Capacitance aging state evaluation method based on improved deep residual

Abstract

The application provides a capacitor aging state evaluation method based on an improved deep residual, which comprises the following steps: S1, obtaining a standardized timing aging feature dataset according to a sub-module switching frequency waveform; S2, establishing an improved deep residual convolutional neural network model, which is sequentially connected by an input layer, a first convolutional down-sampling layer, a maximum pooling layer, two first residual basic modules, N second residual basic modules and N first residual basic modules in alternation, an attention mechanism module, a global average pooling layer and a flattening layer; S3, establishing a capacitor aging state evaluation model based on the improved deep residual; S4, training the model by using the standardized timing aging feature dataset D; and S5, inputting a new sub-module switching frequency waveform to judge the aging state of the capacitor. The application improves the ResNet, introduces a hollow convolution and an attention mechanism, and combines the long short-term memory network, thereby improving the accuracy of the capacitor aging state detection.

Description

Technical Field

[0001] This invention relates to the field of power system product reliability, and in particular to a method for assessing capacitor aging status based on improved deep residuals. Background Technology

[0002] In modern electronic devices, modular multilevel converters (MMCs) are widely used in power electronic systems to store and release energy and balance the voltage and current of the power system. However, during long-term operation, the capacitors in the MMC may age due to factors such as temperature, humidity, electrical stress, and frequency in the operating environment, leading to performance degradation or even failure. Therefore, online monitoring and accurate assessment of the aging degree of MMC capacitors are crucial for maintaining the performance and reliability of the equipment.

[0003] Traditional methods for detecting capacitor aging typically involve experimental testing and parameter analysis. One common approach is to conduct offline experiments on the capacitor, measuring parameters such as capacitance, loss factor, and equivalent series resistance, and then comparing the experimental results with theoretical models for analysis and judgment. However, this method requires power outages and capacitor removal, and is time-consuming and labor-intensive. It necessitates removing the capacitor from the circuit and de-energizing it. Furthermore, this method requires specialized experimental equipment and technicians, and is also time-consuming and labor-intensive. Parameter analysis is another commonly used method. It involves online monitoring of parameters such as capacitance, loss factor, and equivalent series resistance, and comparing the measurement results with those of a new capacitor to determine the degree of aging. However, capacitor aging is usually a gradual process, with parameters changing slowly. Parameter analysis only provides static information about the capacitor and cannot accurately capture the dynamic process of capacitor aging.

[0004] To overcome the limitations of traditional methods, several new technologies and approaches have been introduced and researched in recent years to improve the efficiency and accuracy of capacitor aging detection. Among these, artificial intelligence (AI) technology is widely used in power system fault diagnosis and equipment condition monitoring. Deep learning algorithms such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) excel in signal processing and pattern recognition. These algorithms can automatically extract features from input data and perform pattern recognition and prediction by learning from large amounts of data. By introducing AI technology, reliance on human experience can be reduced, enabling automated aging detection and processing of large amounts of data, thus accelerating aging assessment. Furthermore, this method can reduce the cost of aging detection and decrease the need for specialized equipment and personnel. However, existing network models are often designed for broader applicability, and for specific fields and equipment, they suffer from limitations in fully extracting input feature information. This problem currently exists in the assessment of capacitor aging status in modular multilevel converters. Summary of the Invention

[0005] To overcome the problems that existing network models cannot fully extract input feature information and traditional methods cannot achieve automated detection of capacitor aging, this invention provides a capacitor aging state assessment method based on improved deep residuals. By using an improved deep residual convolutional neural network (ResNet) and a long short-term memory network (LSTM), combined with switching frequency waveforms related to capacitor aging, the method can accurately detect and assess the degree of capacitor aging with good accuracy.

[0006] This invention provides a capacitor aging condition assessment method based on improved deep residuals, comprising:

[0007] S1: Collect actual operating data of flexible DC converter valve using multilevel converter under online monitoring, obtain the switching frequency waveform of submodule as capacitor timing aging feature data, preprocess the capacitor timing aging feature data to obtain standardized timing aging feature dataset D and divide it into training set and test set;

[0008] S2: Establish an improved deep residual convolutional neural network model;

[0009] S21: Train the existing ResNet model using the standardized temporal aging feature dataset D features, and determine the parameters of the existing ResNet model;

[0010] S22: Construct an improved ResNet model;

[0011] S221: Improve the basic residual block to obtain the first basic residual module and the second basic residual module; their specific structures are as follows:

[0012] The specific structure of the first residual basic module is as follows: When data is input into the first residual basic module, it is divided into three paths. The first path is a value-equal mapping, that is, the input is not changed and is directly used as the output result. The second path is the output result after the data is input and passes through the second convolutional layer, the BN layer, the first ReLU activation function, and the second convolutional layer BN layer in sequence. The third path is the output result after passing through the dilated convolutional layer. The output results of the three paths are linearly fused and then input into the second ReLU activation function. The output of the second ReLU activation function is used as the output of the first residual basic module.

[0013] The specific structure of the second residual basic module is as follows: When data is input into the second residual basic module, it is also divided into 3 paths. The first path outputs the result after passing through a 1*1 convolution module; the second path outputs the result after the data is input through the second convolutional layer, BN layer, first ReLU activation function, third convolutional downsampling layer, and BN layer in sequence; the third path outputs the result after passing through a dilated convolutional layer; the results of the 3 paths are linearly fused and then input into the second ReLU activation function, and the output of the second ReLU activation function is used as the output of the second residual basic module.

[0014] S222: Introducing an attention mechanism module;

[0015] An attention mechanism module was added between the final residual basic module and the global average pooling layer;

[0016] S223: The improved ResNet model structure is as follows:

[0017] It consists of an input layer, a first convolutional downsampling layer (Conv), a max pooling layer, two first residual basic modules, N second residual basic modules and N first residual basic modules alternately, an attention mechanism module, a global average pooling layer, and a flattening layer connected in sequence.

[0018] S3: Establish a capacitor aging status assessment model based on improved deep residuals;

[0019] The improved deep residual convolutional neural network model obtained in S2, the existing long short-term memory network, the fully connected layer, and the softmax function are connected in sequence, and the probability sequence obtained by the softmax function is used as the output of the capacitor aging state evaluation model based on the improved deep residual.

[0020] S4: Train the capacitor aging state assessment model based on improved deep residual using the standardized time-series aging feature dataset D;

[0021] S5: Input the new submodule switching frequency waveform to determine the aging status of the capacitor.

[0022] Preferably, the improved ResNet model in S223 is a ResNet-18 model. The structure of the improved ResNet-18 model is as follows: an input layer, a first convolutional downsampling layer, a max pooling layer, a first residual basic module, a first residual basic module, a second residual basic module, a first residual basic module, a second residual basic module, a first residual basic module, a second residual basic module, a first residual basic module, a second residual basic module, a first residual basic module, an attention mechanism module, a global average pooling layer, and a Flatten layer connected in sequence.

[0023] Preferably, in the improved ResNet-18 model, the first convolutional downsampling layer is a 7*7 convolution followed by a downsampling step of 2; the convolutional layers in the first residual basic module and the second convolutional layers in the second residual basic module are both 3*3 convolutional layers; the third convolutional downsampling layer in the second residual basic module is a 3*3 convolution followed by a downsampling step of 2; the dilated convolutional layers in the first and second residual basic modules are both 3*3 convolutional layers, and the first and second ReLU activation functions in the first and second residual basic modules are the same.

[0024] Preferably, the preprocessing of the capacitor time-series aging feature data in S1 includes: processing noise, missing data and outliers, and then normalizing the preprocessed capacitor time-series aging feature data to obtain a standardized time-series aging feature dataset D, which is then divided into a training set and a test set in a 4:1 ratio.

[0025] Preferably, the attention mechanism module in S222 adopts a displacement attention mechanism.

[0026] Preferably, S4 includes:

[0027] S41: Input the training set from the standardized time-series aging feature dataset D into the capacitor aging state assessment model based on improved deep residuals, and train the capacitor aging model. The specific process is as follows:

[0028] The standardized temporal aging feature data of the training set capacitor is input into the improved deep residual convolutional neural network model to extract the feature vector in the sequence. The feature vector is then input into the long short-term memory network for sequence modeling. The sequence model is input into the fully connected layer. The last layer of the fully connected layer is mapped to obtain the probability sequence through the softmax function. During this process, the fully connected layer is subjected to dropout operation to obtain the final classification result. The classification result and the true result are compared with the cross-entropy loss function. The model is optimized through backpropagation until the loss function meets the termination condition. Finally, the softmax layer outputs the result.

[0029] S42: Input the test set into the model to evaluate its performance and predictive ability.

[0030] Preferably, the loss function in S41 is the cross-loss entropy function:

[0031]

[0032] In the formula: A is the sample size; B is the number of categories; t ab Indicates whether the a-th sample belongs to the b-th category; y abThe output of sample a belonging to class b is the softmax output, which represents the probability that the network will classify sample a as class b.

[0033] Preferably, in S41, the Dropout operation is performed on the fully connected layer node with a probability of 0.5.

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

[0035] 1. This invention improves deep residual convolutional neural networks by introducing dilated convolution, which can increase the receptive field of the convolutional kernel without introducing additional parameters. This means that a wider range of features can be extracted more effectively while keeping fewer model parameters, which helps to reduce the computational complexity of the model and the risk of overfitting, making the improved ResNet model easier to generalize to new datasets.

[0036] 2. This invention divides the basic residual module in the deep residual convolutional neural network into two types: one uses the original 1*1 convolution module, and the other uses value-equal mapping. The 1*1 convolution module is used to reduce the dimension of features, while the value-equal mapping is used to extract deep features. The combination of the two further improves the feature extraction capability.

[0037] 3. This invention also introduces an attention mechanism into the deep residual convolutional neural network, which can mine deep information of features and perform feature recombination based on channel optimization.

[0038] 4. This invention combines an improved deep residual convolutional neural network (ResNet) with a long short-term memory network (LSTM), fully leveraging ResNet's excellent feature extraction capabilities in image processing and LSTM's advantages in sequence modeling. This allows the model to better capture key features in time-series data, making it more suitable for capacitor aging state assessment tasks and improving detection accuracy and performance. Through this combination, the model can more comprehensively understand the information in time-series data, more accurately predict the aging state of capacitors, and improve detection accuracy.

[0039] 5. This invention only requires the switching frequency waveform of the sub-module in the modular multilevel converter to diagnose the aging state of the capacitor. The dilated convolution can better capture the details and contextual information in the image. The sub-module switching frequency waveform is easy to obtain. Therefore, this invention does not require additional sensors or measurement equipment, which makes this invention both improve the feature extraction capability and easy to implement. Attached Figure Description

[0040] Figure 1 This is a flowchart of the capacitor aging state assessment method based on improved deep residuals according to the present invention;

[0041] Figure 2 This is a schematic diagram of the existing ResNet-18 architecture;

[0042] Figure 3 This is a schematic diagram of the improved ResNet-18 structure in an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram of the ShuffleAttention principle introduced in the ResNet model in this invention;

[0044] Figure 5 This is a schematic diagram of the feature extraction using the LSTM of the present invention;

[0045] Figure 6 This is a schematic diagram of the capacitor aging state assessment model based on improved deep residuals according to the present invention;

[0046] Figure 7 This is a flowchart of the training process for the capacitor aging state assessment model based on improved deep residuals in this invention. Detailed Implementation

[0047] To better understand the technical solution of the present invention, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The same reference numerals in the drawings indicate elements with the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0048] This invention discloses a capacitor aging condition assessment method based on improved deep residuals, particularly for modular multilevel converter sub-modules. The specific implementation steps are as follows: Figure 1 As shown:

[0049] S1: Collect actual operating data of flexible DC converter valves using multilevel converters (MMC) under online monitoring, obtain the switching frequency waveform of submodules as capacitor timing aging feature data, preprocess the capacitor timing aging feature data to obtain standardized timing aging feature dataset D, and divide it into training set and test set.

[0050] The preprocessing of capacitor time-series aging feature data includes handling noise, missing data, and outliers. Then, the preprocessed capacitor time-series aging feature data is normalized to obtain a standardized time-series aging feature dataset D, which is then divided into training and test sets in a 4:1 ratio.

[0051] In this embodiment, the standardized time-series aging feature dataset D contains a total of 1800 data samples, including 1000 normal samples of submodule capacitors and 200 aging samples of submodule capacitors with capacitance values ​​decreasing by 5%, 10%, 15%, and 20%. Each sample contains 0.025s of switching status information of the submodule. The training set and test set are divided in a 4:1 ratio, so the training set has 1440 samples and the test set has 360 samples.

[0052] In this embodiment, the capacitor timing aging characteristic data is selected as the submodule switching frequency waveform. In fact, the capacitor voltage or bridge arm current waveforms can also be used as timing aging characteristic data.

[0053] S2: Build an improved deep residual convolutional neural network (ResNet) model.

[0054] S21: Train the existing ResNet model using the standardized time-series aging feature dataset D features, and determine the parameters of the existing ResNet model.

[0055] S22: Construct an improved ResNet model.

[0056] A traditional ResNet model consists of an input layer, a first convolutional downsampling layer, a max pooling layer, a residual module, and a global average pooling layer connected sequentially. The residual module contains multiple basic residual modules; based on the number of these basic residual blocks, commonly used ResNet models include ResNet-18 and ResNet-50. In this invention, the convolutional downsampling layer refers to a downsampling layer connected after the convolution.

[0057] The improved ResNet model is based on the traditional ResNet model. The improvements mainly include: improving the basic residual module in the residual module of the existing ResNet model, and adding an attention mechanism module between the final residual module and the global average pooling layer.

[0058] S221: Improve the residual basic block to obtain the first residual basic module and the second residual basic module.

[0059] The improvements to the existing residual basic blocks in this invention are divided into two types, referred to as the first residual basic module and the second residual basic module, respectively. Their specific structures are as follows:

[0060] The specific structure of the first residual basic module is as follows: When data is input into the first residual basic module, it is divided into three paths. The first path is a value-equal mapping, that is, the input is not changed and is directly used as the output result. The second path is where the input data passes through the second convolutional layer (Conv), the BN layer (Batch Normalization), the first ReLU activation function, the second convolutional layer (Conv), and the BN layer in sequence before outputting the result. The third path input passes through a dilated convolutional layer before outputting the result. The results of the three paths are linearly fused and then input into the second ReLU activation function. The output of the second ReLU activation function is used as the output of the first residual basic module.

[0061] The specific structure of the second residual basic module is as follows: When the data is input into the second residual basic module, it is also divided into 3 paths. The first path is the output result after passing through a 1*1 convolution module; the second path is the output result after the data is input and passes through the second convolutional layer (Conv), BN layer, first ReLU activation function, third convolutional downsampling layer, and BN layer in sequence; the third path is the output result after passing through a dilated convolutional layer; the output results of the 3 paths are linearly fused and then input into the second ReLU activation function, and the output of the second ReLU activation function is used as the output of the second residual basic module.

[0062] It can be seen that existing residual basic modules are simply divided into two paths, without needing dilated convolutional layers, and the first path always uses a 1*1 convolutional module. This application, however, divides the residual module into two types: one using the original 1*1 convolutional module, and the other using value-isomapping. This is because the 1*1 convolutional module is used for feature dimensionality reduction, while value-isomapping is used to extract depth features. Traditional convolution operations use a fixed kernel size, which to some extent limits the size of the receptive field. Dilated convolution introduces a fixed interval between the sampling points of the convolution kernel, thus expanding the size of the receptive field. By increasing the dilation rate of the convolution kernel, the receptive field can be expanded. Dilated convolution can increase the receptive field without introducing additional parameters, helping to reduce the computational complexity of the model and reduce the risk of overfitting. Dilated convolution can better capture details and contextual information in the image. For this invention, using the sub-module switching frequency waveform as capacitance time-series aging feature data can improve feature extraction capabilities. By introducing dilated convolutions, ResNet can better handle long-range dependencies in the data, thereby improving the network's expressive power and performance. Therefore, improving the residual basic module can enhance feature extraction capabilities when using the sub-module switching frequency waveform as capacitor time-series aging feature data, further improving the overall model's performance metrics.

[0063] S222: Introducing the attention mechanism module.

[0064] The attention mechanism module can mine deep information from features and perform channel-optimized feature reorganization. In this embodiment, the attention mechanism module adopts the Shuffle Attention mechanism, which introduces interactive information into the channel dimension of the feature map. By rearranging and exchanging channels, the interactivity between channels in the feature map is enhanced, thereby improving the model's expressive power and performance. In this embodiment, Shuffle Attention achieves channel reorganization through channel grouping, channel splitting, and feature fusion. The principle of Shuffle Attention is as follows: Figure 4 As shown, the specific implementation is as follows:

[0065] The feature map is input from the residual module, which includes at least three parameters: number of channels c, height, and width, and may also include the number of batches.

[0066] Channel grouping: Divide the c input channels into g groups (g must be a divisor of c). For example, if c = 64, it can be divided into 8 groups, each containing 8 channels.

[0067] Channel splitting: For each group, the channels within the group are rearranged. This process can be done randomly or using specific rules to rearrange the channel order to increase interaction between channels.

[0068] Feature fusion: The feature maps from the shuffled channels are concatenated or added element-wise to fuse the interaction information between the groups. This process can be achieved using element-wise multiplication or other attention fusion mechanisms.

[0069] S223: Improved ResNet model structure.

[0070] Based on S221 and S222, the improved ResNet model structure is as follows:

[0071] It consists of an input layer, a first convolutional downsampling layer (Conv), a max pooling layer, two first residual basic modules, N second residual basic modules and N first residual basic modules alternately, an attention mechanism module, a global average pooling layer, and a flattening layer connected in sequence.

[0072] The following section will provide a detailed explanation using improvements to the ResNet-18 model as an example. The existing ResNet-18 model is as follows: Figure 2 As shown, the residual module comprises 8 basic residual modules, therefore the structure of the improved ResNet-18 model is as follows: Figure 3As shown, the layers are connected in sequence as follows: input layer, first convolutional downsampling layer, max pooling layer, first residual basic module, first residual basic module, second residual basic module, first residual basic module, second residual basic module, first residual basic module, second residual basic module, first residual basic module, attention mechanism module, global average pooling layer, and Flatten layer.

[0073] In this implementation, the first convolutional downsampling layer is a 7x7 convolution followed by a downsampling step of 2. The convolutional layers in both the first and second residual basic modules are 3x3 convolutional layers. The third convolutional downsampling layer in the second residual basic module is also a 3x3 convolution followed by a downsampling step of 2. The dilated convolutional layers in both the first and second residual basic modules are also 3x3 convolutional layers. The first and second ReLU activation functions in both modules can be the same or different; this embodiment uses both. The Flatten layer flattens the multidimensional data output from the convolutional layers into a one-dimensional vector for use as input to the subsequent Long Short-Term Memory (LSTM) network.

[0074] Taking a 3x224x224 input as an example, the feature extraction process using the improved ResNet-18 model is as follows: After inputting the data, it passes through the first convolutional downsampling layer to obtain a 64x... The data is initially 112x112 pixels, then passes through a max pooling layer to obtain 64x56x56 pixels. Next, it passes through two first residual basic modules. Since the first residual basic module only contains convolutions, the data size remains unchanged at 64x56x56. Then, it passes through a second residual basic module, which contains a third convolutional downsampling layer, thus outputting 128x28x28 pixels. This process continues: after passing through the first residual basic module (data size unchanged), the second residual module outputs 256x14x14 pixels, then the second residual basic module outputs 512x7x7 pixels, then the attention mechanism module (which does not change the data size), and finally, a global average pooling layer outputs 512x1x1 pixels. Finally, a Flatten layer transforms it into 512 one-dimensional pixels.

[0075] S3: Establish a capacitor aging state assessment model based on improved deep residuals.

[0076] The structure of the capacitor aging state assessment model based on the improved deep residual (referred to as the "improved model") is as follows: Figure 6 As shown, specifically:

[0077] The improved deep residual convolutional neural network (ResNet) model obtained in S2, the existing long short-term memory network (LSTM), the fully connected layer, and the softmax function are connected in sequence, and the probability sequence obtained by the softmax function is used as the output of the capacitor aging state assessment model based on the improved deep residual.

[0078] This invention combines an improved ResNet and LSTM, using the feature sequences extracted by the improved ResNet as the input sequence for the LSTM. The improved ResNet is used for feature extraction, while the LSTM handles memorization. This fully leverages ResNet's excellent feature extraction capabilities in image processing and LSTM's advantages in sequence modeling, enabling the model to better capture the dynamic changes of aging states in time-series data. First, a convolutional network is used to extract the "features" of the image, achieving dimensionality reduction. The LSTM's feature extraction is as follows... Figure 5 As shown, LSTM solves the problem that RNNs cannot handle long-range dependencies, and can also solve common problems in neural networks such as gradient explosion or gradient vanishing. The improved ResNet model in this invention, due to the introduction of dilated convolutions and deep feature mining and channel-optimized feature reorganization, can better handle long-range dependencies in data compared to the traditional ResNet. Combined with LSTM, which is suitable for handling long-range dependencies, it can better utilize these long-range dependencies, thereby improving the accuracy of assessing capacitor aging conditions.

[0079] S4: Train the capacitor aging state assessment model based on improved deep residuals using the standardized time-series aging feature dataset D.

[0080] S41: Input the training set from the standardized time-series aging feature dataset D into the capacitor aging state assessment model based on improved deep residuals to train the capacitor aging model, such as... Figure 7 As shown, the specific process is as follows:

[0081] The standardized temporal aging feature data of the capacitor in the training set is input into an improved deep residual convolutional neural network (ResNet) model to extract feature vectors from the sequence. The feature vectors are then input into a long short-term memory network (LSTM) for sequence modeling. The sequence modeling is input into a fully connected layer. The last layer of the fully connected layer is mapped to obtain a probability sequence through a softmax function. During this process, a dropout operation is performed on the fully connected layer (randomly discarding data) to obtain the final classification result. The classification result and the true result are compared with the cross-entropy loss function (the model is optimized through backpropagation) until the loss function meets the termination condition (tends to stationarity). Finally, the softmax layer outputs the result.

[0082] In this embodiment, a Dropout operation is performed on the node with a probability of 0.5, and the loss function is calculated. If the loss function does not meet the termination condition, stochastic gradient descent (SGD) is used to train the convolution kernel and bias matrix of the capacitor aging model, update its parameters, and continue the above calculation until the loss function meets the termination condition, thus obtaining the trained model.

[0083] The loss function used is the cross-loss entropy function:

[0084]

[0085] In the formula: A is the sample size; B is the number of categories; t ab Indicates whether the a-th sample belongs to the b-th category; y ab The output of sample a belonging to class b is the softmax output, which represents the probability that the network will classify sample a as class b.

[0086] In this embodiment, the aging state of the capacitor is directly divided into two categories: unaged capacitor and aged capacitor; therefore, the number of categories B is 2.

[0087] S42: Input the test set into the model to evaluate its performance and predictive ability.

[0088] The test set is input into the trained capacitor state model for testing. Based on the actual labeled capacitor aging data, the model's prediction results are compared and evaluated with the true values. When the predetermined index is reached, the model is considered to be trained successfully; otherwise, the model training continues.

[0089] Table 1 Comparison of performance metrics of multiple models

[0090] Model Name Accuracy / % Recall rate / % Specificity / % RepVGG 93.0 93.3 93.3 ResNet18 94.0 94.1 94.1 EffficientNet 93.0 94.1 94.1 Mobilenetv3 95.0 95.3 95.3 Improved model 95.5 95.6 95.6

[0091] Furthermore, to demonstrate the effectiveness of this invention, comparative training of multiple models was conducted under the same experimental environment. These models included RepVGG, ResNet18, EffficientNet, MobileNetv3, and the improved model. Corresponding model performance metrics were calculated, with precision, recall, and specificity selected to evaluate the model's performance and predictive ability. As shown in Table 1, the improved model achieved a precision of 95.5%, while also exhibiting the highest detection performance in both recall and specificity.

[0092] S5: Input the new submodule switching frequency waveform to determine the aging status of the capacitor.

[0093] The switching frequency waveform of the new submodule is used as the capacitor aging characteristic data and input into the capacitor aging state assessment model based on the improved deep residual, and the result of judging the capacitor aging state is output.

[0094] This invention only requires the switching frequency waveform data of the submodule to determine the aging state of the capacitor, eliminating the need for additional sensors or measuring equipment. Moreover, its accuracy is higher than existing methods, making the diagnosis of capacitor aging state more convenient and accurate, and reducing the complexity of capacitor aging state assessment.

[0095] Finally, it should be noted that the above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A capacitor aging state assessment method based on improved deep residuals, characterized in that: It includes the following steps: S1: Collect actual operating data of flexible DC converter valve using multilevel converter under online monitoring, obtain the switching frequency waveform of submodule as capacitor timing aging feature data, preprocess the capacitor timing aging feature data to obtain standardized timing aging feature dataset D and divide it into training set and test set; S2: Establish an improved deep residual convolutional neural network model; S21: Train the existing ResNet model using the standardized temporal aging feature dataset D features, and determine the parameters of the existing ResNet model; S22: Construct an improved ResNet model; S221: Improve the basic residual block to obtain the first basic residual module and the second basic residual module; their specific structures are as follows: The specific structure of the first residual basic module is as follows: When data is input into the first residual basic module, it is divided into 3 paths. The first path is a value-equal mapping, that is, the input is not changed and is directly used as the output result. The second path is the output result after the data is input and passes through the second convolutional layer, the BN layer, the first ReLU activation function, and the second convolutional layer BN layer in sequence. The third path is the output result after the input passes through the dilated convolutional layer. The results of the three outputs are linearly fused and then input into the second ReLU activation function. The output of the second ReLU activation function is used as the output of the first residual basic module. The specific structure of the second residual basic module is as follows: When the data is input into the second residual basic module, it is also divided into 3 paths. The first path is the output result after passing through a 1*1 convolution module; the second path is the output result after the data is input, passing through the second convolution layer, BN layer, first ReLU activation function, third convolution downsampling layer, and BN layer in sequence; the third path is the output result after passing through a dilated convolution layer. The results of the three outputs are linearly fused and then input into the second ReLU activation function. The output of the second ReLU activation function is used as the output of the second residual basic module. S222: Introducing an attention mechanism module; An attention mechanism module was added between the final residual basic module and the global average pooling layer; S223: The improved ResNet model structure is as follows: It consists of an input layer, a first convolutional downsampling layer (Conv), a max pooling layer, two first residual basic modules, N second residual basic modules and N first residual basic modules alternately, an attention mechanism module, a global average pooling layer, and a flattening layer connected in sequence. S3: Establish a capacitor aging state assessment model based on improved deep residuals; The improved deep residual convolutional neural network model obtained in S2, the existing long short-term memory network, the fully connected layer, and the softmax function are connected in sequence, and the probability sequence obtained by the softmax function is used as the output of the capacitor aging state evaluation model based on the improved deep residual. S4: Train the capacitor aging state assessment model based on improved deep residual using the standardized time-series aging feature dataset D; S5: Input the new submodule switching frequency waveform to determine the aging status of the capacitor.

2. The capacitor aging status assessment method based on improved deep residual as described in claim 1, characterized in that: The improved ResNet model in S223 is the ResNet-18 model. The structure of the improved ResNet-18 model is as follows: input layer, first convolutional downsampling layer, max pooling layer, first residual basic module, first residual basic module, second residual basic module, first residual basic module, second residual basic module, first residual basic module, second residual basic module, first residual basic module, attention mechanism module, global average pooling layer, and Flatten layer are connected in sequence.

3. The capacitor aging status assessment method based on improved deep residual as described in claim 2, characterized in that: In the improved ResNet-18 model, the first convolutional downsampling layer is a 7*7 convolution followed by a downsampling step of 2. The convolutional layers in the first residual basic module and the second convolutional layer in the second residual basic module are both 3*3 convolutional layers. The third convolutional downsampling layer in the second residual basic module is a 3*3 convolution followed by a downsampling step of 2. The dilated convolutional layers in the first and second residual basic modules are both 3*3 convolutional layers, and the first and second ReLU activation functions in the first and second residual basic modules are the same.

4. The capacitor aging status assessment method based on improved deep residual as described in claim 1, characterized in that: The preprocessing of capacitor time-series aging feature data in S1 includes: processing noise, missing data and outliers, and then normalizing the preprocessed capacitor time-series aging feature data to obtain a standardized time-series aging feature dataset D, which is then divided into training set and test set in a 4:1 ratio.

5. The capacitor aging status assessment method based on improved deep residual as described in claim 1, characterized in that: The attention mechanism module in S222 adopts a displacement attention mechanism.

6. The capacitor aging status assessment method based on improved deep residual as described in claim 1, characterized in that: S4 includes: S41: Input the training set from the standardized time-series aging feature dataset D into the capacitor aging state assessment model based on improved deep residuals, and train the capacitor aging model. The specific process is as follows: The standardized temporal aging feature data of the training set capacitor is input into the improved deep residual convolutional neural network model to extract the feature vector in the sequence. The feature vector is then input into the long short-term memory network for sequence modeling. The sequence model is input into the fully connected layer. The last layer of the fully connected layer is mapped to obtain the probability sequence through the softmax function. During this process, the fully connected layer is subjected to dropout operation to obtain the final classification result. The classification result and the true result are compared with the cross-entropy loss function. The model is optimized through backpropagation until the loss function meets the termination condition. Finally, the softmax layer outputs the result. S42: Input the test set into the model to evaluate its performance and predictive ability.

7. The capacitor aging status assessment method based on improved deep residual as described in claim 6, characterized in that: In S41, the loss function is the cross-loss entropy function E, which is calculated as follows: ； In the formula: A is the sample size; B is the number of categories; t ab Indicates whether the a-th sample belongs to the b-th category; y ab The output of sample a belonging to class b is the softmax output, which represents the probability that the network will classify sample a as class b.

8. The capacitor aging status assessment method based on improved deep residual as described in claim 6, characterized in that: In S41, Dropout operation is performed on fully connected layer nodes with a probability of 0.5.

Images