more effectively extract and utilize important features. In addition, the introduction of residual connection in CBAM enables the model to retain the direct path of the original input features during the feature extraction process, so that the features processed by the attention module are fused with the original features to prevent excessive loss of feature information. In this way, an improved CBAM is obtained.

7. The method for diagnosing line faults in a marine ranch power supply system based on MS-2DResNet and ICBAM according to claim 6 is characterized in that: The output formula of CAM is abbreviated as follows: M c =σ(MLP(AvgPool(F))+MLP(MaxPool(F))) Where F represents the input feature map, AvgPool and MaxPool represent global average pooling and global maximum pooling operations respectively, MLP represents multi-layer perceptron, and σ represents the sigmoid activation function; The SAM expression is as follows: M s (F)=σ(f 7×7 ([AvgPool(F);MaxPool(F)])) In the formula, f 7×7 Represents a 7×7 convolution operation, and [AvgPool(F); MaxPool(F)] means concatenating the global average pooling and global maximum pooling results along the channel axis.

8. The method for diagnosing line faults in a marine ranch power supply system based on MS-2DResNet and ICBAM according to claim 1 is characterized in that: The specific implementation steps of this method are as follows: S1. Use MATLAB / Simulink simulation software to build a power supply system network model and collect the original three-phase current signal under fault conditions, i.e., original fault data; S2. Data preprocessing: The collected original fault data is sliced ​​randomly to increase the data samples, and the obtained fault samples are divided into training set, validation set and test set according to the predetermined ratio; S3, constructing a fault diagnosis model based on a multi-scale two-dimensional residual network model and an improved CBAM, and initializing the fault diagnosis model parameters; S4, input the training set into the fault diagnosis model for training, calculate the loss error between the predicted value and the true label through forward propagation, and then use BP back propagation to update the network weights and biases; input the validation set data to evaluate the prediction error of the fault diagnosis model, and when the validation error reaches the preset value or reaches the set maximum number of iterations, the training ends and the trained fault diagnosis model is saved; S5. Load the trained fault diagnosis model, input the test set for testing, and output the test accuracy, t-SNE visualization graph, and confusion matrix.

9. A marine ranch power supply system line fault diagnosis system based on MS-2DResNet and ICBAM, characterized in that: The method comprises a memory, a processor and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps as claimed in any one of claims 1 to 8 can be implemented.

10. A computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, and when the processor executes the computer program instructions, the method steps according to any one of claims 1 to 8 can be implemented.

Images