Memory retention based power converter open circuit fault lifelong learning diagnostic method

Abstract

The application discloses a power converter open-circuit fault lifelong learning diagnosis method based on memory reservation, constructs a power converter open-circuit fault diagnosis model, collects source data sample sets of the power converter to train the power converter open-circuit fault diagnosis model, obtains source model parameters and corresponding model parameter prior distribution, collects data sample sets of a current task and performs variational memory reservation learning to obtain a power converter open-circuit fault diagnosis model of the current task whenever the working condition of the power converter changes and relearning is needed, the loss function in the variational memory reservation learning includes a KL loss item of model parameter distribution of a previous task and a current task and a classification loss of the current task, and then the obtained open-circuit fault diagnosis model is used to perform open-circuit fault diagnosis under a corresponding task. The application adopts the variational memory reservation strategy, realizes lifelong learning of the power converter open-circuit fault diagnosis model, and improves the accuracy of the power converter open-circuit fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of power converter open circuit fault diagnosis technology, and more specifically, relates to a lifelong learning diagnosis method for power converter open circuit faults based on memory retention. Background Technology

[0002] Power converters convert electrical energy between the public power grid, vehicle power supply network, and traction system, while maintaining high efficiency, high power density, and stable operation under high voltage, high power, and harsh environmental conditions. However, excessive power loss and high thermal stress can easily lead to power converter failures, mainly classified into short-circuit faults and open-circuit faults. Compared to short-circuit faults, the early characteristics of open-circuit faults are more easily overlooked, potentially leading to secondary faults and significant economic losses. Therefore, research on open-circuit fault diagnosis technology for power converters is of great significance. Currently, open-circuit fault diagnosis methods for power converters are mainly divided into two categories: model-based methods and data-driven methods.

[0003] In model-based fault diagnosis methods, fault diagnosis is achieved by quantifying the deviation between the measured signal and the predicted output of the system's mathematical model. Mehmood et al. proposed a nonlinear model sensor fault diagnosis scheme for grid-side converters, which can detect and isolate three-phase AC current and DC voltage sensor faults in real time. Zhuo et al. designed a fast diagnostic method based on a Luneburg observer to address the vulnerability of power switching devices in fuel cell interleaved parallel Boost converters and verified its robustness under complex operating conditions. Zhuo et al. studied a robust power switch fault diagnosis method based on a sliding mode observer for floating interleaved parallel Boost converters, and experimentally verified that it can quickly identify open-circuit faults within two switching cycles.

[0004] However, model-based fault diagnosis methods are highly dependent on the accuracy of system parameters. In real-world industrial scenarios, parameter drift caused by power device aging and thermal effects can lead to discrepancies between the preset mathematical model and the actual operating conditions, thus limiting diagnostic performance. Furthermore, to reduce computational burden, model-based methods often idealize and simplify system dynamics, inevitably reducing diagnostic performance under complex dynamic conditions.

[0005] Data-driven fault diagnosis methods can effectively overcome the limitations of model-based methods. These methods rely solely on historical operational data for fault diagnosis, eliminating the need for precise system models. One literature proposes a particle swarm optimization-based diagnostic scheme to improve the performance of support vector machines (SVMs) and effectively overcome the limitations of traditional SVMs in handling small-sample nonlinear datasets. Kou et al. employed a hybrid diagnostic method combining knowledge-driven and random forest algorithms to solve the problem of IGBT open-circuit fault location in NPC inverters, significantly reducing the algorithm's over-reliance on fault data. Kong et al. designed a diagnostic scheme based on the k-nearest neighbor algorithm for fault diagnosis of inverters in doubly-fed induction wind power generation systems.

[0006] In real-world industrial scenarios, the operating conditions of power converters dynamically adjust according to different task requirements, leading to a continuous emergence of new fault diagnosis tasks. However, most existing deep learning methods rely on specific task data for training, making them ill-suited for the continuously evolving tasks and sequentially arriving data in streaming learning scenarios, and also creating significant data storage pressure. When introducing new operating condition tasks, traditional deep learning methods typically need to integrate all historical data to retrain the model, resulting in substantial storage and computational overhead. Consequently, limited by computational and storage resources, it becomes difficult to retain large amounts of historical data, impacting diagnostic performance. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a lifelong learning diagnostic method for open-circuit faults in power converters based on memory retention. It adopts a variational memory retention strategy to realize the lifelong learning of the open-circuit fault diagnosis model of the power converter, thereby improving the accuracy of open-circuit fault diagnosis of the power converter.

[0008] To achieve the above-mentioned objectives, the present invention provides a lifelong learning diagnostic method for open-circuit faults in power converters based on memory retention, comprising the following steps:

[0009] S1: Construct an open-circuit fault diagnosis model for the power converter, including a GRU network and a classifier, where:

[0010] GRU networks are used for timing monitoring signals of power converters. Feature extraction is performed, and the obtained time-series fault features are sent to the classifier. Represents a time variable;

[0011] The classifier is used to predict the current open-circuit fault category of the power converter based on the characteristics of the time-series faults;

[0012] S2: Acquire source data sample set from the power converter Each data sample includes the timing monitoring signal of the power converter and the corresponding open-circuit fault category label, using the source data sample set. The open-circuit fault diagnosis model for the power converter is trained to obtain the source model parameters. And obtain the distribution of model parameters as the prior distribution. ;

[0013] S3: Whenever the power converter's operating conditions change and relearning is required, collect the data sample set for the current task. , Then, the following variational memory retention learning is used to obtain the open-circuit fault diagnosis model for the power converter. :

[0014] S3.1: Let the number of iterations be... Initialize the parameters of the power converter open-circuit fault diagnosis model. ;

[0015] S3.2: From the data sample set Sample the current batch of data. :

[0016] S3.3: Data sample set Each data sample is input into the power converter open-circuit fault diagnosis model. Perform open-circuit fault category prediction, and then calculate the loss function. And update the obtained model parameters loss function The calculation formula is as follows:

[0017] ,

[0018] in, Indicates the first The distribution of model parameters learned for each task. Indicates the first The distribution of model parameters learned for each task This indicates the solution for the KL divergence. Represents KL loss term The weight, Indicates the first Classification loss for each task;

[0019] S3.4: Determine whether the training end condition for the current task has been met. If it has, the training for the current task ends; otherwise, proceed to step S3.5.

[0020] S3.5: Let the number of iterations be... Return to step S3.2;

[0021] S4: When the task When the power converter needs to be diagnosed for open-circuit faults under the corresponding operating conditions, timing monitoring signals are collected and input into the power converter open-circuit fault diagnosis model. In the process, the open circuit fault diagnosis results were obtained.

[0022] This invention relates to a lifelong learning diagnostic method for open-circuit faults in power converters based on memory retention. It constructs an open-circuit fault diagnosis model for the power converter and trains it using a sample set of source data from the power converter to obtain the source model parameters and their corresponding prior distributions. Whenever the operating conditions of the power converter change and relearning is required, the data sample set for the current task is collected and variational memory retention learning is performed to obtain the open-circuit fault diagnosis model for the current task. The loss function in variational memory retention learning includes the KL loss term of the parameter distributions of the previous and current tasks and the classification loss of the current task. Then, the trained open-circuit fault diagnosis model is used to diagnose open-circuit faults under the corresponding task.

[0023] The present invention has the following beneficial effects:

[0024] 1) This invention alleviates the catastrophic forgetting problem in the continuous learning process of new tasks by constructing a variational memory retention strategy, theoretically solves the problem of "variance shrinkage", realizes lifelong learning of the power converter open circuit fault diagnosis model, and improves the accuracy of power converter open circuit fault diagnosis.

[0025] 2) In this invention, a wavelet-structured gated recurrent unit neural network is designed as a feature extraction module to extract fault discrimination features and time-series features. Furthermore, a discriminative wavelet convolution is proposed to achieve end-to-end optimization, improve the representation ability of time-series monitoring signals, and thus enhance fault diagnosis performance. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a specific implementation of the present invention: a lifelong learning diagnostic method for open-circuit faults in power converters based on memory retention.

[0027] Figure 2 This is a structural diagram of the wavelet-gated recurrent unit neural network in this embodiment;

[0028] Figure 3 It is the probability density plot of the three-phase current timing signal under mean statistics;

[0029] Figure 4 This is a flowchart of variational memory retention learning in this invention;

[0030] Figure 5 This is a schematic diagram illustrating the alignment of model parameter distributions based on Gaussian and vMF distributions. Detailed Implementation

[0031] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0032] Example

[0033] Figure 1 This is a flowchart illustrating a specific implementation of the present invention's lifelong learning diagnostic method for open-circuit faults in power converters based on memory retention. Figure 1 As shown, the lifelong learning diagnostic method for open-circuit faults in power converters based on memory retention of the present invention includes the following steps:

[0034] S101: Constructing a power converter open-circuit fault diagnosis model:

[0035] Construct an open-circuit fault diagnosis model for a power converter, including a GRU network and a classifier, wherein:

[0036] GRU networks are used for timing monitoring signals of power converters. Feature extraction is performed, and the obtained time-series fault features are sent to the classifier. Represents a time variable.

[0037] To achieve more comprehensive detection, open-circuit faults in the power converter include single-transistor faults and dual-transistor faults. To improve detection accuracy, each timing monitoring signal data point needs to cover at least two fault cycles. In this embodiment, the timing monitoring signal uses a three-phase current timing signal.

[0038] The classifier is used to predict the current open-circuit fault category of the power converter based on the characteristics of the time-series fault.

[0039] GRU (Gated Recurrent Unit) is a lightweight recurrent neural network specifically designed for processing time-series signals. By dynamically retaining historical state information through update and reset gates, it effectively captures the current waveform changes, fluctuation patterns, and transient characteristics of power converters under normal and open-circuit fault conditions, while avoiding the gradient vanishing problem of traditional RNNs. GRU transforms the original time-series signal into a high-dimensional fault feature vector, providing a stable and compact feature representation for subsequent fault classification. To further enhance the feature extraction capability of the GRU network for time-series monitoring signals, this embodiment proposes an improved GRU network based on wavelet convolution—the Wavelet Structure Gated Recurrent Unit Neural Network (WSGRU). Figure 2 This is a structural diagram of the wavelet-gated recurrent unit neural network in this embodiment. For example... Figure 2As shown, the wavelet-structured gated recurrent unit neural network in this embodiment includes a wavelet convolutional layer, a 1×1 convolutional layer, a reset gate, an update gate, a candidate hidden state calculation unit, and a hidden state fusion unit, wherein:

[0040] Wavelet convolutional layers are used for time-series monitoring signals. Perform wavelet convolution to obtain signal features It is then sent to the reset gate to update the gate and candidate hidden state calculation unit.

[0041] A 1×1 convolutional layer is used to hide the hidden state from the previous time step. Generate historical features by performing 1×1 convolution operations. It is then sent to the reset gate to update the gate and candidate hidden state calculation unit.

[0042] The reset gate is used to characterize the signal. and historical characteristics Processing is performed to obtain the reset door feature. And send it to the candidate hidden state calculation unit to reset the gate feature. The calculation formula is as follows:

[0043] (1)

[0044] in, This represents the Sigmoid activation function.

[0045] Using a reset gate can effectively adjust historical features. Hidden states for subsequent candidates The impact.

[0046] Candidate hidden state computation unit is used to fuse signal features and reset door features Generate candidate hidden states And send it to the hidden state fusion unit, candidate hidden state The calculation formula is as follows:

[0047] (2)

[0048] in, This represents the dot product operation. This represents the hyperbolic tangent activation function.

[0049] Update gates are used to update signal features and historical characteristics Processing is performed to obtain the updated gate features. And send it to the hidden state fusion unit to update the gate feature. The calculation formula is as follows:

[0050] (3)

[0051] The update gate feature is used to adjust the fusion ratio between historical hidden states and candidate hidden states.

[0052] The hidden state fusion unit is used to fuse the hidden states from the previous time step. Candidate hidden state and update gate features Calculate the current hidden state :

[0053] (4)

[0054] The above formulas can be used to further achieve a collaborative representation of fault identification information and time dependence.

[0055] Generally speaking, energy signals can characterize the differences in timing monitoring signals under different faults and can be used as effective fault discrimination information. Figure 3 It is the probability density diagram of the three-phase current timing signal under mean statistics. Figure 3 The mean square value is used as an example to reflect the signal energy intensity, illustrating the probability density of different faults, where each sub-graph corresponds to a phase channel. From Figure 3 As can be seen, within the same phase current, the distribution of different faults exhibits significant differences in both center location and morphology, demonstrating that the energy representation of the signal can reflect different fault categories. Since the monitoring signal of a power converter exhibits obvious time-dependent characteristics when an open-circuit fault occurs, extracting fault discrimination information and timing features from the signal is the core challenge for achieving high-performance diagnosis of open-circuit faults in power converters. Therefore, this embodiment designs a discriminative wavelet convolution function. The expression for enhancing the energy representation in the monitoring signal is as follows:

[0056] (5)

[0057] in, , , This represents the learnable parameters.

[0058] To ensure end-to-end stable optimization of the model, this embodiment addresses the three learnable parameters in the above formula. , , The differentiability is verified, and the specific derivation process is as follows:

[0059] (6)

[0060] (7)

[0061] (8)

[0062] From the above three formulas, it can be seen that the discriminant wavelet convolution with respect to parameters... , , Both are differentiable, verifying the feasibility of the proposed discriminative wavelet convolution function.

[0063] S102: Source Model Training

[0064] Acquire source data sample set of power converter Each data sample includes the timing monitoring signal of the power converter and the corresponding open-circuit fault category label. The source data sample set is used to train the power converter open-circuit fault diagnosis model to obtain the source model parameters. And obtain the distribution of model parameters as the prior distribution. .

[0065] S103: Variational Memory Retention Learning under New Tasks:

[0066] To overcome the limitations of historical data preservation, this invention proposes a novel variational memory retention strategy (VMP) from the perspective of Bayesian optimization, which effectively solves the problem of "variance shrinkage".

[0067] Regarding the two task data provided sequentially and Model parameters The posterior distribution can be represented in Bayesian terms as follows:

[0068] (9)

[0069] Generally speaking, there is no correlation between task data, which means and They are mutually independent. Therefore, the above formula can be further expressed as follows:

[0070] (10)

[0071] From the above formula, we can see that when Task 2 is trained, This serves as the new prior distribution, meaning the posterior distribution of Task 1 is the prior distribution of Task 2. Furthermore, With model parameters Since it is irrelevant, it can be considered a constant, and thus equation (10) can be further expressed as:

[0072] (11)

[0073] in, It indicates a proportional relationship.

[0074] Therefore, we can conclude that the posterior distribution is evaluated. Equivalent to evaluating prior distribution Similar to a variational autoencoder strategy, a distribution can be learned through a neural network. To make it approximate the prior distribution Thus, the posterior distribution can be determined. An approximate estimate. The Kullback-Leibler (KL) divergence is introduced to evaluate... and Differences between them:

[0075] (12)

[0076] in This represents the expectation. According to Bayes' theorem, It can be written in the following form:

[0077] (13)

[0078] in Indicates the purpose of training tasks The distribution of model parameters.

[0079] Substituting equation (13) into equation (12) yields:

[0080] (14)

[0081] Equation (14) consists of the following four parts: 1) The first part is used to measure the distribution of network learning in Equation (12). With the prior distribution in equation (13) The differences; the second and third parts indicate about The negative expected log-likelihood; the fourth part is a constant that can be ignored during parameter updates. Therefore, equation (14) can be expressed as the following loss function. :

[0082] (15)

[0083] in, , These represent the classification losses for tasks 1 and 2, respectively. This represents the weight of the KL loss term, used to balance learning for the current task and retaining past knowledge.

[0084] make For the first The classification loss for each task can be calculated using the cross-entropy classification loss formula, as follows:

[0085] (16)

[0086] in, Indicates the first Data samples from each task Open circuit fault category The true label, , Indicates the first Number of data samples for each task , Indicates the number of open-circuit fault categories. Indicates the first Data samples from each task Open circuit fault category The predicted probability.

[0087] It is worth noting that in a lifelong learning scenario, when learning Task 2, the knowledge from Task 1 can be retained, that is... If the value is close to 0, then equation (15) can be rewritten as:

[0088] (17)

[0089] Furthermore, the above formula can be generalized to two consecutive tasks. and When learning the task At that time, rewriting equation (17) yields the task. loss function :

[0090] (18)

[0091] in, , They represent the training of the first and second training sessions. The, the The distribution of model parameters learned for each task. This indicates the solution for the KL divergence. Represents KL loss term The weight, Indicates the first Classification loss for each task.

[0092] Equation (18) shows that the optimization objective achieves the retention of past task knowledge by minimizing the difference between the current parameter distribution and the historical parameter distribution; at the same time, it ensures the model's ability to discriminate new task data through classification loss. Therefore, this method can reduce the dependence on historical storage while taking into account model performance.

[0093] Whenever the power converter's operating conditions change and relearning is required, a data sample set for the current task is collected. , Then, the following variational memory retention learning is used to obtain the open-circuit fault diagnosis model for the power converter. . Figure 4 This is a flowchart of variational memory retention learning in this invention. For example... Figure 4 As shown, the specific steps of variational memory retention learning in this invention include:

[0094] S401: Initialize iteration parameters.

[0095] Let the number of iterations Initialize the parameters of the power converter open-circuit fault diagnosis model. ;

[0096] S402: Sample the current batch of data;

[0097] From the data sample set Sample the current batch of data. .

[0098] S403: Calculate the loss function and update the model parameters:

[0099] Data sample set Each data sample is input into the power converter open-circuit fault diagnosis model. Perform open-circuit fault category prediction, and then calculate the loss function. And update the obtained model parameters Loss function The calculation formula is as follows:

[0100] (19)

[0101] in, Indicates the first The distribution of model parameters learned for each task. Indicates the first The distribution of model parameters learned for each task This indicates the solution for the KL divergence. Represents KL loss term The weight, Indicates the first Classification loss for each task.

[0102] Model parameters The update formula is as follows:

[0103] (20)

[0104] in, This represents the learning rate.

[0105] Existing methods typically employ a Gaussian distribution. The parameters are modeled, but due to the inadequacy of this distribution, this embodiment attempts to introduce the von Mises-Fisher (vMF) distribution. Figure 5 This is a schematic diagram illustrating the alignment of model parameter distributions based on Gaussian and vMF distributions. Figure 5 As shown, Gaussian distributions are prone to variance shrinkage during optimization, leading to gradual distribution degradation and a "false memory" phenomenon, which weakens the model's memory retention capacity. Therefore, this embodiment introduces... Figure 5 von Mises-Fisher (vMF) distribution Model the parameters, where It is the direction of the mean. It is a concentration parameter. Unlike the Gaussian distribution, the vMF distribution is defined on a unit hypersphere and can be expressed through... Characterize the location of parameter distribution and utilize This effectively adjusts the dispersion of the distribution and avoids the problem of "variance shrinkage"; therefore, in equation (18) It can be rewritten as:

[0106] (twenty one)

[0107] in, This indicates the dimension of the output features of the GRU network. , They represent the order as and The modified Bessel formula, , They represent the first The, the Concentration parameters of the von Mises distribution for each task. , They represent the first The, the The mean direction vector of the parameter distribution of each task model.

[0108] To theoretically verify that the vMF distribution can effectively solve the above-mentioned "variance shrinkage" problem, the following mathematical derivation is given. First, equation (19) is rewritten as:

[0109] (twenty two)

[0110] (twenty three)

[0111] (twenty four)

[0112] The key to solving the variance shrinkage problem lies in adjusting the central tendency of the parameter distribution. In the vMF distribution, this characteristic is only related to the concentration parameter. Relevant. Therefore, when analyzing the impact of KL divergence on the degree of distribution concentration, without loss of generality, we can assume that the mean direction parameters of adjacent tasks are basically aligned, i.e. Then equation (22) can be simplified to:

[0113] (25)

[0114] At this point, the above equation regarding The derivative can be expressed as follows:

[0115] (26)

[0116] Combining the recurrence relation of the modified Bessel formula with equation (25), the above equation can be further simplified to:

[0117] (27)

[0118] in The larger the value, the more concentrated the distribution. Monotonically increasing, Prior distribution parameters It is a fixed value. If Equation (25) is monotonically decreasing. Increasing the value will align the current distribution with the prior distribution; if Equation (25) is monotonically increasing. Increasing the variance will result in a stronger KL divergence penalty. Therefore, the VMP strategy can effectively avoid the variance shrinkage problem.

[0119] Based on the above analysis and equations (18) and (22), this embodiment addresses the task... The loss function can be written as:

[0120] (28)

[0121] S404: Determine whether the training end condition for the current task has been met. If it has, the training for the current task ends; otherwise, proceed to step S405.

[0122] S405: Set the iteration count Return to step S402;

[0123] S104: Open-circuit fault diagnosis under new tasks:

[0124] When the task When the power converter needs to be diagnosed for open-circuit faults under the corresponding operating conditions, timing monitoring signals are collected and input into the power converter open-circuit fault diagnosis model. In the process, the open circuit fault diagnosis results were obtained.

[0125] To better illustrate the technical effects of the present invention, specific examples are used to experimentally verify the invention. This embodiment employs a self-built hardware experimental platform, which simulates open-circuit faults and dynamic characteristics of power converters under actual industrial operating conditions. The hardware experimental platform in this embodiment consists of several key subsystems, including a power grid simulation power supply, a Tektronix MDO34 oscilloscope, an electronic load, a power circuit, an industrial control computer, and a control sampling system. During the experiment, the power grid simulation power supply and the electronic load jointly adjust the input voltage and dynamic impedance to simulate various operating conditions from low to high load power. Under these conditions, six types of single-phase open-circuit faults are generated by precisely controlling the switching actions in the power circuit. Table 1 is the fault type information table in this embodiment.

[0126] <![CDATA[T1]]> 0 <![CDATA[T4]]> 3 <![CDATA[T2]]> 1 <![CDATA[T5]]> 4 <![CDATA[T3]]> 2 <![CDATA[T6]]> 5

[0127] Table 1

[0128] Finally, the control sampling system collects monitoring signals and stores the acquired fault data in the industrial control computer to construct a dataset. Table 2 shows the key parameters of the experimental platform in this embodiment.

[0129] Rated DC voltage 700V Power grid AC voltage measurement 220V / 50Hz Controller Model DSP283773 Rated power 10kW Filter capacitors 480uF Filter Inductor 4.3mL Sampling frequency 800kHz Control frequency 20kHz

[0130] Table 2

[0131] The collected samples ranged in load power from 3.4 kW to 26.9 kW, covering a wide operating range and accurately reflecting actual industrial conditions, thus facilitating a more reliable evaluation of the proposed method. Using the aforementioned experimental platform, this embodiment designed non-overlapping incremental tasks. Under this setting, the operating power ranges of adjacent tasks were completely separated, and the ratio of the training set to the test set was 4:1. Table 3 shows the task partitioning table for the non-overlapping incremental tasks in this embodiment.

[0132] 1 5.4-7.0kW 3 9.1-12.4kW 2 7.07-9.0kW 4 12.8-26.9kW

[0133] Table 3

[0134] To comprehensively evaluate the diagnostic performance and memory retention capability of the proposed method in lifelong learning fault diagnosis scenarios, this paper adopts three general evaluation metrics:

[0135] Classification accuracy (Acc, the higher the value, the better) is used to evaluate diagnostic performance.

[0136] The Forgotten Measure (FGT, lower is better) is used to evaluate the average performance degradation of the model across all learned diagnostic tasks. After training is complete, FGT can be represented as:

[0137]

[0138] in, Indicating in diagnostic tasks The highest accuracy rate achieved. Indicating in the task After training, the model performs well on the task. The current accuracy rate.

[0139] Average incremental accuracy (AIA, higher is better) is used to evaluate the overall performance of the model throughout the lifelong learning diagnostic process, and it is defined as follows:

[0140]

[0141] in It refers to the number of tasks.

[0142] The effectiveness of the present invention was then verified based on the overlapping incremental tasks in Table 3, and compared with several typical methods under the same experimental settings. The comparison methods included deep learning methods DualGaT, LTCD, and TCN-SE, as well as lifelong learning methods Memory-Aware Synapse (MAS) and Elastic Weight Fixation (EWC). MAS assesses the importance of model parameters based on their impact on the output and mitigates catastrophic forgetting by constraining important parameters; EWC introduces a regularization term to penalize changes in important parameters. Table 4 compares the experimental results of the present invention and the comparison methods in this embodiment under non-overlapping incremental tasks.

[0143]

[0144] Table 4

[0145] As shown in Table 4, under non-overlapping settings, this invention maintains a high Acc value (94.69%), a low FGT value (1.76%), and a high AIA performance (99.33%) on Task 1. These results demonstrate that the proposed method not only effectively mitigates interference between new and old tasks and reduces the forgetting of early task information, but also maintains high overall diagnostic performance during continuous learning, thus validating the effectiveness of the method. Furthermore, the proposed method achieves the aforementioned performance improvements without requiring additional storage of historical data, thereby reducing the data storage burden in practical engineering applications.

[0146] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A memory retention based power converter open circuit fault lifelong learning diagnostic method, characterized by, Includes the following steps: S1: Construct an open-circuit fault diagnosis model for the power converter, including a GRU network and a classifier, where: GRU network for time series monitoring signals of a power converter performing feature extraction, and sending the obtained time series fault features to a classifier, denotes a time variable; The classifier is used to predict the current open-circuit fault category of the power converter based on the characteristics of the time-series faults; S2: Acquire source data sample set from the power converter Each data sample includes the timing monitoring signal of the power converter and the corresponding open-circuit fault category label, using the source data sample set. The open-circuit fault diagnosis model for the power converter is trained to obtain the source model parameters. And obtain the distribution of model parameters as the prior distribution. ; S3: Whenever the power converter's operating conditions change and relearning is required, collect the data sample set for the current task. , Then, the following variational memory retention learning is used to obtain the open-circuit fault diagnosis model for the power converter. : S3.1: Let the number of iterations be... Initialize the parameters of the power converter open-circuit fault diagnosis model. ; S3.2: From the data sample set Sample the current batch of data. : S3.3: Data sample set Each data sample is input into the power converter open-circuit fault diagnosis model. Perform open-circuit fault category prediction, and then calculate the loss function. And update the obtained model parameters loss function The calculation formula is as follows: ， in, Indicates the first The distribution of model parameters learned for each task. Indicates the first The distribution of model parameters learned for each task This indicates the solution for the KL divergence. Represents KL loss term The weight, Indicates the first Classification loss for each task; S3.4: Determine whether the training end condition for the current task has been met. If it has, the training for the current task ends; otherwise, proceed to step S3.

5. S3.5: Let the number of iterations be... Return to step S3.2; S4: When the task When the power converter needs to be diagnosed for open-circuit faults under the corresponding operating conditions, timing monitoring signals are collected and input into the power converter open-circuit fault diagnosis model. In the process, the open circuit fault diagnosis results were obtained.

2. The power converter open-circuit fault diagnosis method according to claim 1, characterized in that, The open-circuit faults include single-tube faults and dual-tube faults.

3. The power converter open-circuit fault diagnosis method according to claim 1, characterized in that, The timing monitoring signal is a three-phase current timing signal.

4. The power converter open-circuit fault diagnosis method according to claim 1, characterized in that, The data points of each timing monitoring signal need to cover at least two fault cycles.

5. The lifelong learning diagnostic method for open-circuit faults in power converters according to claim 1, characterized in that, The GRU network includes wavelet convolutional layers, 1×1 convolutional layers, reset gates, update gates, candidate hidden state calculation units, and hidden state fusion units, wherein: Wavelet convolutional layers are used for time-series monitoring signals. Perform wavelet convolution to obtain signal features And send it to the reset gate to update the gate and candidate hidden state calculation unit; A 1×1 convolutional layer is used to hide the hidden state from the previous time step. Generate historical features by performing 1×1 convolution operations. And send it to the reset gate to update the gate and candidate hidden state calculation unit; The reset gate is used to characterize the signal. and historical characteristics Processing is performed to obtain the reset door feature. And send it to the candidate hidden state calculation unit to reset the gate feature. The calculation formula is as follows: ， in, This represents the Sigmoid activation function; Candidate hidden state computation unit is used to fuse signal features and reset door features Generate candidate hidden states And send it to the hidden state fusion unit, candidate hidden state The calculation formula is as follows: ， in, This represents the dot product operation. Represents the hyperbolic tangent activation function; Update gates are used to update signal features and historical characteristics Processing is performed to obtain the updated gate features. And send it to the hidden state fusion unit to update the gate feature. The calculation formula is as follows: ； The hidden state fusion unit is used to fuse the hidden states from the previous time step. Candidate hidden state and update gate features Calculate the current hidden state : 。 6. The lifelong learning diagnostic method for open-circuit faults in power converters according to claim 5, characterized in that, The discriminant wavelet convolution function used in the wavelet convolution layer as follows: ， in, , , This represents the learnable parameters.

7. The lifelong learning diagnostic method for open-circuit faults in power converters according to claim 1, characterized in that, The classification loss The cross-entropy classification loss is used, and the calculation formula is as follows: ， in, Indicates the first Data samples from each task Open circuit fault category The true label, , Indicates the first Number of data samples for each task , Indicates the number of open-circuit fault categories. Indicates the first Data samples from each task Open circuit fault category The predicted probability.

8. The lifelong learning diagnostic method for open-circuit faults in power converters according to claim 1, characterized in that, The KL loss item The calculation formula is as follows: ， ， ， in, This indicates the dimension of the output features of the GRU network. , They represent the order as and The modified Bessel formula, , They represent the first The, the Concentration parameters of the von Mises distribution for each task. , They represent the first The, the The mean direction vector of the parameter distribution of each task model.

Images