Direct current support capacitor online state monitoring method and device based on PINN

Abstract

The embodiment of the application discloses a DC support capacitor online state monitoring method and device based on PINN, the method comprises the following steps: obtaining a plurality of groups of charge-discharge data of voltage change with time under different capacitance values of DC support capacitor, and determining a training set and a verification set; using the AdamW algorithm to train a preset PINN model according to the training set, adaptively updating the loss function weight and the network weight until a first preset iteration number is reached, and obtaining the model parameters of the trained PINN model; using the L-BFGS algorithm to update the model parameters, and verifying the final PINN model according to the verification set; online obtaining pre-charge data of the DC support capacitor to be measured, and applying the final PINN model to monitor the state of the DC support capacitor to be measured. Through the above method, the embodiment of the application can effectively reduce the influence of the measurement signal-to-noise ratio on the prediction result, and is suitable for capacitor value prediction in a wide range of capacitor intervals under the same topology.

Description

Technical Field

[0001] This invention relates to the field of electronic component technology, specifically to a method and apparatus for online status monitoring of DC support capacitors based on PINN. Background Technology

[0002] The onboard traction converter system, as the core equipment for the train's kinetic energy conversion, is the train's powerful "power source heart." Any malfunction in the converter severely restricts the normal operation of the entire train. The DC-link capacitor, a key component in the intermediate DC link of the converter used for power conversion, balances the instantaneous power difference between the input source and the output load. Therefore, with current technology, it maximizes the absorption of high-pulse current and buffering of high-pulse voltage in the intermediate DC link, effectively filtering ripple. Currently, traction converter systems often use metallized polypropylene film capacitors (MFCs) with low impedance, large capacitance, non-polarity, strong overvoltage and reverse pulse voltage withstand capabilities, and long service life. Statistics show that DC-link capacitor failures account for the largest proportion of onboard converter system faults. Therefore, online monitoring of the degradation characteristics of traction converter capacitors is reasonable and has considerable scientific and practical value for train maintenance.

[0003] Currently, based on the importance of prior physical knowledge in identification methods, existing online identification methods can be broadly categorized into three types: physics-based formula-driven identification methods based on capacitance and impedance physical models, pure data-driven identification methods based on intelligent black-box algorithms, and hybrid identification methods that combine physical information with data-driven approaches. The application of pure physics-based model-driven methods stems from years of research accumulation in various engineering fields. As long as accurate pre-parameters are obtained and the corresponding physical model is correctly solved, sufficiently accurate prediction results can be achieved. However, in practical applications, many traditional physics-driven methods are based on partial differential equations, often encountering mathematical bottlenecks in solving high-dimensional problems, solving complex geometric boundaries in conjunction with actual situations, and solving inverse problems, leading to slow progress. The basic idea of ​​pure data-driven identification methods based on intelligent black-box algorithms is to train a supervised optimal machine learning model to establish a latent function mapping between the acquired input data and the output data to be monitored, thereby bypassing the physical model to indirectly estimate the parameters to be measured. However, this pure data-driven identification method based on the intelligent black box algorithm relies too heavily on existing data. The data acquisition accuracy, quantity, and frequency in power electronic equipment are all greatly limited, which may make it difficult to provide high-quality datasets for machine learning tools to play a good role. This will impact the basic requirement of consistent distribution of training and test data in machine learning methods, resulting in reduced generalization ability of the model, weakened robustness to external interference, and difficulty in meeting actual engineering needs.

[0004] The approach of using data-driven, physically-constrained machine learning (PIML) based on physical knowledge has gained widespread acceptance. Among these, the Physics-Informed Neural Network (PINN), as a type of supervised data training model, benefits from the automatic differentiation technique of neural networks. It effectively embeds physical information into the model training loss calculation during the training process, greatly improving the algorithm's efficiency and practicality. For example, the PINN algorithm solves the high-frequency (HF) modeling problem of induction motors, using the impedance of each phase of the motor to deduce the values ​​of each circuit component. The PINN algorithm has also been applied to synchronous generator modeling, enabling the model to accurately represent the magnetic saturation characteristics of the synchronous generator at various stages, even with small datasets. However, current technology has not yet applied the PINN algorithm to the online status monitoring of the DC support capacitor in traction converters. Summary of the Invention

[0005] In view of the above problems, embodiments of the present invention provide a method and apparatus for online status monitoring of DC support capacitors based on PINN, which overcomes or at least partially solves the above problems.

[0006] According to one aspect of the present invention, a method for online state monitoring of a DC support capacitor based on PINN is provided. The method includes: acquiring multiple sets of charging and discharging data of voltage changes over time for different capacitance values ​​of the DC support capacitor, and determining a training set and a validation set; using the AdamW algorithm to train a preset PINN model based on the training set, adaptively updating the loss function weights and linear network weights of each component until a first preset number of iterations is reached, and acquiring the model parameters of the trained PINN model, wherein the PINN model is a feedforward fully connected deep neural network, and its model parameters are adjusted according to the solution of the partial differential equation constructed based on the pre-charging model of the DC support capacitor; using the L-BFGS algorithm to update the model parameters of the trained PINN model, and validating it according to the validation set to obtain the final PINN model; acquiring the pre-charging data of the DC support capacitor under test online, and applying the final PINN model to monitor the state of the DC support capacitor under test based on the pre-charging data.

[0007] Optionally, acquiring charging and discharging data of voltage variations over time under different values ​​of the DC support capacitor, and determining the training set and validation set, includes: simulating the aging process of the DC support capacitor by connecting different numbers of thin-film capacitors of the same capacity in parallel outside the DC support capacitor; acquiring ordered voltage arrays and ordered time arrays with the same sampling interval under different values ​​of the DC support capacitor as charging and discharging data of voltage variations over time; replacing the DC support capacitor and acquiring ordered voltage arrays and ordered time arrays with the same sampling interval under different values ​​of the DC support capacitor; and determining the training set and validation set based on the ordered voltage arrays and ordered time arrays with different values ​​of the DC support capacitor.

[0008] Optionally, the step of training the preset PINN model based on the training set until a first preset number of iterations is reached, and obtaining the model parameters of the trained PINN model, includes: inputting sample data from the training set into the preset PINN model, obtaining the output results and the current network structure; calculating the total loss of the PINN model, and updating the model weights using the AdamW algorithm until the first preset number of iterations is reached.

[0009] Optionally, the step of calculating the total loss of the preset PINN model and updating the model weights using the AdamW algorithm includes: constructing an adaptive loss function based on the partial differential equation to obtain the total loss of the PINN model; and simultaneously updating the adaptive weights of each loss component in the adaptive loss function and the model weights of the PINN model using the AdamW algorithm.

[0010] Optionally, the step of constructing an adaptive loss function based on the partial differential equation to obtain the total loss of the PINN model includes: obtaining the data loss by calculating the mean square error between the output value and the true value of the feedforward fully connected deep neural network; inputting random numbers in an optional domain into the network to obtain the initial loss and boundary loss; constructing a partial differential equation with the output value and input value of the feedforward fully connected deep neural network to obtain the partial differential equation residual loss; and adaptively weighting and summing the data loss, the initial loss, the boundary loss, and the partial differential equation residual loss to obtain the total loss of the PINN model.

[0011] Optionally, after calculating the total loss of the preset PINN model and updating the model weights using the AdamW algorithm, the method further includes: determining whether the number of iterations has reached a second preset number of iterations, where the second preset number of iterations is less than the first preset number of iterations; if the number of iterations reaches the second preset number of iterations and the data loss is less than a preset threshold, applying a conditional generative adversarial network to expand the training set; updating the total loss of the PINN model to a weighted sum of data loss, initial loss, boundary loss, partial differential residual loss, and generated data loss, wherein the generated data loss is the mean square error between the output value and the true value of the feedforward fully connected deep neural network based on the generated data.

[0012] Optionally, the application of a conditional generative adversarial network to expand the training set includes: grouping the sample data in the training set; obtaining a calculated voltage vector based on the experimental voltage vector; using a random noise vector following a probability distribution as input and passing it through a generator in the conditional generative adversarial network to randomly generate a noise vector; adding the calculated voltage vector and the noise vector to obtain a negative sample; using the experimental voltage vector as a real sample and the negative sample as a fake sample, applying a discriminator in the conditional generative adversarial network to determine a reliable result vector; concatenating all the reliable result vectors obtained by applying the conditional generative adversarial network and their corresponding time vectors into two corresponding matrices and adding them to the training set to obtain the expanded training set.

[0013] Based on the same inventive concept, an online state monitoring device for DC support capacitors based on PINN is provided, comprising: a dataset acquisition unit, used to acquire multiple sets of charging and discharging data of voltage changes over time under different capacitance values ​​of DC support capacitors, and to determine a training set and a validation set; a model training unit, used to train a preset PINN model according to the training set until a first preset number of iterations is reached, and to acquire the model parameters of the trained PINN model, wherein the PINN model is a feedforward fully connected deep neural network, and its model parameters are adjusted according to the solution of the partial differential equation constructed based on the pre-charging model of the DC support capacitor; a model validation unit, used to update the model parameters of the trained PINN model using the L-BFGS algorithm, and to validate it according to the validation set to obtain the final PINN model; and an online monitoring unit, used to acquire the pre-charging data of the DC support capacitor under test online, and to monitor the state of the DC support capacitor under test by applying the final PINN model according to the pre-charging data.

[0014] Based on the same inventive concept, this invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned method.

[0015] Based on the same inventive concept, embodiments of the present invention also propose a computer storage medium storing at least one executable instruction that causes a processor to execute the aforementioned method.

[0016] This invention acquires charging and discharging data of voltage variations over time for multiple sets of DC support capacitors with different capacitance values, and determines training and validation sets. The AdamW algorithm is used to train a pre-defined PINN model based on the training set, adaptively updating the loss function weights and linear network weights of each component until a first preset number of iterations is reached. The trained PINN model's parameters are then obtained. The PINN model is a feedforward fully connected deep neural network, and its parameters are adjusted based on solving the partial differential equations constructed from the pre-charging model of the DC support capacitor. The L-BFGS algorithm is used to update the trained PINN model's parameters, and the model is validated using the validation set to obtain the final PINN model. The pre-charging data of the DC support capacitor under test is acquired online, and the final PINN model is applied to monitor the state of the DC support capacitor under test based on the pre-charging data. This method has lower requirements for the voltage sensor's sampling frequency and does not require strict alignment of the capacitor pre-charging voltage curve with the time axis, effectively reducing the impact of the measurement signal-to-noise ratio on the prediction results. It is suitable for predicting capacitance values ​​over a wide range of capacitance intervals within the same topology.

[0017] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0019] Figure 1 A flowchart illustrating the online status monitoring method for DC support capacitors based on PINN provided in an embodiment of the present invention is shown.

[0020] Figure 2 A schematic diagram of the parallel structure of the DC support capacitor side according to an embodiment of the present invention is shown;

[0021] Figure 3 A schematic diagram of the pre-charging model of the DC support capacitor in the middle of the traction converter according to an embodiment of the present invention is shown.

[0022] Figure 4 A schematic diagram of the PINN model according to an embodiment of the present invention is shown;

[0023] Figure 5 A schematic diagram illustrating the training process of the adaptive PINN model according to an embodiment of the present invention is shown.

[0024] Figure 6 A schematic diagram of the original CGAN network structure is shown;

[0025] Figure 7 A schematic diagram of the CGAN network structure according to an embodiment of the present invention is shown;

[0026] Figure 8 This diagram illustrates the verification process of the CGAN model results in the training of the PINN network according to an embodiment of the present invention.

[0027] Figure 9 The experimental charging voltage curve and schematic diagram of each group of labels of an embodiment of the present invention are shown.

[0028] Figure 10 This diagram illustrates a comparison between the predicted capacitance values ​​of each algorithm in this embodiment and the actual capacitance values.

[0029] Figure 11A schematic diagram of the structure of the PINN-based DC support capacitor online status monitoring device provided in an embodiment of the present invention is shown.

[0030] Figure 12 A schematic diagram of an electronic device according to an embodiment of the present invention is shown. Detailed Implementation

[0031] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0032] Figure 1 A schematic flowchart of the online status monitoring method for DC support capacitors based on PINN provided in an embodiment of the present invention is shown. Figure 1 As shown, the online status monitoring method for DC support capacitors based on PINN includes:

[0033] Step S11: Obtain charging and discharging data of voltage changes over time for multiple sets of DC support capacitors with different capacitance values, and determine the training set and validation set.

[0034] In this embodiment of the invention, the PINN-based online status monitoring method for DC support capacitors is applied to a server, which can be any computer, PC, or other electronic device. In step S11, optionally, the aging process of the DC support capacitor is simulated by connecting different numbers of film capacitors of the same capacity in parallel outside the DC support capacitor. This obtains ordered voltage arrays and ordered time arrays with the same sampling interval for different values ​​of the DC support capacitor, serving as charge / discharge data of voltage changing over time. The DC support capacitor is then replaced, and multiple sets of ordered voltage arrays and ordered time arrays with the same sampling interval for different values ​​of the DC support capacitor are obtained. Based on these multiple sets of ordered voltage arrays and ordered time arrays for different values ​​of the DC support capacitor, a training set and a validation set are determined. The sample data in the training and validation sets include charging curves showing the voltage changes over time for a large number of DC support capacitors with different capacitance values.

[0035] In this embodiment of the invention, a method of connecting different numbers of thin-film capacitors of the same capacity in parallel to the DC support capacitor side is adopted for a simplified topology of the DC circuit in the pre-charging stage. This simulates the change process of capacitance C decreasing and equivalent series resistance R_ESR increasing during the actual capacitor aging process. The parallel structure diagram of the DC support capacitor side is shown below. Figure 2As shown, 10 external small capacitors are connected in parallel with the converter capacitor C0, and each small capacitor is connected in series with a switch. When all 10 switches are closed, the total capacitance is at its maximum. Then, as each switch is opened sequentially, the total capacitance gradually decreases, thus simulating the actual capacitor aging process. Based on this, several other sets of data are obtained by appropriately replacing the main DC support capacitor C0. This yields a training set with the same capacitance value interval within each group but discontinuous intervals between groups. This training set is used to study a training method for a capacitor monitoring model with a wide capacitance range but uneven label distribution.

[0036] This invention utilizes an RC experimental platform to simulate the pre-charging process. An external oscilloscope is used to obtain the voltage-time charging and discharging curves for different capacitance values. An LCR analyzer is used to measure the state characteristic parameters of the capacitor at the DC support capacitor location, primarily focusing on the capacitance value C. Simultaneously, to ensure that the capacitance measurement using the LCR analyzer is not interfered with by other parts of the topology, the original support capacitor is disconnected and replaced with an external thin-film capacitor connected in parallel at the same location. This allows for obtaining more independent and realistic capacitance parameters, and facilitates easier adjustment of the capacitance range, resulting in a richer training set.

[0037] Step S12: Use the AdamW algorithm to train the preset PINN model based on the training set, adaptively update the loss function weights of each component and the linear network weights until the first preset number of iterations is reached, and obtain the model parameters of the trained PINN model. The PINN model is a feedforward fully connected deep neural network, and its model parameters are adjusted according to the solution of the partial differential equations constructed based on the pre-charging model of DC support capacitor.

[0038] In step S12, optionally, the sample data from the training set is input into a preset PINN model to obtain the output result and the current network structure; the total loss of the PINN model is calculated, and the model weights are updated using the AdamW algorithm until a first preset number of iterations is reached. Specifically, an adaptive loss function is constructed based on the partial differential equation to obtain the total loss of the PINN model; then, the AdamW algorithm is used to simultaneously update the adaptive weights of each loss component in the adaptive loss function and the model weights of the PINN model. The PINN model is a feedforward fully connected deep neural network, which is a linear network. The model parameters of the PINN model are adjusted based on the solution of the partial differential equation constructed from the pre-charge model based on the DC-DC supporting capacitor.

[0039] In this embodiment of the invention, the basic modeling idea of ​​PINN is to construct a feedforward fully connected deep neural network, using partial differential equations and initial conditions. and boundary conditions Residuals are constructed and added to the loss function according to relative weights. The automatic differentiation technique in deep neural networks is used to minimize the residuals, thereby optimizing the network parameters. This approximates the solution of the differential equation within an effective number of iterations, yielding reliable prediction results and a prediction model. The equation is as follows:

[0040]

[0041] Assume its solution function is , That is, a pair Functions that perform differentiation operations For parameter variables in a functional, For spatial variables, For time variables, It is a subset of Euclidean space. This is the termination boundary time. From the perspective of the PINN deep learning algorithm, it is defined as follows: To approximate the function For a feedforward fully connected deep neural network, the above equation is combined with its initial conditions. With boundary conditions It can be represented as:

[0042]

[0043] in, Here is the weight matrix of the feedforward fully connected deep neural network. This represents the bias vector of a feedforward fully connected deep neural network. Representing network parameters The set of values. The loss function is minimized using optimization algorithms such as the conjugate gradient method, gradient descent method, or Newton's method, thus optimizing the internal weight parameters of the network. Define the network output as , To optimize the objective, As the loss function, the optimization process can be expressed as:

[0044]

[0045]

[0046] In this embodiment of the invention, the partial differential equations and corresponding initial and boundary conditions are obtained based on the pre-charge model of the DC support capacitor in the middle of the traction converter. The pre-charge model of the DC support capacitor in the middle of the traction converter is as follows: Figure 3As shown, before the traction converter begins formal operation, the DC support capacitor needs to be charged through a pre-charging unit with a resistor. This ensures that a large peak current is not instantaneously generated due to the input voltage being directly applied to the uncharged support capacitor, which could have unnecessary negative impacts on the safety of the traction converter and the capacitor's lifespan. At the start of the pre-charging phase, the pulse width modulation inverter is disconnected from the intermediate DC link to reduce circuit interference. Only after the pre-charging phase is completed (generally, when the voltage difference between the input voltage and the charging voltage is less than 50V) will the contactor on the input line be closed, connecting the pulse width modulation inverter to begin normal operation.

[0047] According to Thevenin's theorem and Norton's theorem, and based on the formula for a first-order zero-response circuit with a capacitor, we can obtain the following formula for a known charging input DC voltage: Charging time and pre-charge resistor Under the premise that the capacitor voltage divider is obtained during pre-charging, the voltage divider can be obtained. for:

[0048]

[0049] The mathematical formula for calculating the capacitance C of the DC-link capacitor is as follows:

[0050]

[0051] As can be seen from the above formula, it seems that as long as a certain point in the pre-charge voltage curve is used as a reference, and the voltage across the capacitor in adjacent sampling periods is obtained simultaneously, the capacitance value C and the equivalent series resistance can be derived mathematically. However, in practice, the identification results obtained by mathematical calculations often have significant errors. These errors are due to several reasons, including but not limited to: the formulas contain the natural constant e, and solving systems of equations with transcendental numbers can easily lead to the accumulation of complex errors; the pre-charge voltage curves of the capacitors monitored online have significant noise, and directly using the formulas for calculation may result in prediction errors due to difficulty in obtaining the optimal feature points. Considering these reasons, this embodiment of the invention will use a method combining physical law-driven and data-driven approaches to monitor the state of the DC support capacitors of the traction converter.

[0052] In practical engineering scenarios, the charge-discharge curve is a curve controlled by multiple variables. However, for a fixed converter system with an unchanged topology, the pre-charge curve of the capacitor voltage obtained after each charge-discharge process is affected by other factors in the same way. Meanwhile, the capacitance (C) and capacitance (C) caused by capacitor aging... The change is the main reason for the variation in the voltage pre-charge curve. From the perspective of the physical model, the charge-discharge curve has a clear characteristic of varying the charging voltage. Charging time t, capacitance value C, and equivalent resistance The connecting mathematical formulas. Obtaining a differential equation model that can be simultaneously propagated in both forward and backward directions using automatic differentiation techniques is the primary core step of the PINN algorithm.

[0053] Compared to the input resistance, the equivalent series resistance of the capacitor ( The effect of capacitance at the measured frequency is negligible, and its impact on the charge-discharge curve is even smaller. Due to experimental limitations, we will only analyze and predict the capacitance value C, a characteristic parameter of capacitor lifetime. The following modifications can be made:

[0054]

[0055] Assuming the charging voltage E fluctuates relatively little, the output physical quantity is determined to be the capacitance value C, and the input physical quantity is an ordered array of a segment of the charging voltage curve based on the sampling interval. And a discrete arithmetic-time ordered array t that conforms to the sampling interval. The constructed partial differential equation (PDE) is as follows:

[0056]

[0057] Determining the initial and boundary conditions of a partial differential equation is a sufficient condition for the equation to obtain a solution; it is a hard constraint method. The system of equations consisting of the initial and boundary conditions of the partial differential equation is shown below:

[0058]

[0059] Thus, the structure of the PINN model in this embodiment of the invention is as follows: Figure 4 The image shows a feedforward fully connected deep neural network, whose optimization objective is... The goal is to minimize the loss function based on the partial differential equation (PDE), initial conditions, boundary conditions, and data residuals.

[0060] The loss function of the PINN algorithm is mainly composed of data loss, partial derivative residual loss, initial loss, and boundary loss, summed according to the weights of each component. To briefly describe the network input-output relationship, we use... Characterizing the input ordered voltage array The result is predicted by the PINN network using the input ordered time array t, where i represents the ordinal number, the subscript indicates the source of the input data, the subscript d indicates that the data comes from the experiment or the reliable generation result of the validation stage of the generative adversarial network, and the subscript p indicates the random number generated within the range of initial and boundary conditions.

[0061] The loss function is the mean squared error loss function, which is defined as follows:

[0062]

[0063] In this embodiment of the invention, the data loss is obtained by calculating the mean square error between the output value and the true value of the feedforward fully connected deep neural network.

[0064]

[0065] Input random numbers from the selectable domain into the network to obtain the initial loss and boundary loss, as shown in the following equations:

[0066]

[0067]

[0068] The output and input values ​​of the feedforward fully connected deep neural network are used to construct a partial differential equation, and the residual loss of the partial differential equation is obtained:

[0069]

[0070] in, The label indicates the actual measured capacitance value. This represents the boundary conditions under the physical formula. This represents the residual obtained by substituting the network results into the partial differential equation (PDE). The logical relationship is as follows:

[0071]

[0072] The total loss of the PINN model is obtained by adaptively weighting and summing the data loss, the initial loss, the boundary loss, and the partial differential residual loss. Therefore, the adaptive PINN total loss function to be optimized can be expressed as:

[0073]

[0074] In this embodiment of the invention, the training process of the adaptive PINN model is as follows: Figure 5 As shown, it includes:

[0075] Step 100: Obtain the preprocessed training set and validation set.

[0076] Step 101: Select appropriate network layers, network nodes, activation functions, optimization algorithms, number of iterations, and adaptive initial weight values.

[0077] Step 102: Input the training set into the PINN model and obtain the output results and the current network structure.

[0078] Step 103: Calculate the mean squared error between the output and the true value to obtain the data loss.

[0079] Step 104: Input random numbers from the feasible region into the network to obtain the initial loss and boundary loss.

[0080] Step 105: Construct a partial differential equation with the output value and the input value to obtain the PDE residual loss.

[0081] Step 106: Weight each loss component, sum them to obtain the total loss, and backpropagate to train the network.

[0082] Step 107: Use the AdamW algorithm to simultaneously update the adaptive weights of each loss component and the model weights.

[0083] Step 108: Determine if the first preset number of iterations has been reached. If yes, proceed to step 109; otherwise, return to step 102.

[0084] Step 109: Update the model weights again using the L-BFGS algorithm.

[0085] Step 110: Determine if the network performs well on the validation set. If yes, proceed to step 111; otherwise, return to step 101.

[0086] Step 111: Obtain reliable prediction results and the PINN model.

[0087] In this embodiment of the invention, considering the uneven distribution and insufficient quantity of training set sample data, a Conditional Generation Adversarial Network (CGAN) is used to expand the original training sample dataset. By expanding the data with a limited number of collected samples, the minimum required training set sample size for the PINN deep learning algorithm is met. This effectively improves the performance of the PINN network, achieves good training results, and reduces the requirement for a small dataset size.

[0088] Generative Adversarial Networks (GANs) are modeled based on zero-sum games in game theory. They consist of a generator to generate samples and a discriminator to distinguish between real and generated samples. Through adversarial training, the performance of each component is continuously optimized. Generally, when the discriminator struggles to differentiate between real and generated samples, a Nash equilibrium is reached, and the GAN model is considered to have been successfully trained. Conditional Generative Adversarial Networks (CGANs) are a common variant of GANs. Their key feature is the addition of the same conditional variable y to both the generator and discriminator, increasing the constraints on model training and improving model reliability. The original CGAN network structure is shown below. Figure 6As shown, the noise data z is generated by the generator under the constraints of the conditional data y. Real data x and generated results The joint conditional data y constitutes the input to the discriminator. The discriminator output is a scalar. The generator's goal is to make the generated results as pleasing as possible. Conclusions on the discriminator Conclusions on the discriminator based on real data As close as possible. This process uses value functions. It can be represented as:

[0089]

[0090] In this embodiment of the invention, the training process of the PINN model further includes: determining whether the number of iterations has reached a second preset number of iterations. The second preset number of iterations Less than the first preset number of iterations If the number of iterations reaches the second preset number of iterations and the data loss is less than the preset threshold. The training set is expanded using a conditional generative adversarial network (GAN). The total loss of the PINN model is updated to a weighted sum of data loss, initial loss, boundary loss, partial differential residual loss, and generated data loss, where the generated data loss is the mean squared error between the output value and the true value of the feedforward fully connected deep neural network based on the generated data. The PINN network reaches the second preset number of iterations. After adjusting for the prediction accuracy on the test set, the validation-stage reliable generation results obtained by combining with the CGAN network are similar in characteristics to the experimental data; therefore, the same weight settings are used. The loss for this data generation can be expressed as follows:

[0091]

[0092] The generated data is { }, the corresponding generating matrix is ​​{ , , }

[0093] When the PINN network simultaneously reaches the second preset number of iterations When considering the prediction accuracy of the test set, the total loss function can be updated as follows:

[0094]

[0095] In this embodiment of the invention, when using conditional generative adversarial networks to expand the training set, optionally, the sample data in the training set can be grouped, and the calculated voltage vector can be obtained based on the experimental voltage vector. A random noise vector z, following a probability distribution, is taken as input and processed by a generator in a conditional generative adversarial network to randomly generate a noise vector. The calculated voltage vector With the noise vector Adding them together yields negative samples. ; the experimental voltage vector The negative sample serves as a true sample. As fake samples, a discriminator in a conditional generative adversarial network is used to distinguish them and determine the reliable result vector. The trusted result vectors and corresponding time vectors obtained by the applied conditional generative adversarial network are concatenated into two corresponding matrices and added to the training set to obtain the expanded training set.

[0096] Specifically, the CGAN network in this embodiment of the invention is as follows: Figure 7 As shown, the basic process is as follows:

[0097] (1) First stage: First, the original training set is randomly divided into m groups of equal number according to the label (m is a random natural number), and the experimental voltage vector is... The voltage vector is obtained according to the physical formula.

[0098]

[0099]

[0100] It is important to note that the current calculation of the voltage vector... The results are only mathematical derivations of the experimental voltage vectors, not highly accurate and completely reliable predictions, and are only used to assist the CGAN network in generating data.

[0101] (2) Second stage: A set of random noise vectors z that follow a probability distribution is used as input and a generator is used to randomly generate noise vectors. Adding the calculation of voltage vector Then treat it as a whole as a negative sample At the same time The partial differential residual is obtained as input to the current PINN network, and is used as the directional conditional data y for the generator and discriminator in the next loop.

[0102] (3) Third stage: Real experimental voltage vector As a true sample, As fake samples, they are combined with conditional data y to train the discriminator.

[0103] In this embodiment of the invention, during the training of the training set using the PINN network, when the number of iterations exceeds a second preset number of iterations... And the mean square error (MSE) of the test set is less than a set small number. Only when the time vector is set will the CGAN model be used to generate data. At this point, the current network weights, adaptive weights, and training set are saved, along with the current iteration number, and a fixed time vector is generated. and all provisionally reliable result vectors obtained in the CGAN model The corresponding matrices are concatenated into two matrices. and , two matrices and Add these two matrices to the training set and use their predictions under the current PINN network as their capacitance labels. The training results after adding the generated training set are monitored by the accuracy of the test set after each iteration. If the accuracy decreases within a fixed iteration interval, the generated training set is considered unsatisfactory, and the generation process is restarted from the last saved iteration number. If the accuracy remains unchanged or improves within a fixed iteration interval, the generated training set is considered reliable and effective, and the generated dataset is updated before the next fixed iteration interval begins. At this time, the current generated dataset replaces the original training set derived computation data.

[0104] The validation process of CGAN model results in PINN network training is as follows: Figure 8 As shown, it includes:

[0105] Step 200: Train the training set using the PINN network.

[0106] Step 201: Determine whether the second preset number of iterations has been reached and the data loss is less than 10%. If yes, proceed to step 202; otherwise, return to step 200.

[0107] Step 202: Apply the conditional generative adversarial network (GAN) generator matrix and .

[0108] Step 203: Save the current network model and iteration count, and generate the matrix. and Add it to the training set for training.

[0109] Step 204: Determine whether the accuracy remains unchanged or improves. If yes, proceed to step 205; otherwise, return to step 202.

[0110] Determine whether the accuracy remains unchanged or improves within a fixed iteration interval. If the accuracy remains unchanged or improves within the fixed iteration interval, the generated training set is considered reliable and effective; otherwise, the generated training set is considered unsatisfactory.

[0111] Step 205: Return to the previously saved network model and iteration count.

[0112] Step 206: Determine if the first preset number of iterations has been reached. If yes, proceed to step 207; otherwise, return to step 200.

[0113] Step 207: End.

[0114] This invention utilizes the Conditional Generative Adversarial Network (CGAN) algorithm to augment data, enabling the PINN-based online state monitoring method for DC support capacitors to be applied to capacitor value prediction over a wide range of capacitor intervals under the same topology, thus reducing the training requirements of the PINN model.

[0115] Step S13: Update the model parameters of the trained PINN model using the L-BFGS algorithm, and validate it according to the validation set to obtain the final PINN model.

[0116] In this embodiment of the invention, for the adaptive PINN part, the AdamW algorithm is first used to simultaneously update the adaptive weights of each loss component and the weights of the linear network, until a first preset number of iterations is reached. Next, the L-BFGS algorithm is used to iteratively optimize the linear network weights of the current PINN network, but the adaptive weights are not updated. Then, the trained PINN model is validated using the validation set. If the PINN model performs well on the validation set, it indicates that the trained PINN model can obtain reliable prediction results, and this trained PINN model is the final PINN model.

[0117] Step S14: Acquire the pre-charge data of the DC support capacitor under test online, and apply the final PINN model to monitor the state of the DC support capacitor under test based on the pre-charge data.

[0118] In this embodiment of the invention, the voltage curve data and sampling interval of the DC support capacitor under test during the pre-charging stage are collected online using existing sensors in the traction converter to obtain the pre-charging data of the DC support capacitor under test. This pre-charging data is then input into the final PINN model obtained earlier to obtain the capacitance value of the DC support capacitor under test output by the PINN model, thereby realizing online status monitoring of the DC support capacitor under test.

[0119] To verify the effectiveness and feasibility of the PINN-based online state monitoring method for DC support capacitors in this invention, a laboratory platform was used for verification. Specifically, an RC experimental platform was used to simulate the pre-charging process, and an external oscilloscope was used to obtain the voltage-time charging and discharging curves for different capacitance values. An LCR analyzer was used to measure the state characteristic parameters of the capacitor at the DC support capacitor location, primarily focusing on the capacitance value C. Simultaneously, to ensure that the capacitance measurement using the LCR analyzer is not affected by interference from other parts of the topology, the original support capacitor was disconnected and replaced with an external thin-film capacitor connected in parallel at the same location. This allows for obtaining more independent and realistic capacitance parameters, and facilitates easier adjustment of the capacitance range, resulting in a richer training set.

[0120] Because it is necessary to collect charging curves of a large number of capacitors with different capacitance values ​​C, but due to limitations in experimental and practical engineering measurement conditions, it is difficult to obtain measurement data with continuous intervals for capacitance values ​​C, this experiment is based on... Figure 2 The parallel circuit on the DC support capacitor side shown uses a combination of large and small capacitors (the three large capacitors have capacitance values ​​of 1.14537mF, 0.41026mF, and 0.41026mF respectively, and the ten small capacitors are all 0.01mF). This yields five groups of measurement data with equal intervals within each group and larger intervals between groups, satisfying the requirement for a wide capacitance range. The specific grouping of the five capacitance intervals is shown in Table 1, where the capacitance value interval within each interval is 0.01mF, and the maximum capacitance value of 2.06589mF is 5.036 times the minimum capacitance value of 0.41026mF.

[0121] Table 1. Specific details of capacitor interval grouping

[0122]

[0123] First, the capacitance characteristic parameters of the capacitor under test were measured using an LCR analyzer. To reduce measurement error, each capacitor was measured three times and the average value was taken. After connecting the experimental circuit and the measurement circuit, the capacitor under test was connected in parallel and the initial experimental information was recorded. The contactor was disconnected to perform a pre-charging process, the charging voltage curve was measured and saved, and the capacitor was discharged. If the capacitance value under test has not been completely measured at this time, the capacitance value was adjusted and the above experiment was repeated. If all measurements were completed, the experimental platform was cleaned up and the experiment was ended. The main experimental parameters are shown in Table 2.

[0124] Table 2 Main Parameters of the Experiment

[0125]

[0126] The adaptive PINN network was built on a Windows system using Python 3.9, with PyTorch as the backend. Training and testing were performed on an NVIDIA GeForce RTX 3050 Ti Laptop GPU. Assuming a small fluctuation in the charging voltage E, the output physical quantity was determined to be the capacitance C. The input physical quantities were an ordered array u(t) of the charging voltage curve based on the sampling interval and an ordered array t of discrete arithmetic time intervals conforming to the sampling interval. The original training data size, the generated training data size, and the test data size were 50, 50, and 5, respectively. After multiple trials, analyses, and corrections, the adaptive PINN model was configured with 5 hidden layers, each with 50 neurons, and the tanh activation function. The AdamW optimizer was used to optimize the network weights and adaptive weights, with an initial adaptive weight value of 1 and an initial learning rate of 10⁻⁴. The number of iterations was... =50,000 epochs, of which At 30,000 epochs, the CGAN network is introduced, and then the L-BFGS algorithm is used to perform a new round of iterative optimization on the linear network weight part of the current PINN model. The number of iterations is 500 epochs, and the learning rate is 10⁻⁵.

[0127] Five capacitor values ​​(1.16537mF, 1.63563mF, 2.03589mF, 0.85052mF, 0.51026mF) were randomly selected from the range as the test dataset and training / test set, with the remaining 50 values ​​used as the training dataset. To test the applicability of this method across extended capacitance ranges, a sixth test dataset was set up with a capacitance value outside the five ranges, with a value of 1.4432mF. The ratio of the training set to the test set was approximately 10:1. Since time axis alignment is not required, the experiment theoretically only requires measuring the charging voltage curve and the corresponding capacitance value C. Simultaneously, the first data point was selected from the ordered array of charging voltage curves for each capacitor. (0.5,1), the remaining data volume The continuous subarray is used to construct a discrete arithmetic sequence with elements conforming to the sampling interval, based on the length of the subarray; this is the proposed time axis. Due to the PINN algorithm's fusion of physical information, accurate optimization can be achieved even if the time axis cannot be perfectly matched.

[0128] Considering the computer operating costs, the charging voltage curve reaches the first time constant with the smallest measured capacitance value. Based on 2%E), until the third time constant ( (%E) End, select the charging voltage curve values ​​of all capacitors during this time period. Using the corresponding time axis t as a feature quantity, a data sample set is constructed with the output physical quantity as the capacitance value C. The time range selected is... The experimental charging voltage curve and the labels for each group are as follows: Figure 9 As shown, 1, 2, 3, 4, 5, and 6 represent capacitors with different numbers. There are four main performance evaluation indicators, among which... Represents the true value. This represents the predicted value. Mean Squared Error (MSE) :

[0129]

[0130] Mean squared error describes the deviation between the model's predictions and the actual results. It is mainly used for internal loss calculation and phased evaluation of the model on the test set.

[0131] Mean Absolute Percentage Error (MAPE) :

[0132]

[0133] Mean absolute percentage error (MASE) can describe the accuracy and robustness of a prediction model and is mainly used to present the model's prediction performance.

[0134] The absolute error (Δ) of the prediction result and relative error percentage ( ) :

[0135]

[0136]

[0137] The absolute error of the prediction result can intuitively show the deviation value of the prediction result, while the relative error can intuitively show the degree of deviation of the prediction result.

[0138] The smaller the above performance indicators, the higher the prediction accuracy, the better the optimization effect, and the more reliable the model and prediction results.

[0139] After a complete experimental process of the adaptive PINN method, the prediction results of the test set in the training model are shown in Table 3.

[0140] Table 3. Prediction results of the test set in the trained model

[0141]

[0142] As shown in the table, when the parallel small capacitor of 0.01mF accounts for about 1% of the actual capacitance value of the tested capacitor, the relative error is around 0.15%, which meets the actual prediction requirements in engineering. When the proportion is close to 2%, the relative error increases, with the relative error of capacitor number 5 being much higher than that of other tested capacitors. However, the absolute error is still not much different from other test samples. This is mainly due to the limitations of experimental conditions, where the capacitance value interval within the range is fixed at 0.01mF, accounting for as much as 1.95% of the actual capacitance of capacitor number 5. This indicates that the initial error of 0.01mF is relatively large compared to the range in the training set, which is also the main reason for the large relative error of the prediction result. At the same time, the actual capacitance of capacitor number 5 accounts for only 6.04% of the entire capacitance range, belonging to the marginal sample set. Therefore, the relative error of 0.48% is acceptable from the perspective of relative accuracy, and the method is reliable. As a test data far from the training set, capacitor number 6 has a capacitance value identification error of less than 1%, which proves that the method has reference value in constructing a training model with a wide range of capacitance intervals but uneven label distribution.

[0143] Three methods were compared. The pure BP algorithm and the original PINN algorithm share the same initial parameter configuration as the improved PINN method in this example, facilitating variable control. The prediction results and relative errors of each algorithm are shown in Table 4. To visually demonstrate the differences between the algorithms, the actual capacitor values ​​for each number are taken as the nominal value 1, and the comparison of the predicted capacitor values ​​relative to the actual capacitor values ​​is shown below. Figure 10 As shown in the figure. The experimental results clearly show that the original PINN network provides better prediction results than the BP network, while the improved adaptive PINN algorithm in this example performs the best. Therefore, the improved adaptive PINN algorithm in this example can significantly improve the prediction accuracy.

[0144] Table 4. Prediction Results and Relative Errors of Each Algorithm

[0145]

[0146] The PINN-based online state monitoring method for DC support capacitors in this invention is based on an adaptive physical information deep neural network (PINN) and a capacitor pre-charging model, supplemented by a conditional generative adversarial network (CGAN) to expand the data. As long as a reasonable and effective data sample set is established initially, the pre-trained model can be used for capacitance value monitoring during application, improving parameter monitoring efficiency. With appropriate modifications, it can be applied to most train DC-Link capacitor application scenarios, exhibiting strong universality. It can predict capacitance values ​​over a wide range of capacitance intervals under the same topology, with the maximum value reaching up to five times the minimum value. Simultaneously, it overcomes the problem of sparse and uneven distribution caused by limitations in data sample training set creation, reducing model training requirements. It is practical, simple, and highly feasible in engineering. No additional sensors are needed, nor are modifications to the original circuit structure and system control algorithm required. Only existing sensor measurement data (voltage curve data and sampling intervals) from the pre-charging stage are needed. Preprocessing does not require data fitting or strict time axis alignment to obtain relatively accurate capacitance state monitoring results.

[0147] In summary, the online status monitoring method for DC support capacitors based on PINN in this embodiment of the invention achieves the desired results.

[0148] Multiple sets of charging and discharging data of voltage variations over time for different capacitance values ​​of DC support capacitors are acquired, and training and validation sets are determined. A preset PINN model is trained using the training set until a first preset number of iterations is reached. The model parameters of the trained PINN model are obtained. The PINN model is a feedforward fully connected deep neural network, and its model parameters are adjusted based on solving the partial differential equations constructed using a pre-charging model of the DC support capacitor. The L-BFGS algorithm is used to update the model parameters of the trained PINN model, and the model is validated using the validation set to obtain the final PINN model. Pre-charging data of the DC support capacitor under test is acquired online, and the final PINN model is applied to monitor the state of the DC support capacitor under test based on the pre-charging data. This method has lower requirements for the sampling frequency of the voltage sensor and does not require strict alignment of the capacitor pre-charging voltage curve with the time axis, effectively reducing the impact of the measurement signal-to-noise ratio on the prediction results. It is suitable for predicting capacitance values ​​over a wide range of capacitance intervals under the same topology.

[0149] The foregoing has described specific embodiments of the present invention. In some cases, the actions or steps described in the embodiments of the present invention may be performed in a different order than that shown in the embodiments and the desired results may still be achieved. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0150] Based on the same concept, this invention also provides an online status monitoring device for DC support capacitors based on PINN. It is applied to servers. (See attached document.) Figure 11 As shown, the PINN-based online status monitoring device for DC support capacitors includes: a dataset acquisition unit, a model training unit, a model validation unit, and an online monitoring unit. Among them,

[0151] The dataset acquisition unit is used to acquire charging and discharging data of voltage changes over time under different capacitance values ​​of DC support capacitors, and to determine the training set and validation set.

[0152] The model training unit is used to train the preset PINN model according to the training set until the first preset number of iterations is reached, and to obtain the model parameters of the trained PINN model. The PINN model is a feedforward fully connected deep neural network, and its model parameters are adjusted according to the solution of the partial differential equation constructed based on the pre-charging model of DC support capacitor.

[0153] The model validation unit is used to update the model parameters of the trained PINN model using the L-BFGS algorithm and to validate the model according to the validation set to obtain the final PINN model.

[0154] An online monitoring unit is used to acquire the pre-charge data of the DC support capacitor under test online, and to monitor the state of the DC support capacitor under test by applying the final PINN model based on the pre-charge data.

[0155] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing the embodiments of the present invention, the functions of each module can be implemented in one or more software and / or hardware.

[0156] The apparatus of the above embodiments is applied to the corresponding methods in the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0157] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method described in any of the above embodiments.

[0158] This invention provides a non-volatile computer storage medium storing at least one executable instruction that can execute the method described in any of the above embodiments.

[0159] Figure 12This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1201, a memory 1202, an input / output interface 1203, a communication interface 1204, and a bus 1205. The processor 1201, memory 1202, input / output interface 1203, and communication interface 1204 are interconnected internally via the bus 1205.

[0160] The processor 1201 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention.

[0161] The memory 1202 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1202 can store the operating system and other application programs. When the technical solution provided by the method embodiment of the present invention is implemented by software or firmware, the relevant program code is stored in the memory 1202 and is called and executed by the processor 1201.

[0162] Input / output interface 1203 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0163] The communication interface 1204 is used to connect the communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0164] Bus 1205 includes a pathway for transmitting information between various components of the device, such as processor 1201, memory 1202, input / output interface 1203, and communication interface 1204.

[0165] It should be noted that although the above-described device only shows the processor 1201, memory 1202, input / output interface 1203, communication interface 1204, and bus 1205, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of the present invention, and does not necessarily include all the components shown in the figures.

[0166] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present invention as described above, which are not provided in detail for the sake of brevity.

[0167] This application is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of all embodiments. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this invention should be included within the protection scope of this disclosure.

Claims

1. A method for online status monitoring of DC support capacitors based on PINN, characterized in that, The method includes: Acquire charging and discharging data of voltage changes over time for multiple sets of DC support capacitors with different capacitance values, and determine the training set and validation set; The AdamW algorithm is used to train the preset PINN model based on the training set, adaptively updating the loss function weights and linear network weights of each component until the first preset number of iterations is reached. The model parameters of the trained PINN model are then obtained. The PINN model is a feedforward fully connected deep neural network. Its model parameters are adjusted based on the solution of the partial differential equation constructed by the pre-charging model based on the DC support capacitor. The basic modeling idea of ​​PINN is to construct a feedforward fully connected deep neural network, construct residuals using partial differential equations, initial conditions, and boundary conditions, and add them to the loss function according to relative weights. The automatic differentiation technique in deep neural networks is used to minimize the residuals, thereby optimizing the network parameters. The solution of the differential equation is approximated within an effective number of iterations to obtain reliable prediction results and prediction models. The L-BFGS algorithm is used to update the model parameters of the trained PINN model, and the model is validated according to the validation set to obtain the final PINN model. The pre-charge data of the DC support capacitor under test is acquired online, and the final PINN model is applied based on the pre-charge data to monitor the state of the DC support capacitor under test.

2. The method according to claim 1, characterized in that, The process of acquiring charging and discharging data of voltage variations over time at different values ​​of the DC support capacitor, and determining the training and validation sets, includes: By simulating the aging process of DC support capacitors by connecting different numbers of thin-film capacitors of the same capacity in parallel outside the DC support capacitor, the voltage ordered array and time ordered array of the same sampling interval under different values ​​of DC support capacitor are obtained as charging and discharging data of voltage changing with time. By changing the DC support capacitor, obtain multiple ordered arrays of voltage and time data with the same sampling interval under different values ​​of the DC support capacitor; The training set and validation set are determined based on ordered voltage and ordered time arrays for different DC support capacitor values.

3. The method according to claim 1, characterized in that, The step of training a preset PINN model based on the training set until a first preset number of iterations is reached, and then obtaining the model parameters of the trained PINN model, includes: Input the sample data from the training set into the preset PINN model to obtain the output results and the current network structure; Calculate the total loss of the PINN model and update the model weights using the AdamW algorithm until the first preset number of iterations is reached.

4. The method according to claim 3, characterized in that, The calculation of the total loss of the preset PINN model and the updating of the model weights using the AdamW algorithm include: An adaptive loss function is constructed based on the partial differential equation to obtain the total loss of the PINN model; The AdamW algorithm is used to simultaneously update the adaptive weights of each loss component in the adaptive loss function and the model weights of the PINN model.

5. The method according to claim 4, characterized in that, The step of constructing an adaptive loss function based on the partial differential equation to obtain the total loss of the PINN model includes: The data loss is obtained by calculating the mean square error between the output value and the true value of the feedforward fully connected deep neural network. Input random numbers from the selectable domain into the network to obtain the initial loss and boundary loss; The output value and input value of the feedforward fully connected deep neural network are used to form a partial differential equation to obtain the partial differential equation residual loss. The total loss of the PINN model is obtained by adaptively weighting and summing the data loss, the initial loss, the boundary loss, and the partial differential residual loss.

6. The method according to claim 5, characterized in that, After calculating the total loss of the preset PINN model and updating the model weights using the AdamW algorithm, the method further includes: Determine whether the number of iterations has reached a second preset number of iterations, where the second preset number of iterations is less than the first preset number of iterations; If the number of iterations reaches the second preset number of iterations and the data loss is less than the preset threshold, the adversarial network is augmented with a training set under the applied conditions. The total loss of the updated PINN model is a weighted sum of data loss, initial loss, boundary loss, partial differential residual loss, and generated data loss, wherein the generated data loss is the mean square error between the output value and the true value of the feedforward fully connected deep neural network based on the generated data.

7. The method according to claim 6, characterized in that, The application condition-based generative adversarial network augmentation training set includes: The sample data in the training set are grouped, and the calculated voltage vector is obtained based on the experimental voltage vector. A random noise vector following a probability distribution is used as input and then processed by a generator in a conditional generative adversarial network to generate a random noise vector. The calculated voltage vector is added to the noise vector to obtain a negative sample; The experimental voltage vector is used as the true sample and the negative sample is used as the false sample. The discriminator in the conditional generative adversarial network is used to make a judgment to determine the reliable result vector. All the trusted result vectors and corresponding time vectors obtained by applying the conditional generative adversarial network are concatenated into two corresponding matrices and added to the training set to obtain the expanded training set.

8. A PINN-based online status monitoring device for DC support capacitors, characterized in that, The device includes: The dataset acquisition unit is used to acquire charging and discharging data of voltage changes over time under different capacitance values ​​of DC support capacitors, and to determine the training set and validation set. The model training unit is used to train the preset PINN model using the AdamW algorithm based on the training set, adaptively updating the loss function weights and linear network weights of each component until the first preset number of iterations is reached, and obtaining the model parameters of the trained PINN model. The PINN model is a feedforward fully connected deep neural network, and its model parameters are adjusted according to the solution of the partial differential equation constructed based on the pre-charging model of DC support capacitor. The basic modeling idea of ​​PINN is to construct a feedforward fully connected deep neural network, construct residuals using partial differential equations, initial conditions and boundary conditions, and add them to the loss function according to relative weights. The automatic differentiation technique in deep neural networks is used to minimize the residuals to optimize the network parameters, approximate the solution of the differential equation under the effective number of iterations, and obtain reliable prediction results and prediction models. The model validation unit is used to update the model parameters of the trained PINN model using the L-BFGS algorithm and to validate the model according to the validation set to obtain the final PINN model. An online monitoring unit is used to acquire the pre-charge data of the DC support capacitor under test online, and to monitor the state of the DC support capacitor under test by applying the final PINN model based on the pre-charge data.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-7.

10. A computer storage medium, characterized in that, The storage medium stores at least one executable instruction that causes the processor to perform the method as described in any one of claims 1-7.

Images