A method for fault diagnosis of energy storage converter based on semi-supervised domain adaptive learning

Abstract

This invention discloses a fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning, belonging to the field of power electronic equipment condition monitoring technology. First, a quantitative function model is established through IGBT accelerated aging experiments. Then, source domain simulation and target domain experimental fault datasets are constructed. After preprocessing and multi-domain feature extraction, the source domain is divided into a source domain and a target domain containing a small number of labeled samples and a large number of unlabeled samples. A model consisting of a feature extractor and a classifier is constructed. The model is optimized by combining supervised training with cross-entropy loss and adversarial training with gradient inversion and entropy regularization loss. After evaluating and saving the optimal parameters, the system outputs the fault category and location based on the input signal to be diagnosed. The system includes seven modules, including aging experiments and dataset construction, which work together to implement the above method. This invention integrates the advantages of simulation and experimental data, achieves cross-domain feature alignment, fully utilizes unlabeled samples, and significantly improves the model's diagnostic accuracy and generalization ability.

Description

Technical Field

[0001] This invention belongs to the field of power electronic equipment condition monitoring technology, specifically relating to a fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning. Background Technology

[0002] Power conversion systems (PCS) are the core components of energy storage systems, and their reliability directly impacts the stability and security of the power grid. Insulated-gate bipolar transistors (IGBTs), as the main power switching devices in the converter, are subjected to long-term electrical and thermal stresses, making them a high-risk area for failure. Their failure modes mainly include sudden open-circuit and short-circuit faults, as well as gradual performance degradation due to aging. Traditional fault diagnosis methods are mostly based on threshold judgment or simple signal analysis, which are insufficient to handle early and subtle faults under complex operating conditions.

[0003] In recent years, data-driven deep learning-based diagnostic methods have shown great potential. However, these methods face two major challenges in practical engineering applications: First, obtaining fault data for energy storage converters in actual operating systems is difficult, and directly acquiring fault samples of large-capacity energy storage converters through experimental simulation is very costly. Second, the models lack generalization ability. Models trained solely on simulation data often perform poorly on actual devices due to the differences between simulation and experiment. Existing technologies attempt to combine simulation and experiment, but these often suffer from problems such as unreasonable data partitioning leading to weak feature engineering or model hyperparameters, and reliance on human experience, affecting the reliability, reproducibility, and engineering practicality of the diagnostic solutions.

[0004] Furthermore, the aging of IGBT devices is divided into chip-level aging and package-level aging. Package-level aging involves the progressive degradation process of various materials inside the power device. Its mechanism includes voids in the solder layer under electrothermal cycling stress, fatigue of the bonding wires, and failure of the oxide insulation inside the cell. Chip junction temperature is one of the key parameters. Since chip junction temperature is difficult to measure directly online inside the package, traditional temperature-based aging monitoring methods have limited accuracy and are difficult to implement.

[0005] If model training relies solely on labeled fault samples collected from experimental environments, the limited number of available labeled samples makes it difficult to meet the training sample scale requirements of data-driven models. Introducing fault samples generated in simulation environments as supplementary data presents significant differences in feature structure and data distribution between simulation models and experimental devices due to differences in parameters, boundary conditions, and noise characteristics. Furthermore, experimental samples are influenced by complex factors such as the response of actual measurement and control systems and sensor characteristics. Therefore, when collaboratively utilizing simulation and experimental samples to construct a unified diagnostic model, it is crucial to address the issues of feature alignment and cross-domain distribution adaptation between simulation and experimental data. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning. This method establishes a semi-supervised domain adaptive learning framework, integrates operational fault data obtained from simulation models with operational fault data obtained from small-capacity experimental simulations, and drives the model to learn discriminative features of different operating conditions through gradient inversion and entropy optimization adversarial learning mechanisms. Under the condition of scarce target domain labels, this method improves the diagnostic accuracy and generalization ability for multiple types of faults in energy storage converters, thus solving the technical problems of scarce and costly experimental samples for energy storage converter faults and insufficient generalization ability of diagnostic models.

[0007] The present invention adopts the following technical solution: A fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning includes the following steps: S1. Construct an IGBT accelerated aging test platform, apply periodic electrothermal stress to the IGBT device, and collect the collector-emitter saturation voltage drop of the IGBT device. Aging data of IGBT devices are obtained by using drift data of accumulated stress time; based on the aging data, a quantitative function model describing the changes of key parameters of IGBT with equivalent aging time is established. S2. Based on the quantitative function model extracted in step S1, establish a digital simulation model of the energy storage converter, inject IGBT faults into the simulation model, collect fault waveform data, and construct a fully labeled source domain simulation fault dataset; build an experimental prototype of the energy storage converter, install the IGBT devices aged in step S1 on the experimental prototype, reproduce the IGBT faults, collect measured fault waveform data, and construct a target domain experimental fault dataset. S3. Preprocess and extract features from the current and voltage signals in the source domain simulation fault dataset and the target domain experimental fault dataset to obtain a feature matrix; use the fully labeled source domain simulation fault dataset as the source domain dataset, and divide the target domain experimental fault dataset into labeled samples and unlabeled samples, which together serve as the target domain dataset. S4. Construct a fault diagnosis model consisting of a feature extractor G and a classifier F1: the feature extractor G is used to map the input feature vector to a high-dimensional feature space; the classifier F1 is connected after the feature extractor G and is used to map the high-dimensional features to fault categories. S5. Merge labeled samples from the source domain and labeled samples from the target domain as training data, obtain prediction results through feature extractor G and classifier F1, calculate classification loss using cross-entropy loss function and perform backpropagation to update the parameters of feature extractor G and classifier F1; for unlabeled samples in the target domain, obtain prediction results through feature extractor G and classifier F1, calculate the entropy of the prediction distribution as entropy regularization loss, and multiply the gradient of feature extractor G by a negative coefficient through gradient inversion layer during backpropagation to minimize the entropy regularization loss and update the parameters of feature extractor G. S6. During training, periodically evaluate the current model using the target domain test set and save the parameters of the model with the best performance. S7. Input the real-time electrical signal of the energy storage converter to be diagnosed into the trained fault diagnosis model, and output the fault category and location prediction results.

[0008] Preferably, in step S1, applying periodic electrothermal stress to the IGBT device specifically includes: The IGBT device is controlled to periodically switch between the on and off states, so that the junction temperature of the device fluctuates cyclically between a preset high temperature threshold and a low temperature threshold. The on-state voltage drop between the collector and emitter of the IGBT device is monitored and recorded in real time until the increase of the on-state voltage drop relative to the initial value reaches the preset failure criterion. The experiment is then stopped and complete aging trajectory data is obtained. Establishing a quantitative functional model describing the changes of key IGBT parameters with equivalent aging time specifically includes: The collected aging data were fitted and analyzed to determine the nonlinear mapping relationship between the conduction voltage drop drift and the cumulative stress time or the number of switching cycles. Construct an IGBT physical model or behavioral model that includes aging factors, and embed the nonlinear mapping relationship into the model parameters so that the electrical characteristics output by the model can change dynamically with the increase of equivalent aging time.

[0009] Preferably, in step S2, injecting IGBT faults into the simulation model specifically includes: Simulating IGBT open-circuit fault: By modifying the drive signal of the IGBT device in the simulation model, the gate drive signal is forced to be low or the current path is disconnected. Simulating IGBT short-circuit fault: By directly short-circuiting the collector and emitter of the IGBT device in the simulation model; Simulate device performance degradation faults: Call the quantitative function model established in step S1 to update the on-resistance or on-voltage drop parameters of the IGBT device in the simulation model in real time.

[0010] Preferably, the energy storage converter prototype adopts a three-level NPC topology. The reproduction of IGBT faults includes reproducing IGBT open-circuit faults by changing the drive signal and reproducing IGBT short-circuit faults by combining the short-circuit protection function of the drive chip. The collected measured fault waveform data includes transient waveform data and steady-state waveform data before and after the fault occurs.

[0011] Preferably, in step S3, the preprocessing includes DC bias removal processing of the current signal and robust normalization processing based on median and interquartile range. The feature extraction includes extracting signal features in the time domain, frequency domain, wavelet domain, and Hilbert transform domain respectively. After concatenating the extracted features into a feature vector, Z-score normalization is performed to obtain the feature matrix.

[0012] Preferably, in step S4, the feature extractor G adopts a network structure with three modules connected in series. The first and second modules each contain fully connected transformation, batch normalization, ReLU activation function and Dropout mechanism in sequence, and the third module contains fully connected transformation, batch normalization and ReLU activation function in sequence. The classifier adopts a convolutional neural network structure, which includes a preprocessing layer, two one-dimensional convolutional layers, a flattening layer and two fully connected layers in sequence.

[0013] Preferably, in step S5, multiplying the gradient of the feature extractor by a negative coefficient through the gradient inversion layer specifically includes: During the forward propagation, the gradient inversion layer performs an identity mapping operation, transferring the features output by the feature extractor to the classifier without loss. During backpropagation, the gradient inversion layer multiplies the gradient returned from the classifier by a negative adaptive coefficient and passes the gradient multiplied by the negative coefficient to the feature extractor. This allows the feature extractor to optimize its parameters in a way that maximizes the difficulty for the classifier to distinguish between samples in the source and target domains, thereby achieving the extraction of domain-invariant features.

[0014] Preferably, the formula for calculating the entropy regularization loss is:

[0015] in, Weights for entropy loss. The expected entropy value of the predicted distribution of unlabeled samples in the target domain.

[0016] Preferably, in step S6, the periodic evaluation of the current model using the target domain test set specifically includes: Set an evaluation interval step number, and input labeled samples from the target domain test set into the current model every time the evaluation interval step number is reached; The model's overall accuracy, macro-average F1 score, precision, and recall for each fault category are calculated on the target domain test set. The current performance metrics are compared with the historical best performance metrics. If the current performance is better than the historical best performance, the current model parameters and corresponding performance records are overwritten and saved.

[0017] Secondly, embodiments of the present invention provide a fault diagnosis system for energy storage converters based on semi-supervised domain adaptive learning, comprising: The aging test module is used to build an accelerated aging test platform for IGBTs, apply periodic electrothermal stress to the IGBT devices, and collect the collector-emitter saturation voltage drop of the IGBT devices. Aging data is obtained by using drift data with accumulated stress time, and a quantitative function model describing the changes of key IGBT parameters with equivalent aging time is established based on the aging data. The dataset construction module, connected to the aging experiment module, is used to establish a digital simulation model of the energy storage converter based on the quantitative function model, inject IGBT faults and collect fault waveform data to construct a source domain simulation fault dataset. It is also used to build an energy storage converter experimental prototype, install aged IGBT devices and reproduce IGBT faults, and collect measured fault waveform data to construct a target domain experimental fault dataset. The data processing module, connected to the dataset construction module, is used to preprocess and extract features from the current and voltage signals in the source domain simulation fault dataset and the target domain experimental fault dataset to obtain a feature matrix. It is also used to set the source domain simulation fault dataset as the source domain dataset and to divide the target domain experimental fault dataset into labeled samples and unlabeled samples and set them together as the target domain dataset. The model building module, connected to the data processing module, is used to build a fault diagnosis model consisting of a feature extractor G and a classifier F1. The feature matrix is ​​input into the fault diagnosis model, the feature extractor G maps the input feature vector to a high-dimensional feature space, and the classifier F1 maps the high-dimensional features to fault categories. The model training module, connected to the model building module, is used to merge labeled samples from the source domain and labeled samples from the target domain as training data, obtain prediction results through feature extractor G and classifier F1, calculate classification loss using the cross-entropy loss function and backpropagate to update the parameters of feature extractor G and classifier F1, and also to obtain prediction results for unlabeled samples in the target domain through feature extractor G and classifier F1, calculate the entropy of the prediction distribution as entropy regularization loss, and multiply the gradient of feature extractor G by a negative coefficient through a gradient inversion layer during backpropagation to minimize the entropy regularization loss and update the parameters of feature extractor G. The model evaluation module, which is connected to the model training module and the dataset construction module respectively, is used to periodically evaluate the current model using the target domain test set during the model training process and save the model parameters with the best performance. The fault diagnosis module, connected to the model evaluation module, is used to input the real-time electrical signals of the energy storage converter to be diagnosed into the trained fault diagnosis model and output the fault category and location prediction results.

[0018] Thirdly, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described energy storage converter fault diagnosis method based on semi-supervised domain adaptive learning.

[0019] Fourthly, embodiments of the present invention provide a computer-readable storage medium including a computer program, which, when executed by a processor, implements the steps of the above-described energy storage converter fault diagnosis method based on semi-supervised domain adaptive learning.

[0020] Fifthly, a chip includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described energy storage converter fault diagnosis method based on semi-supervised domain adaptive learning.

[0021] In a sixth aspect, embodiments of the present invention provide an electronic device including a computer program, wherein when the computer program is executed by the electronic device, it implements the steps of the above-described energy storage converter fault diagnosis method based on semi-supervised domain adaptive learning.

[0022] Compared with the prior art, the present invention has at least the following beneficial effects: A fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning is proposed. This method utilizes a minimum-maximum entropy semi-supervised domain adaptive learning approach, integrating simulation and experimental data from different operating conditions to construct a fault sample library. This allows for the establishment and optimization of a fault diagnosis model, overcoming the limitations of single data sources in terms of sample diversity and realism. First, a digital simulation model of the energy storage converter is established, pre-setting typical fault modes for power switching devices. Aging state data of the devices is acquired through an experimental platform, and faults are reproduced. The parameter mapping function of the simulation model is optimized using measured data, forming an experimentally calibrated simulation model. Extended simulation data is generated using this model and used together with experimental data as the training basis. A unified feature extraction process and a data source-based partitioning strategy are employed to construct a dataset. The hyperparameter-optimized diagnostic model is trained, and its performance is verified using independent experimental data, thereby achieving fault diagnosis of the energy storage converter.

[0023] Furthermore, the electrothermal stress application method and quantitative function model establishment method of IGBT accelerated aging experiment were refined. By controlling the cyclic fluctuation of IGBT junction temperature to simulate the actual aging process, the on-state voltage drop drift was used as the aging characteristic. A nonlinear mapping relationship was established through fitting analysis and embedded into the IGBT model. Complete aging trajectory data was obtained through preset failure criteria, so that the established quantitative function model can accurately reflect the change law of key IGBT parameters with aging time. This provides real and accurate aging parameter support for subsequent fault injection in the simulation model, making the aging fault data generated by simulation closer to the fault characteristics of actual devices.

[0024] Furthermore, by modifying the drive signal, shorting devices, and dynamically updating aging parameters, faults are injected from all directions. The constructed source domain simulation fault dataset is not only large in scale, but also truly reflects the electrical characteristics of various faults, enriching the diversity of fault samples. It provides sufficient and realistic sample support for the model to learn the discriminative features of different faults, while reducing the cost and safety risks of experimentally reproducing dangerous faults such as short circuits.

[0025] Furthermore, open-circuit faults are reproduced by modifying the drive signal, and short-circuit faults are safely reproduced using the drive chip's protection function. Simultaneously, transient and steady-state waveforms are collected, closely matching the actual engineering application topology of the energy storage converter, making the experimental data more practically valuable. The safe reproduction of short-circuit faults through the drive chip's protection function solves the safety hazards associated with short-circuit fault reproduction in experiments. Collecting transient and steady-state waveforms before and after a fault compensates for the shortcomings of traditional experiments that only collect steady-state data. This allows the constructed target domain experimental fault dataset to contain more comprehensive fault feature information, providing measured data support for the model to learn the complete characteristics of faults and improving the model's ability to identify real-world faults.

[0026] Furthermore, the preprocessing employs DC bias removal and robust normalization. Feature extraction covers the time domain, frequency domain, wavelet domain, and Hilbert transform domain. The feature matrix is ​​obtained through feature concatenation and Z-score normalization. Robust normalization effectively suppresses the influence of outliers caused by sensor noise and operating condition fluctuations in experiments and simulations. Multi-domain feature extraction fully exploits the statistical features in the time domain, the spectral features in the frequency domain, the time-frequency features in the wavelet domain, and the envelope / phase features in the Hilbert transform domain of current and voltage signals. This enables the extracted features to comprehensively and accurately characterize the fault type, significantly improving the discriminative power of the features.

[0027] Furthermore, the feature extractor adopts a three-module cascaded structure with the addition of a Dropout mechanism. The fault classifier uses a convolutional neural network structure, which works by progressively abstracting and mining the time-frequency features of the input signal through multiple convolutions, followed by feature fusion and dimensionality reduction via fully connected layers, ultimately outputting a highly discriminative feature representation. Specifically, the preprocessing layer normalizes the input through batch normalization to accelerate model convergence. The first convolutional layer extracts the local time-frequency response of the signal, and the second convolutional layer further aggregates these to form higher-level combined features. Subsequently, a flattening operation is used to achieve dimensionality transformation between the convolutional and fully connected layers. The first fully connected layer then performs feature compression and introduces a Dropout mechanism to suppress overfitting. Finally, the second fully connected layer maps the features to a specified dimension for output. This architecture fully utilizes the local correlation of the signal, combined with regularization techniques such as batch normalization and Dropout, effectively reducing the risk of overfitting and improving the model's diagnostic accuracy and generalization ability under target conditions.

[0028] Furthermore, by implementing an adversarial game between the feature extractor and the classifier through gradient inversion, the feature extractor is optimized in the direction of maximizing the difficulty of the classifier to distinguish between different domains. This allows the feature extractor to learn domain-invariant features common to both the source and target domains, achieving cross-domain alignment of feature distributions. The adversarial mechanism of the gradient inversion layer enables the feature extractor to no longer rely on domain-related features, but instead learn domain-invariant features that can simultaneously characterize simulation and experimental faults. This effectively reduces the difference in feature distribution between simulation and experimental data, allowing the model to achieve accurate fault classification even on unlabeled samples in the target domain. This significantly improves the model's cross-domain generalization ability and fully utilizes the value of unlabeled samples in the target domain.

[0029] Furthermore, by leveraging entropy regularization loss to drive the model to maximize the prediction uncertainty of unlabeled samples in the target domain, and combining it with a gradient reversal layer to enable the feature extractor to generate more generalizable features, we achieve synergistic optimization of supervised training and adversarial training. This solves the problem of unlabeled samples being difficult to utilize effectively in semi-supervised learning. Entropy regularization loss can effectively mine fault feature information in unlabeled samples in the target domain. By flexibly adjusting the intensity of adversarial training through weight λ, the model can dynamically balance the proportion of supervised training and adversarial training according to the training process, avoiding excessive adversarial training that could cause the model to lose its fault classification ability.

[0030] Furthermore, by conducting real-time monitoring of model training effects through comprehensive performance evaluation, avoiding model overfitting through interval evaluation, and saving optimal parameters to ensure optimal performance of the final model, we can achieve refined management and control of model training. By comparing historical performance and saving optimal model parameters, we can ensure that the final model used for fault diagnosis has optimal performance in the target domain.

[0031] It is understood that the beneficial effects of the second to sixth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0032] In summary, this invention constructs a highly realistic multi-source fault dataset through precise IGBT aging simulation and fault reproduction. Combining multi-domain feature extraction and deep network models, it achieves cross-domain feature alignment through semi-supervised domain adaptive learning with gradient inversion and entropy regularization. It makes full use of unlabeled samples, effectively solves the problems of scarce experimental samples and domain offset, and significantly improves the model's fault diagnosis accuracy and generalization ability in actual working conditions.

[0033] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the overall process of the fault diagnosis method of the present invention. Figure 2 This is a flowchart of the fault data processing and feature extraction process for the energy storage converter of the present invention. Figure 3 This is a structural diagram of the energy storage converter fault diagnosis model of the present invention; Figure 4 This is a flowchart of the training process for the semi-supervised domain adaptive diagnostic model of the energy storage converter of the present invention; Figure 5 A schematic diagram of a computer device provided in an embodiment of the present invention; Figure 6 This is a block diagram of a chip provided according to an embodiment of the present invention.

[0035] Among them, 60. Computer equipment; 61. Processor; 62. Memory; 63. Computer program; 600. Electronic device; 610. Processing unit; 620. Storage unit; 6201. Random access memory unit; 6202. Cache memory unit; 6203. Read-only memory unit; 6204. Program / utility; 6205. Program module; 630. Bus; 640. Display unit; 650. Input / output interface; 660. Network adapter; 700. External device. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0038] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0039] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0040] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0041] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0042] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0043] This invention provides a fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning. An accelerated aging experimental platform for IGBTs is built, applying periodic electrothermal stress to the IGBT devices to accelerate their performance degradation. The drift trajectory of parameters such as saturation voltage drop is monitored online to obtain IGBT device aging data. Then, a digital simulation model of the energy storage converter is constructed based on the measured aging parameters to obtain waveform data under different operating / fault conditions. A small-capacity energy storage converter experimental platform is built to simulate short-circuit and open-circuit faults and aging states of the IGBT devices, collecting various characteristic waveforms to form a fault database. A semi-supervised domain adaptive learning framework is established, using the energy storage converter simulation data as the source domain and the experimental data as the target domain. A feature extractor and a fault classifier are constructed, and adversarial training is performed using a gradient inversion layer and entropy optimization strategy. This learns domain-invariant feature representations that are insensitive to changes in operating conditions, optimizing the diagnostic model parameters and achieving fault diagnosis applicable to the on-site operating conditions of energy storage power stations. This invention is based on semi-supervised adaptive learning and effectively utilizes simulation and experimental fault samples to construct a fault diagnosis model for energy storage converters suitable for actual working conditions, solving the problems of difficulty in obtaining field fault samples and high experimental simulation costs for energy storage converters.

[0044] Please see Figure 1 The present invention discloses a fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning, comprising the following steps: S1. IGBT Accelerated Aging Experiment and Aging Law Extraction: An IGBT accelerated aging experimental platform was built, consisting of an aging main circuit, cooling device, sampling circuit, drive circuit, data acquisition card and online monitoring system. By applying periodic electrothermal stress, the bonding wire fracture and packaging material fatigue of the device were accelerated.

[0045] The main circuit consists of a DC current source, a fuse, an IGBT device, and an electronic load, forming the power flow loop for the IGBT device under test.

[0046] The sampling circuit acquires the collector-emitter voltage of the IGBT and the current flowing through the IGBT device under test through a Hall sensor; the temperature of the IGBT device is measured by attaching a PT1000 platinum resistance thermometer to the surface of the IGBT device, and then converted into a voltage signal by the conditioning circuit.

[0047] The driving circuit is mainly responsible for transmitting the driving signal to the gate of the IGBT device after optocoupler isolation, so as to realize the turn-on and turn-off of the IGBT device.

[0048] The data acquisition card is mainly responsible for transmitting the signals from the sampling circuit to the host computer, and receiving instructions from the host computer to transmit drive signals to the drive circuit.

[0049] The online monitoring system primarily uses the control program and data acquisition card sampling information to generate real-time temperature curves, on-state voltage drop curves, alarms for abnormal operating conditions, and data storage. It also automatically adjusts the drive signal for the IGBT devices based on temperature data and control strategies. Simultaneously, it uses the collector-emitter saturation voltage drop... As a characteristic parameter, its drift data with accumulated stress over time is periodically collected to obtain the aging trajectory.

[0050] The cooling system consists of a thermoelectric cooler, a water pump, a fan, and a water tank. The IGBT devices are tightly bonded to the cooling system using thermal grease. Upon receiving a heat dissipation command, the fan starts, and the thermoelectric cooler begins to operate, conducting heat from the IGBT devices and transferring it to the water tank via the pump, thus achieving cooling. Based on this experimental data, a quantitative functional model describing the changes in key parameters such as the IGBT on-state resistance with equivalent aging time is established, providing aging data input for subsequent simulations and experiments.

[0051] In the experiment, the IGBT device is turned on, at which point the main circuit is activated, and the IGBT device receives a preset DC current source, causing it to heat up. Once a certain temperature is reached, the online monitoring program issues a shutdown command to the IGBT device based on the set upper temperature limit, disconnecting the main circuit and controlling the cooling device to operate. Cooling is achieved through water cooling and air cooling until the lower temperature limit is reached, at which point the device returns to the heating stage. This power cycle continues, and due to uneven thermal stress distribution, stress accumulates at the bonding wires, eventually causing the bonding wires to warp or break. Under repeated thermal stress cycles, the solder layer gradually develops cracks, which continue to expand, eventually causing the solder layer to detach from the silicon chip. This power cycle achieves the purpose of aging the IGBT device at the package level.

[0052] Accelerated aging experiments are underway. The on-state voltage drop between the collector and emitter of the IGBT device is used as an aging characteristic parameter. The on-state voltage drop under different aging degrees is recorded as an aging criterion, forming a mapping relationship between the on-state voltage drop and aging time, providing data support for subsequent experiments.

[0053] S2, Construction of energy storage converter fault dataset; S201. Simulation Modeling, Fault Injection, and Source Domain Dataset Generation: Based on the aging characteristic change function extracted in step S1, a digital simulation model of the energy storage converter is established. The simulation model includes the main circuit and control system. The main circuit consists of a battery model, a DC regulated capacitor, a three-level NPC circuit, an LCL filter circuit, and an AC power grid. The control strategy adopts a dual closed-loop control with an outer loop for AC active and reactive power and an inner loop for inductor current.

[0054] In the simulation model, open-circuit faults of IGBTs are simulated by modifying the drive signal; short-circuit faults are simulated by shorting devices; and different degrees of performance degradation of the devices are simulated by embedding an aging function model of the IGBT devices. The system collects waveform data such as DC-side and AC-side current and voltage; then, the position and number of the IGBT devices under test in the topology are changed, and different fault behaviors are set to realize a complete simulation fault database; finally, the waveform features collected under different bridge arms, different positions, different numbers, and different faults in the simulation are integrated to construct a source domain simulation fault dataset, which provides a large number of samples for the fault diagnosis model.

[0055] S202. PCS Experimental Platform Setup, Fault Reproduction, and Target Domain Dataset Generation: An experimental prototype of the energy storage converter was built for physical verification and data acquisition. The specific architecture of this experimental prototype is as follows: Its main circuit adopts a three-level NPC topology. The DC side is powered by an adjustable regulated power supply, and after passing through a DC support capacitor, it is connected to a three-phase inverter bridge composed of IGBT modules. The inverter output is connected to a three-phase symmetrical load through an LCL filter.

[0056] The core of the experimental platform is a DSP-based digital controller, which is responsible for generating PWM signals according to the dual closed-loop control strategy, protecting system logic, and communicating with the host computer.

[0057] The acquisition of the main electrical quantities is accomplished by a measurement system consisting of voltage / current sensors, signal conditioning circuits, and a high-speed data acquisition card. It can simultaneously capture transient and steady-state waveforms before and after a fault occurs, providing measured data for constructing the target domain dataset.

[0058] Fault reproduction experiments were conducted using an experimental prototype. First, IGBT devices aged through the S1 process were installed on the experimental platform, and their operational data under known degradation conditions were collected. Then, faults were reproduced on the platform: an open-circuit fault was induced in the IGBT device by changing the drive signal, and steady-state waveforms were recorded; a short-circuit fault was safely triggered by changing the drive signal in conjunction with the short-circuit protection function of the drive chip, and transient waveforms were captured. Next, the position and number of the IGBT devices under test in the topology were changed, and different fault behaviors were set to create a complete experimental fault database.

[0059] Finally, the measured waveform features collected from different bridge arms, positions, quantities, and faults in the simulation were integrated to construct a target domain experimental fault dataset for performance evaluation and generalization capability verification of the fault diagnosis model.

[0060] S3. Data Preprocessing and Feature Extraction: Acquire AC and DC current and voltage signals of the energy storage converter under various operating conditions, and use the fully labeled simulation data as the source domain dataset. The experimental data was divided into a small number of labeled samples and a large number of unlabeled samples, which together served as the target domain dataset. For the energy storage converter fault data processing and feature extraction workflow, please refer to [link / reference]. Figure 2 .

[0061] In the preprocessing module, DC bias removal is first performed on the current signal of each phase. Then, a robust normalization method based on the median and interquartile range is used to normalize the waveforms of each phase to unify the dimensions and suppress the influence of outliers. To enhance data diversity, data augmentation operations such as limited-amplitude Gaussian noise, amplitude scaling, and time shifting can also be applied to the waveforms of the training set.

[0062] In the feature extraction stage, in the time domain extraction, the mean, standard deviation, root mean square value, skewness, kurtosis, peak-to-peak value, interquartile range, and mean absolute value of each phase current are calculated; in the frequency domain extraction, a fast Fourier transform is performed on the signal to extract its dominant frequency, dominant frequency amplitude, spectral centroid, spectral entropy, and total power; in the wavelet domain extraction, multi-level wavelet decomposition is used to obtain the energy, entropy, and statistical characteristics of the coefficients at each level; in the Hilbert transform domain extraction, the statistical characteristics of the envelope, phase, and instantaneous frequency are calculated by analyzing the signal.

[0063] Furthermore, the features from multiple domains are concatenated sequentially to form a feature vector; the initial feature vector is then standardized using Z-score so that the mean of each feature dimension is 0 and the standard deviation is 1; thus, the final feature matrix is ​​obtained.

[0064] S4. Constructing a fault diagnosis model: Constructing a model using a feature extractor. Classifier For a fault diagnosis model consisting of [various components], please refer to [the relevant documentation]. Figure 3 The structure diagram of the fault diagnosis model for energy storage converter is as follows: S401, Constructing a Feature Extractor Feature extractor A network structure consisting of three concatenated modules is employed. The first and second modules each contain fully connected transformations, batch normalization, ReLU activation, and Dropout mechanisms, respectively. The third module contains fully connected transformations, batch normalization, and ReLU activation, mapping the input features to a high-dimensional feature space. (Classifier) A convolutional neural network structure is employed, consisting of: a preprocessing layer, two one-dimensional convolutional layers, a flattening layer, and two fully connected layers. Feature extractor. and classifier Through collaborative optimization in two phases—supervised training and adversarial training—parameters are updated using cross-entropy loss during supervised training and again using adversarial entropy loss during adversarial training, thereby achieving domain-adaptive fault diagnosis.

[0065] S402, Constructing a classifier It adopts a convolutional neural network structure, which includes a preprocessing layer, two one-dimensional convolutional layers, a flattening layer, and two fully connected layers.

[0066] S403, Feature Extractor and classifier Through coordinated optimization in two phases—supervised training and adversarial training—parameters are updated once using cross-entropy loss during supervised training and again using entropy regularization loss during adversarial training, thus jointly achieving domain adaptive fault diagnosis.

[0067] S5. Model Training: Train the diagnostic model based on supervised learning and entropy optimization. Please refer to [link / reference]. Figure 4 The training flowchart for the semi-supervised adaptive diagnostic model is as follows, with specific training steps: S501. Set training hyperparameters, including maximum number of training steps. Preheating steps Training steps Entropy loss weights ; S502, Initializing Network Parameters and Constructing the Feature Extractor With classifier , respectively, feature extractor and classifier Configure the SGD optimizer; S503. Data preparation stage: Importing source domain data. and target domain data Next, the target domain data is divided into labeled data, unlabeled data, and a test set to provide a data foundation for subsequent training. S504. Model training is divided into two stages: supervised training and adversarial training. The domain adaptability of the fault diagnosis model is achieved through the joint optimization of classification loss and entropy regularization loss. S505. For labeled samples in the source domain and labeled samples in the target domain, first merge them to form training samples, and then use a feature extractor... Obtain feature representations, and classify them. The classification loss is calculated using the cross-entropy loss function after obtaining the prediction results. Then backpropagation is performed, and the gradient flows from the classification loss to the classifier. Reflow to Feature Extractor Update parameters and optimize the feature extractor. and classifier The parameters enable the model to learn basic fault classification capabilities, providing a stable feature representation basis for adversarial training; S506. For unlabeled samples in the target domain, a gradient inversion layer is used to refine the feature extractor during backpropagation. The gradient is multiplied by a negative coefficient so that the feature extractor receives a gradient signal in the opposite direction to the classifier's optimization direction. This is used to calculate the entropy regularization loss in adversarial training. The feature extractor parameters are updated with the goal of minimizing the loss function, thereby maximizing the classifier's prediction distribution entropy. This prompts the feature extractor to generate feature representations that lead to high-uncertainty predictions from the classifier, while the classifier optimizes normally to improve its recognition ability. This achieves an adversarial game between the feature extractor minimizing entropy and the classifier maximizing entropy, where the adaptive coefficients... The intensity of the adversarial training is dynamically adjusted according to the training process. Combined with the warm-up phase and cosine annealing learning rate scheduling strategy, the supervised training and adversarial training are optimized in a coordinated manner.

[0068] The operating conditions of energy storage converters are complex and varied, and the available fault samples under different fault conditions are limited and mostly lack manual labeling. This invention constructs a semi-supervised training data partitioning method: using fully labeled simulation data as the source domain; and dividing experimental data into a small number of labeled samples and a large number of unlabeled samples, which together serve as the target domain. This approach addresses the situation in engineering practice where source domain data is abundant but target domain labels are scarce, making model training settings more realistic.

[0069] To address the challenge of directly using unlabeled samples from the target domain for supervised training, this invention introduces a domain adaptation method based on entropy optimization. This method calculates the entropy of the probability distribution of unlabeled samples in the target domain on the classifier output and optimizes by maximizing this entropy. A gradient reversal layer then backpropagates the corresponding gradients to the feature extractor. , making the feature extractor Learning to generate features that enable classifiers to make high-uncertainty predictions prompts feature extractors to learn and generate feature representations that make classification decision boundaries difficult to distinguish, thereby aligning the feature distributions of the source and target domains and improving the model's classification and discrimination capabilities in the target domain using unlabeled data.

[0070] For feature extractors With classifier To address the complexity of the convergence process during training, this invention employs a learning rate adjustment strategy based on preheating and cosine annealing. This strategy gradually increases the learning rate during the linear preheating phase to stabilize the initial training phase, and then smoothly decreases the learning rate during the cosine annealing phase to promote model convergence. This adjustment method helps to reconcile the classification loss. Optimization and Entropy Regularization Loss Optimize and improve the efficiency and stability of the training process.

[0071] S6. Model Evaluation and Saving: During training, the current model performance is evaluated using the target domain test set every 100 steps. The specific evaluation steps are as follows: S601. Prepare test data, load the target domain test set, and ensure it is isolated from the training data; S602, Forward inference: Input the test samples into the trained model to obtain the prediction results. ; S603, Performance Calculation: Calculates overall accuracy, macro average F1 score, precision, recall, and F1 score for each category; S604. Model saving: Compare the current performance with the historical best performance, and save the parameters of the model with the best performance on the validation set. S605, Training monitoring: Records the training loss curve and learning rate change trajectory, and monitors overfitting. S606. Repeat steps S601-S605 until training is complete, and output the best model and performance report.

[0072] S7. Domain Alignment Quality Evaluation: After training, calculate the domain alignment quality index, including class conditional distribution difference, domain discrimination error, and intra-class and inter-class distance ratio. The specific steps for evaluating the domain alignment quality are as follows: S701, Class Conditional Distribution Difference Calculation: Extract the features of samples belonging to the same fault category in the source domain and the target domain respectively, calculate the maximum mean difference between the two, and average the results of all categories to obtain the average class conditional MMD value.

[0073] in, It's a kernel function. It is the expected value of the kernel function of the samples in X; It is the expectation of the kernel function between sample pairs in Y; It is the expected value of the kernel function for sample pairs between X and Y.

[0074] S702. Domain discrimination error calculation: Input the feature representations of all samples from the source domain and the target domain into the domain classifier. Calculate the error rate of its domain label prediction.

[0075]

[0076] in, n It is the total number of samples; It is a feature processor; It is the input sample; It is a classifier; These are the real domain labels of the samples.

[0077] S703. Calculation of intra-class and inter-class distance ratio: Calculate the average Euclidean distance of features of the same class in the source domain samples as the intra-class distance, calculate the average Euclidean distance between feature centers of different classes as the inter-class distance, and calculate the ratio of intra-class distance to inter-class distance.

[0078] in, It is a collection of categories; It is a category index; It is the first Class sample set; It is a sample feature vector generator; It is the first Sample center.

[0079] This method establishes a technical process from simulation models based on IGBT accelerated aging experimental data, building an experimental platform for fault reproduction, to training and validating diagnostic models. IGBT accelerated aging experimental data is used to provide the simulation model with a time-varying mapping of aging states and to validate the fault diagnosis model. The adjusted simulation model provides data support for training, and the trained fault diagnosis model's performance is validated on experimental data, forming a structural closed loop.

[0080] In one embodiment of the present invention, a fault diagnosis system for energy storage converters based on semi-supervised domain adaptive learning is provided. This system can be used to implement the above-mentioned fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning. Specifically, the fault diagnosis system for energy storage converters based on semi-supervised domain adaptive learning includes an aging experiment module, a dataset construction module, a data processing module, a model construction module, a model training module, a model evaluation module, and a fault diagnosis module.

[0081] The aging test module is used to build an accelerated aging test platform for IGBTs, apply periodic electrothermal stress to the IGBT devices, and collect the collector-emitter saturation voltage drop of the IGBT devices. Aging data is obtained by using drift data with accumulated stress time, and a quantitative function model describing the changes of key IGBT parameters with equivalent aging time is established based on the aging data. The dataset construction module, connected to the aging experiment module, is used to establish a digital simulation model of the energy storage converter based on the quantitative function model, inject IGBT faults and collect fault waveform data to construct a source domain simulation fault dataset. It is also used to build an energy storage converter experimental prototype, install aged IGBT devices and reproduce IGBT faults, and collect measured fault waveform data to construct a target domain experimental fault dataset. The data processing module, connected to the dataset construction module, is used to preprocess and extract features from the current and voltage signals in the source domain simulation fault dataset and the target domain experimental fault dataset to obtain a feature matrix. It is also used to set the source domain simulation fault dataset as the source domain dataset and to divide the target domain experimental fault dataset into labeled samples and unlabeled samples and set them together as the target domain dataset. The model building module, connected to the data processing module, is used to build a fault diagnosis model consisting of a feature extractor G and a classifier F1. The feature matrix is ​​input into the fault diagnosis model, the feature extractor G maps the input feature vector to a high-dimensional feature space, and the classifier F1 maps the high-dimensional features to fault categories. The model training module, connected to the model building module, is used to merge labeled samples from the source domain and labeled samples from the target domain as training data, obtain prediction results through feature extractor G and classifier F1, calculate classification loss using the cross-entropy loss function and backpropagate to update the parameters of feature extractor G and classifier F1, and also to obtain prediction results for unlabeled samples in the target domain through feature extractor G and classifier F1, calculate the entropy of the prediction distribution as entropy regularization loss, and multiply the gradient of feature extractor G by a negative coefficient through a gradient inversion layer during backpropagation to minimize the entropy regularization loss and update the parameters of feature extractor G. The model evaluation module, which is connected to the model training module and the dataset construction module respectively, is used to periodically evaluate the current model using the target domain test set during the model training process and save the model parameters with the best performance. The fault diagnosis module, connected to the model evaluation module, is used to input the real-time electrical signals of the energy storage converter to be diagnosed into the trained fault diagnosis model and output the fault category and location prediction results.

[0082] This invention provides a terminal device comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or function. The processor described in this embodiment can be used in the operation of a fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning, including: An accelerated aging test platform for IGBTs was built, and periodic electrothermal stress was applied to the IGBT devices to collect the collector-emitter saturation voltage drop. Aging data of IGBT devices is obtained by analyzing drift data over accumulated stress time. Based on this aging data, a quantitative function model describing the changes of key IGBT parameters with equivalent aging time is established. Based on the extracted quantitative function model, a digital simulation model of an energy storage converter is established. IGBT faults are injected into the simulation model, fault waveform data is collected, and a fully labeled source domain simulation fault dataset is constructed. An experimental prototype of the energy storage converter is built, and the aged IGBT devices are installed on the prototype to reproduce IGBT faults. Measured fault waveform data is collected, and a target domain experimental fault dataset is constructed. The current and voltage signals in the source domain simulation fault dataset and the target domain experimental fault dataset are preprocessed and feature extracted to obtain a feature matrix. The fully labeled source domain simulation fault dataset is used as the source domain dataset, and the target domain experimental fault dataset is divided into labeled samples and unlabeled samples, which together form the target domain dataset. A fault diagnosis model consisting of a feature extractor G and a classifier F1 is constructed. Feature extractor G maps the input feature vector to a high-dimensional feature space. Classifier F1, connected after feature extractor G, maps the high-dimensional features to fault categories. Labeled samples from the source and target domains are merged as training data. Prediction results are obtained through feature extractor G and classifier F1. The classification loss is calculated using the cross-entropy loss function, and backpropagation is performed to update the parameters of feature extractor G and classifier F1. For unlabeled samples in the target domain, prediction results are obtained through feature extractor G and classifier F1. The entropy of the prediction distribution is calculated as the entropy regularization loss. During backpropagation, the gradient of feature extractor G is multiplied by a negative coefficient through a gradient reversal layer to minimize the entropy regularization loss and update the parameters of feature extractor G. During training, the current model is periodically evaluated using the target domain test set, and the parameters of the best-performing model are saved. The real-time electrical signal of the energy storage converter to be diagnosed is input into the trained fault diagnosis model, and the fault category and location prediction results are output.

[0083] Please see Figure 5 The terminal device is a computer device. In this embodiment, the computer device 60 includes a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When executed by the processor 61, the computer program 63 implements the energy storage converter fault diagnosis method based on semi-supervised domain adaptive learning in this embodiment. To avoid repetition, details are omitted here. Alternatively, when executed by the processor 61, the computer program 63 implements the functions of each model / unit in the energy storage converter fault diagnosis system based on semi-supervised domain adaptive learning in this embodiment. To avoid repetition, details are omitted here.

[0084] Computer device 60 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. Computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will understand that... Figure 5 This is merely an example of computer device 60 and does not constitute a limitation on computer device 60. It may include more or fewer components than shown, or combine certain components, or different components. For example, computer device may also include input / output devices, network access devices, buses, etc.

[0085] The processor 61 may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0086] The memory 62 can be an internal storage unit of the computer device 60, such as a hard disk or memory of the computer device 60. The memory 62 can also be an external storage device of the computer device 60, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on the computer device 60.

[0087] Furthermore, the memory 62 may include both internal storage units of the computer device 60 and external storage devices. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 can also be used to temporarily store data that has been output or will be output.

[0088] Please see Figure 6 The terminal device is an electronic device 600, which is manifested in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different platform components (including storage unit 620 and processing unit 610), a display unit 640, etc.

[0089] The storage unit stores program code, which can be executed by the processing unit 610 to perform the steps described in the method section of this specification according to various exemplary embodiments of the present invention. For example, the processing unit 610 can perform actions such as... Figure 1 The steps are shown in the figure.

[0090] Storage unit 620 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 6201 and / or cache memory 6202, and may further include a read-only memory (ROM) 6203.

[0091] Storage unit 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0092] Bus 630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the multiple bus structures.

[0093] Electronic device 600 can also communicate with one or more external devices 700 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem). This communication can be performed via input / output interface 650. Furthermore, electronic device 600 can also communicate with one or more networks (e.g., local area network, wide area network, and / or public network, such as the Internet) via network adapter 660. Network adapter 660 can communicate with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0094] Example 4 This invention also provides a storage medium, specifically a computer-readable storage medium, which is a memory device in a terminal device for storing programs and data. It is understood that the computer-readable storage medium here can include both built-in storage media in the terminal device and extended storage media supported by the terminal device; it can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). More specific examples of the computer-readable storage medium include: an electrical connection with one or more wires, a portable disk, a hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical fiber, portable compact disk read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0095] Computer-readable storage media also include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium can also be any readable medium other than a readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, radio frequency, etc., or any suitable combination thereof.

[0096] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0097] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the energy storage converter fault diagnosis method based on semi-supervised domain adaptive learning in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by the processor to perform the following steps: An accelerated aging test platform for IGBTs was built, and periodic electrothermal stress was applied to the IGBT devices to collect the collector-emitter saturation voltage drop. Aging data of IGBT devices is obtained by analyzing drift data over accumulated stress time. Based on this aging data, a quantitative function model describing the changes of key IGBT parameters with equivalent aging time is established. Based on the extracted quantitative function model, a digital simulation model of an energy storage converter is established. IGBT faults are injected into the simulation model, fault waveform data is collected, and a fully labeled source domain simulation fault dataset is constructed. An experimental prototype of the energy storage converter is built, and the aged IGBT devices are installed on the prototype to reproduce IGBT faults. Measured fault waveform data is collected, and a target domain experimental fault dataset is constructed. The current and voltage signals in the source domain simulation fault dataset and the target domain experimental fault dataset are preprocessed and feature extracted to obtain a feature matrix. The fully labeled source domain simulation fault dataset is used as the source domain dataset, and the target domain experimental fault dataset is divided into labeled samples and unlabeled samples, which together form the target domain dataset. A fault diagnosis model consisting of a feature extractor G and a classifier F1 is constructed. Feature extractor G maps the input feature vector to a high-dimensional feature space. Classifier F1, connected after feature extractor G, maps the high-dimensional features to fault categories. Labeled samples from the source and target domains are merged as training data. Prediction results are obtained through feature extractor G and classifier F1. The classification loss is calculated using the cross-entropy loss function, and backpropagation is performed to update the parameters of feature extractor G and classifier F1. For unlabeled samples in the target domain, prediction results are obtained through feature extractor G and classifier F1. The entropy of the prediction distribution is calculated as the entropy regularization loss. During backpropagation, the gradient of feature extractor G is multiplied by a negative coefficient through a gradient reversal layer to minimize the entropy regularization loss and update the parameters of feature extractor G. During training, the current model is periodically evaluated using the target domain test set, and the parameters of the best-performing model are saved. The real-time electrical signal of the energy storage converter to be diagnosed is input into the trained fault diagnosis model, and the fault category and location prediction results are output.

[0098] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0099] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0100] Fault diagnosis application The AC current signal of the energy storage converter to be diagnosed is input into the trained fault diagnosis model, and the fault category and location prediction results are output.

[0101] This invention is illustrated using the open-circuit fault diagnosis of IGBT devices in an energy storage converter as an example. Typical fault modes are summarized, and IGBT open-circuit faults in different bridge arms and locations are taken as diagnostic objects. The open-circuit fault types of IGBT devices in a type I three-level energy storage converter are divided into 78 types, including single-transistor open-circuit faults, in-phase dual-transistor open-circuit faults, and out-of-phase dual-transistor open-circuit faults, as shown in Table 1. Table 1 Classification of Open Circuit Faults

[0102] A fault dataset is constructed by collecting fault waveform data generated by the simulation model and fault waveform data measured on the experimental platform.

[0103] The simulation dataset covers 42 categories.

[0104] Open-circuit fault modes: The experimental dataset contains 78 open-circuit fault modes, which are used as the source and target domains for domain adaptation training. The trained SSDA-MME model is used to diagnose faults on the experimental test set. The test set covers 78 open-circuit fault modes, with a total of 2028 samples.

[0105] Calculations show that the overall diagnostic accuracy of the model reaches 95.96%, the macro-average F1 score is 94.27%, and the average class conditional MMD is reduced to 0.8272, a relative reduction of 40.57%. This significantly reduces the feature distribution difference between similar fault samples in the source and target domains. The domain classifier error is 44.71%, demonstrating good domain confusion effect. This model achieves fault diagnosis and location of energy storage converters with high accuracy.

[0106] The simulation and experimental data above fully demonstrate that the present invention effectively solves the core technical problems in fault diagnosis of energy storage converters, such as the scarcity of experimental samples, the offset between simulation and experimental data domains, the insufficient generalization ability of the model, and the low fault location accuracy. Compared with traditional diagnostic methods, it has achieved significant improvements in diagnostic accuracy, generalization ability, fault coverage, and location accuracy, and has extremely high engineering practical value.

[0107] In summary, this invention presents a fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning. By using simulation and experimental data, it combines the advantages of simulation data in terms of scale and controllability with the value of experimental data in terms of authenticity and reliability. The step of adjusting key parameters in the simulation model using experimental data improves the accuracy of the simulation data in representing the actual physical process, which helps to improve the generalization ability of the subsequently trained diagnostic model. Through an entropy optimization mechanism based on gradient reversal, it utilizes unlabeled data in the target domain to drive the model to learn discriminative features that are invariant across operating conditions, thereby improving the adaptability and generalization performance of the diagnostic model under the target operating conditions.

[0108] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0109] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0110] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0111] In the embodiments provided by this invention, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0112] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0113] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0114] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random-access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0115] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0116] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0118] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning, characterized in that, Includes the following steps: S1. Construct an IGBT accelerated aging test platform, apply periodic electrothermal stress to the IGBT device, and collect the collector-emitter saturation voltage drop of the IGBT device. Aging data of IGBT devices are obtained by using drift data of accumulated stress time; based on the aging data, a quantitative function model describing the changes of key parameters of IGBT with equivalent aging time is established. S2. Based on the quantitative function model extracted in step S1, establish a digital simulation model of the energy storage converter, inject IGBT faults into the simulation model, collect fault waveform data, and construct a fully labeled source domain simulation fault dataset. An experimental prototype of an energy storage converter was built. The IGBT devices aged in step S1 were installed on the experimental prototype to reproduce IGBT faults. Measured fault waveform data were collected to construct an experimental fault dataset in the target domain. S3. Preprocess and extract features from the current and voltage signals in the source domain simulated fault dataset and the target domain experimental fault dataset to obtain a feature matrix; The fully labeled source domain simulation fault dataset is used as the source domain dataset, and the target domain experimental fault dataset is divided into labeled samples and unlabeled samples, which together are used as the target domain dataset. S4. Construct a fault diagnosis model consisting of a feature extractor G and a classifier F1: the feature extractor G is used to map the input feature vector to a high-dimensional feature space; the classifier F1 is connected after the feature extractor G and is used to map the high-dimensional features to fault categories. S5. Merge labeled samples from the source domain and labeled samples from the target domain as training data, obtain prediction results through feature extractor G and classifier F1, calculate classification loss using cross-entropy loss function and perform backpropagation to update the parameters of feature extractor G and classifier F1; for unlabeled samples in the target domain, obtain prediction results through feature extractor G and classifier F1, calculate the entropy of the prediction distribution as entropy regularization loss, and multiply the gradient of feature extractor G by a negative coefficient through gradient inversion layer during backpropagation to minimize the entropy regularization loss and update the parameters of feature extractor G. S6. During training, periodically evaluate the current model using the target domain test set and save the parameters of the model with the best performance. S7. Input the real-time electrical signal of the energy storage converter to be diagnosed into the trained fault diagnosis model, and output the fault category and location prediction results.

2. The fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning according to claim 1, characterized in that, In step S1, applying periodic electrothermal stress to the IGBT device specifically includes: The IGBT device is controlled to periodically switch between the on and off states, so that the junction temperature of the device fluctuates cyclically between a preset high temperature threshold and a low temperature threshold. The on-state voltage drop between the collector and emitter of the IGBT device is monitored and recorded in real time until the increase of the on-state voltage drop relative to the initial value reaches the preset failure criterion. The experiment is then stopped and complete aging trajectory data is obtained. Establishing a quantitative functional model describing the changes of key IGBT parameters with equivalent aging time specifically includes: The collected aging data were fitted and analyzed to determine the nonlinear mapping relationship between the conduction voltage drop drift and the cumulative stress time or the number of switching cycles. Construct an IGBT physical model or behavioral model that includes aging factors, and embed the nonlinear mapping relationship into the model parameters so that the electrical characteristics output by the model can change dynamically with the increase of equivalent aging time.

3. The fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning according to claim 1, characterized in that, In step S2, injecting IGBT faults into the simulation model specifically includes: Simulating IGBT open-circuit fault: By modifying the drive signal of the IGBT device in the simulation model, the gate drive signal is forced to be low or the current path is disconnected. Simulating IGBT short-circuit fault: By directly short-circuiting the collector and emitter of the IGBT device in the simulation model; Simulate device performance degradation faults: Call the quantitative function model established in step S1 to update the on-resistance or on-voltage drop parameters of the IGBT device in the simulation model in real time.

4. The energy storage converter fault diagnosis method based on semi-supervised domain adaptive learning according to claim 3, characterized in that, The energy storage converter prototype adopts a three-level NPC topology. The reproduction of IGBT faults includes reproducing IGBT open-circuit faults by changing the drive signal and reproducing IGBT short-circuit faults by combining the short-circuit protection function of the drive chip. The collected measured fault waveform data includes transient waveform data and steady-state waveform data before and after the fault occurs.

5. The energy storage converter fault diagnosis method based on semi-supervised domain adaptive learning according to claim 1, characterized in that, In step S3, the preprocessing includes DC bias removal processing of the current signal and robust normalization processing based on median and interquartile range. The feature extraction includes extracting signal features in the time domain, frequency domain, wavelet domain, and Hilbert transform domain respectively. After concatenating the extracted features into a feature vector, Z-score normalization is performed to obtain the feature matrix.

6. The fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning according to claim 1, characterized in that, In step S4, the feature extractor G adopts a network structure with three modules connected in series. The first and second modules each contain fully connected transformation, batch normalization, ReLU activation function and Dropout mechanism in sequence. The third module contains fully connected transformation, batch normalization and ReLU activation function in sequence. The classifier adopts a convolutional neural network structure, which includes a preprocessing layer, two one-dimensional convolutional layers, a flattening layer and two fully connected layers in sequence.

7. The fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning according to claim 1, characterized in that, In step S5, multiplying the gradient of the feature extractor by a negative coefficient through the gradient inversion layer specifically includes: During the forward propagation, the gradient inversion layer performs an identity mapping operation, transferring the features output by the feature extractor to the classifier without loss. During backpropagation, the gradient inversion layer multiplies the gradient returned from the classifier by a negative adaptive coefficient and passes the gradient multiplied by the negative coefficient to the feature extractor. This allows the feature extractor to optimize its parameters in a way that maximizes the difficulty for the classifier to distinguish between samples in the source and target domains, thereby achieving the extraction of domain-invariant features.

8. The fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning according to claim 1, characterized in that, The formula for calculating the entropy regularization loss is as follows: in, Weights for entropy loss. The expected entropy of the predicted distribution of unlabeled samples in the target domain.

9. The fault diagnosis method for energy storage converters based on semi-supervised domain adaptive learning according to claim 1, characterized in that, In step S6, the periodic evaluation of the current model using the target domain test set specifically includes: Set an evaluation interval step number, and input labeled samples from the target domain test set into the current model every time the evaluation interval step number is reached; The model's overall accuracy, macro-average F1 score, precision, and recall for each fault category are calculated on the target domain test set. The current performance metrics are compared with the historical best performance metrics. If the current performance is better than the historical best performance, the current model parameters and corresponding performance records are overwritten and saved.

10. A fault diagnosis system for energy storage converters based on semi-supervised domain adaptive learning, characterized in that, include: The aging test module is used to build an accelerated aging test platform for IGBTs, apply periodic electrothermal stress to the IGBT devices, and collect the collector-emitter saturation voltage drop of the IGBT devices. Aging data is obtained by using drift data with accumulated stress time, and a quantitative function model describing the changes of key IGBT parameters with equivalent aging time is established based on the aging data. The dataset construction module, connected to the aging experiment module, is used to establish a digital simulation model of the energy storage converter based on the quantitative function model, inject IGBT faults and collect fault waveform data to construct a source domain simulation fault dataset. It is also used to build an energy storage converter experimental prototype, install aged IGBT devices and reproduce IGBT faults, and collect measured fault waveform data to construct a target domain experimental fault dataset. The data processing module, connected to the dataset construction module, is used to preprocess and extract features from the current and voltage signals in the source domain simulation fault dataset and the target domain experimental fault dataset to obtain a feature matrix. It is also used to set the source domain simulation fault dataset as the source domain dataset and to divide the target domain experimental fault dataset into labeled samples and unlabeled samples and set them together as the target domain dataset. The model building module, connected to the data processing module, is used to build a fault diagnosis model consisting of a feature extractor G and a classifier F1. The feature matrix is ​​input into the fault diagnosis model, the feature extractor G maps the input feature vector to a high-dimensional feature space, and the classifier F1 maps the high-dimensional features to fault categories. The model training module, connected to the model building module, is used to merge labeled samples from the source domain and labeled samples from the target domain as training data, obtain prediction results through feature extractor G and classifier F1, calculate classification loss using the cross-entropy loss function and backpropagate to update the parameters of feature extractor G and classifier F1, and also to obtain prediction results for unlabeled samples in the target domain through feature extractor G and classifier F1, calculate the entropy of the prediction distribution as entropy regularization loss, and multiply the gradient of feature extractor G by a negative coefficient through a gradient inversion layer during backpropagation to minimize the entropy regularization loss and update the parameters of feature extractor G. The model evaluation module, which is connected to the model training module and the dataset construction module respectively, is used to periodically evaluate the current model using the target domain test set during the model training process and save the model parameters with the best performance. The fault diagnosis module, connected to the model evaluation module, is used to input the real-time electrical signals of the energy storage converter to be diagnosed into the trained fault diagnosis model and output the fault category and location prediction results.

Images