A device for power electronic device lifetime evaluation test

Abstract

The application discloses a kind of for power electronic device life evaluation test device, belong to power electronic device life evaluation field.The device includes: data acquisition module is used to obtain the full life cycle of power electronic device under the condition of power cycle test and double pulse test Multiple source degradation data;Data processing module is used to carry out data cleaning, feature extraction and multi-dimensional feature splicing to original degradation data;Life analysis and evaluation module is used to fuse the loss constraint of physical information consistency to the life prediction network based on CNN-BiLSTM, and combines CCLF incremental learning method, realizes the phase evaluation of device life state and life trend prediction.The application improves the physical interpretability of prediction result by introducing monotonicity and boundary constraint, enhances the cross-condition generalization ability by using domain adversarial feature alignment, effectively suppresses catastrophic forgetting by using CCLF incremental learning, significantly improves the accuracy, stability and continuous learning ability of life evaluation.

Description

Technical Field

[0001] This invention relates to the field of power electronic device life assessment, and more particularly to a test apparatus for power electronic device life assessment. Background Technology

[0002] Power electronic devices are core components of power electronics technology, primarily used in main circuits to process and convert or control electrical energy. These devices are typically made of silicon and can handle power ranging from milliwatts to megawatts. Their applications are extensive, including but not limited to industrial machinery, electric vehicles, motor control systems, CNC machine tools, and power systems. For example, in the new energy sector, they are used in the development and utilization of renewable energy sources such as solar and wind power, enabling precise control and regulation of power quality, voltage, and frequency. In power systems, they are applied in high-voltage direct current transmission, flexible alternating current transmission, and reactive power compensation. Power electronic devices typically operate in high-voltage, high-current, and high-frequency environments, which can lead to overheating, degradation, and failure. Furthermore, failures in power electronic devices can severely impact not only individual devices but also the entire electronic system. Therefore, it is necessary to design power electronic device lifespan assessment test equipment to improve device performance and lifespan, reduce maintenance costs, and ensure the stable operation of electronic systems.

[0003] To ensure the reliable and stable operation of power electronic devices under adverse conditions, researchers typically employ online monitoring and data analysis techniques. This allows them to comprehensively analyze and model the degradation process of power electronic devices, thereby predicting their remaining useful life (RUL) and ensuring the safety and reliability of the entire system. Deep learning, as an important branch of machine learning, is also increasingly being applied to RUL prediction for power electronic devices due to its ability to automatically learn features. However, power electronic devices operate under coupled electrical and thermal stress scenarios, making accurate real-time lifetime assessment difficult using traditional data-driven methods. Therefore, by monitoring the device's operating status in real time, combined with historical data and simulation models, the device's health condition can be assessed, and its remaining useful life can be predicted.

[0004] Chinese patent CN111257720A discloses a thyristor life assessment test device that obtains expected lifespan information by collecting thyristor parameter information without manual intervention. However, this scheme only analyzes thyristor monitoring data, neglecting the impact of its actual physical degradation process on lifespan assessment, and also struggles to achieve lifespan status assessment and lifespan trend prediction based on data collected during the test process. Therefore, it is necessary to research a lifespan assessment test device for power electronic devices to achieve lifespan status assessment and lifespan trend prediction based on data collected during the test process. Summary of the Invention

[0005] To at least partially address one of the technical problems existing in the prior art, the present invention aims to provide a test device for lifetime assessment of power electronic devices. This device organically combines physical information consistency loss constraints, domain adversarial feature alignment, and the Correlation Continuous Learning Framework (CCLF) incremental learning method, achieving high-precision, high-generalization-ability lifetime status assessment and trend prediction that supports continuous updates.

[0006] The technical solution adopted in this invention is as follows: A test apparatus for life assessment of power electronic devices, comprising: The data acquisition module is used in conjunction with the power electronic device power cycle degradation test bench and dual-pulse test circuit to collect multi-source degradation data throughout the entire life cycle, characterizing the degradation state of the device. The data processing module is used to perform data cleaning, feature extraction, and multi-dimensional feature concatenation on the multi-source degraded data to obtain multi-dimensional degraded features; and The lifetime analysis and assessment module is used to integrate the physical information consistency loss constraint into the lifetime prediction network based on CNN-BiLSTM, and combine it with the CCLF incremental learning method of the association continuous learning framework to perform phased assessment of the lifetime status and lifetime trend prediction of power electronic devices.

[0007] Furthermore, the data acquisition module includes: a current parameter acquisition device, a voltage parameter acquisition device, a thermal parameter acquisition device, a heat dissipation device, a DC power supply, a Speedgoat control device, and a temperature chamber; The device includes a current parameter acquisition unit for real-time acquisition of conduction current; a voltage parameter acquisition unit for real-time acquisition of voltage across the device and gate voltage; a thermal parameter acquisition unit for real-time acquisition of junction temperature and case temperature; a heat dissipation device for rapid cooling of the device during power cycling degradation testing; a DC power supply that serves as a heating current source during power cycling degradation testing and a voltage source during double-pulse testing; a Speedgoat control unit for automatic switching between power cycling testing and double-pulse testing; and a temperature chamber to ensure that the power electronic device under test operates under set temperature and humidity conditions.

[0008] Furthermore, the data processing module includes a PC host computer; the Speedgoat control device transmits some degradation parameters during the power cycling degradation test to the PC host computer in real time via an Ethernet interface. The PC host computer is used to perform data cleaning, feature extraction, and multi-dimensional feature splicing on the original degradation data, and to construct a training dataset using aging data under different operating conditions, dividing it into a source domain dataset and a target domain dataset.

[0009] Furthermore, the lifespan analysis and assessment module includes: The physical information consistency loss module is used to construct an objective loss function that satisfies physical constraints during the training of the lifetime prediction network. A domain adversarial feature alignment module is used to extract the spatiotemporal features of degradation parameters and achieve feature alignment between different operating conditions; and The incremental learning module employs the CCLF incremental learning method to continuously update the model and suppress catastrophic forgetting as new experimental data arrives.

[0010] Furthermore, the target loss function constructed by the physical information consistency loss module includes a mean squared error term, a monotonicity constraint term, and a boundary constraint term; The monotonicity constraint term is used to penalize non-decreasing trends in the remaining lifetime prediction sequence; The boundary constraint term is used to penalize outputs where the normalized remaining lifetime prediction value exceeds the [0,1] interval.

[0011] Furthermore, the objective loss function expression of the physical information consistency loss module is:

[0012] in, For data fitting terms, For monotonically decreasing constraint terms, For boundary constraint terms, These are the weighting coefficients.

[0013] Furthermore, the domain adversarial feature alignment module includes: The feature extractor, using a CNN-BiLSTM structure, is used to extract the spatiotemporal features of the degradation parameters and output a fixed-dimensional intermediate representation. The label predictor, consisting of a fully connected layer and a ReLU activation function, is used to predict remaining lifetime values ​​based on the intermediate output; and The domain discriminator consists of a three-layer fully connected network and uses the Sigmoid function as the output to determine whether the input features belong to the source domain or the target domain.

[0014] Furthermore, the overall optimization objective of the domain adversarial feature alignment module is to simultaneously minimize the label prediction loss and maximize the domain discrimination loss in order to achieve cross-condition feature alignment.

[0015] Furthermore, the incremental learning module employs the CCLF incremental learning method, which includes: a long-term potentiation plasticity adjustment mechanism and an asciative replay mechanism; The long-term enhancement plasticity adjustment mechanism calculates the average sensitivity of model parameters to prediction errors as the parameter contribution and adaptively adjusts the learning rate of each parameter in the incremental update stage based on the parameter contribution. The associated playback mechanism includes a sample playback and a knowledge buffer unit, which is used to select representative samples from the current stage data and update the knowledge buffer according to the three criteria of information content, diversity and novelty, so as to realize the playback of historical knowledge and the integration of new knowledge.

[0016] Furthermore, the contribution of the parameters Defined as:

[0017] in This is a dataset of degraded samples from historical periods. It refers to the predicted value output by the deep learning model. It is the input to the model. These are the parameters of the model. It is to predict the true value. The loss function is the first... The gradient of each parameter.

[0018] Furthermore, the information content criterion is measured based on the degree of difference between the gradient of a single sample and the overall average gradient; the diversity criterion is measured based on the gradient differences between samples; and the novelty criterion is measured based on the difference between the current sample gradient and the gradients of existing samples in the knowledge buffer.

[0019] Furthermore, the multi-source degradation data includes: static degradation parameters obtained through power cycling tests and dynamic degradation parameters obtained through double-pulse tests; The static degradation parameters include on-state voltage, gate voltage, on-state current, junction temperature, and case temperature; The dynamic degradation parameters include turn-on delay, turn-off delay, rise time, fall time, turn-on time, turn-off time, and switching energy loss.

[0020] Furthermore, the workflow of the lifetime analysis and assessment module includes an offline training phase and an experimental monitoring phase; During the offline training phase, the CNN-BiLSTM-based lifetime prediction network is trained using historical degradation data, and the physical information consistency loss constraint and domain adversarial feature alignment are introduced. During the experimental monitoring phase, when new experimental data arrives, the CCLF incremental learning method continuously updates the model and predicts subsequent degradation trends.

[0021] Furthermore, the CNN-BiLSTM lifetime prediction network includes one-dimensional convolutional layers, batch normalization layers, activation layers, bidirectional LSTM layers, and fully connected layers, which are used to extract spatiotemporal features from the time series of voltage, current, and temperature and output the remaining lifetime prediction value.

[0022] Compared with the prior art, the beneficial effects of the present invention are at least as follows: 1) High degree of automation: The data acquisition module realizes the automatic acquisition, transmission and storage of multi-source degradation data under power cycle test and dual-pulse test conditions, providing a reliable data foundation for subsequent life assessment.

[0023] 2) Comprehensive feature representation: The data processing module extracts and fuses features of multi-source degradation parameters in the time domain, frequency domain, and switching characteristics, which can more accurately characterize the degradation process of power electronic devices.

[0024] 3) Strong physical interpretability: By using the physical information consistency loss module, the monotonic degradation law and normalization boundary constraints are integrated into the deep network training, so that the prediction results can simultaneously meet the data fitting accuracy and physical laws, thus improving the credibility of the model.

[0025] 4) Excellent cross-condition generalization ability: The domain adversarial feature alignment module realizes the alignment of degenerate features under different conditions, which significantly improves the prediction accuracy of the model under unseen conditions.

[0026] 5) Supports continuous learning: The CCLF incremental learning method is adopted, which effectively suppresses catastrophic forgetting in multi-stage incremental updates through plasticity adjustment and associated replay mechanism, thus ensuring the stability of long-term evaluation. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following description is provided with accompanying drawings of the relevant technical solutions in the embodiments of the present invention or the prior art. It should be understood that the accompanying drawings described below are only for the purpose of clearly illustrating some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the overall structure of a power electronic device life assessment test device provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the data acquisition module provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the data processing module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the life analysis and evaluation module provided in an embodiment of the present invention; Figure 5 This is a comparison chart of domain adversarial and monotonic constraint ablation experiments provided in the embodiments of the present invention, showing the prediction curves and error performance of different module combinations (DA+Mon, Mon only, DA only, baseline) in lifetime prediction tasks. Detailed Implementation

[0029] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0030] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0031] In the description of this invention, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0032] In the description of this invention, unless otherwise explicitly defined, terms such as "setting," "installing," and "connecting" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0033] Please see Figure 1 The present invention provides a test apparatus for life evaluation of power electronic devices, which specifically includes the following modules: Component 1: Data Acquisition Module. Required measuring devices include a high-precision oscilloscope, voltage probe, current probe, fiber optic temperature measurement system, and thermocouple sensor, used to acquire voltage, current, and temperature parameters, respectively. The required circuitry includes a power cycling circuit and a dual-pulse test circuit. The power cycling circuit accelerates the degradation of power electronic devices and constructs a lifetime assessment model based on changes in electrical parameters during the degradation process. Its sampling rate can be set to 10 Hz. Its working principle involves applying periodic temperature cycling to the power electronic device: during the heating phase, a large current is applied to the device to raise the junction temperature; after reaching the set heating time, the large current is disconnected, and the device is rapidly cooled during the cooling phase using a heat dissipation component. This cycle repeats, causing the device junction temperature to fluctuate at a fixed period. Due to the inconsistent thermal expansion coefficients of the materials, alternating stress is generated in the internal connection structure, leading to degradation. The dual-pulse test circuit evaluates the switching performance of power electronic devices and monitors their dynamic characteristics as they change during the aging process. Relevant test parameters can be set according to the rated operating conditions of the device under test, including the first pulse width, second pulse width, pulse interval, sampling interval, and data sampling length. Static degradation parameters such as on-state voltage, gate voltage, on-state current, junction temperature, and case temperature can be acquired through power cycling tests, while dynamic degradation parameters such as on-state voltage, gate voltage, and on-state current waveforms can be obtained through dual-pulse tests.

[0034] Component Two: Data Processing Module. This module cleans, smooths, and normalizes the electrical and temperature signals acquired from power cycling tests, extracting time-domain and frequency-domain features. Time-domain features include mean, standard deviation, root square amplitude, root mean square (RMS) amplitude, peak-to-peak value, and skewness. It performs Fast Fourier Transform (FFT) on the raw voltage, current, and temperature waveforms to extract frequency-domain features, including centroid frequency, standard deviation frequency, RMS frequency, frequency kurtosis, frequency skewness, and the mean, variance, skewness, and kurtosis of the amplitude. Finally, it analyzes the double-pulse waveforms acquired from double-pulse tests to obtain the switching characteristic parameters of power electronic devices, including conduction delay. Shutdown delay Ascent time descent time On-time Shutdown time The calculation of the conduction delay is included. With rise time It is necessary to pay attention to the electrical parameter waveforms near the rising edge of the second pulse, including the gate voltage. Rise to 10% of peak value to turn-on voltage The time interval during which the peak value decreases to 90% is the conduction delay. On-state voltage The time interval between the peak value dropping from 90% to 10% is the rise time. ; Calculate the shutdown delay With descent time It is important to pay attention to the electrical parameter waveforms near the falling edge of the first pulse, including the gate voltage. The voltage drops to 90% of its peak value at the turn-on voltage. The time interval during which the peak value rises to 10% is the shutdown delay. On-state voltage The time interval from 10% of the peak to 90% is the decline time. And conduction time The calculation formula is Shutdown time The calculation formula is Turn-on voltage based on double-pulse test , conduction current Waveform calculation of energy loss during power device switching, including conduction loss. The calculation formula is: Among them, shutdown losses The calculation formula is: .in The gate voltage corresponds to the moment near the rising edge of the second pulse. The moment when it drops to 90% of its peak value; Near the falling edge of the first pulse, the gate voltage The moment when it rises to 10% of the peak value.

[0035] Component 3: Lifetime Analysis and Assessment Module. This includes a Physical Information Consistency Loss Module, a Domain Adversarial Feature Alignment Module, and an Incremental Learning Module. The Physical Information Consistency Loss Module is used to construct a target loss function that satisfies physical constraints during the training phase of the lifetime prediction network. The Domain Adversarial Feature Alignment Module uses a CNN-BiLSTM structure to extract spatiotemporal features of degradation parameters. The label predictor consists of a fully connected network and a ReLU activation function, and the domain discriminator consists of a three-layer fully connected network and a Sigmoid output, used to achieve alignment and generalization of degradation features under different operating conditions. The Incremental Learning Module adopts the CCLF (Correlated Continual Learning Framework) incremental learning method, which uses long-term potentialization plasticity adjustment and associative replay mechanism to achieve continuous learning and updating under the condition of newly added experimental data.

[0036] In one embodiment, the lifetime analysis and evaluation module can be divided into an offline training phase and an experimental monitoring phase. During the offline training phase, the multi-source degradation features output from the data processing module are input into the domain adversarial feature alignment module, and the physical information consistency loss module is invoked to construct a combined loss function. This allows the network to minimize prediction errors while satisfying the monotonically decreasing degradation law and lifetime boundary constraints. The physical consistency loss function uses weights... The importance of the three constraints is adjusted so that the lifetime prediction network can simultaneously consider data fitting ability and physical consistency during training. The loss function constituting the physical information consistency loss module consists of a data fitting term, a monotonically decreasing constraint term, and a boundary constraint term. The specific expression of the loss function is as follows:

[0037] in, , , These are the weighting coefficients. This represents the true RUL value for power devices. The RUL prediction value for power devices. For the sample size, Let be the sample sequence number. The optimization objective is: The aim is to minimize the loss function This makes the life prediction results closer to the actual value, thereby improving the accuracy of life assessment.

[0038] Among them, data fitting term The expression is:

[0039] It is used to constrain the model output to be as close as possible to the true lifetime label.

[0040] Among them, the monotonically decreasing constraint term The expression is:

[0041] The lifetime prediction sequence should decrease with the degradation process to satisfy the condition. When predicted life expectancy increases Punishment will be imposed at that time.

[0042] Boundary constraint terms The expression is:

[0043] Therefore, the normalized RUL prediction value should satisfy... When the predicted value Penalties will be imposed if the situation falls outside this range.

[0044] The domain adversarial feature alignment module uses a CNN-BiLSTM network as the basic structure for feature extraction. An adversarial network is connected in series after its output to achieve feature alignment and domain generalization between different working conditions (source domain and target domain). This module mainly includes: (1) Feature extractor This paper employs a CNN-BiLSTM fusion structure, combining a one-dimensional convolutional neural network and a bidirectional LSTM, to extract spatiotemporal features from degenerate sequences such as voltage, current, and temperature. Specifically, it includes four convolutional-normalization-activation sequences and a bidirectional LSTM layer, finally compressing the high-dimensional temporal features into a fixed-dimensional intermediate representation. (2) Tag predictor In the feature extractor output vector Based on this, and through a fully connected layer and a nonlinear activation function, the remaining lifetime prediction for the corresponding time step or sample is output. (3) Domain discriminator The domain adversarial sub-network employs a three-layer fully connected structure with a sigmoid output to determine whether the current feature belongs to the source domain or the target domain. The formula for calculating the domain probability is as follows: ,in It is the ReLU activation function. For the Sigmoid function, The parameters of the domain discriminator, the domain discriminator Binary cross-entropy loss is used as the domain discriminant loss function, and the source and target domain samples are uniformly expressed as follows: , Let be the predicted probability of the domain, then the domain discrimination loss is:

[0045] The operation flow of the domain adversarial feature alignment module is as follows: During training, for datasets under different working conditions, the source domain dataset is labeled as... The target domain dataset is labeled as During the forward propagation process, samples from the source domain and the target domain are input into the feature extractor respectively. The source domain features are input into the label predictor to obtain the lifetime prediction results. Simultaneously, features from both the source and target domains are input into the domain discriminator. , .right use ,right use ,Right now To optimize the objective, among which To counteract the weighting.

[0046] The physical information consistency loss module is responsible for constructing a lifetime prediction loss that includes mean squared error, monotonicity constraints, and boundary constraints. The domain adversarial feature alignment module, through the aforementioned structure and mechanism, constructs a domain discrimination loss. Taking into account lifetime prediction accuracy, physical consistency, and cross-domain generalization performance, the overall objective function of the lifetime analysis and evaluation module in this embodiment can be expressed as:

[0047] in To adjust the weighting coefficient of the domain adversarial loss in the overall objective, parameters are continuously updated via backpropagation. , , Subsequently, the network not only outputs lifetime prediction results that satisfy monotonically decreasing and boundary constraints, but also possesses good domain generalization ability across operating conditions and degradation conditions. The parameter update methods are as follows: , , ,in For learning rate, To counteract the weighting.

[0048] The incremental learning module is used within the lifetime analysis and evaluation module to achieve data coupling and dynamic updates between the physical information consistency loss module and the domain adversarial feature alignment module. After network training, the dynamic learning rate adjustment unit assesses the importance of each parameter in the domain adversarial feature alignment module and adaptively adjusts the learning rate to ensure long-term preservation of key degradation features. The sample replay and knowledge buffer unit selects representative samples from the current stage data based on the model gradient characteristics and updates the knowledge buffer, enabling historical knowledge replay and new knowledge fusion.

[0049] Specifically, the learning rate dynamic adjustment unit and the sample playback and knowledge buffer unit can be divided into two working stages.

[0050] For the first working stage of the learning rate dynamic adjustment unit, after the domain adversarial feature alignment module has been trained, the initial dynamic learning rate is calculated based on the trained model. This is used for updating parameters during backpropagation of the prediction model in the next stage, based on the dynamic learning rate. Perform personalized parameter updates and dynamic learning rate. The expression is:

[0051] in , The parameter contribution determines the direction and magnitude of the learning rate adjustment for each parameter in the next stage of the deep learning model. This parameter reflects the average sensitivity to prediction errors; a larger value indicates that the parameter is more important to the learned degradation patterns. It is defined as the first in the initial stage The contribution of each parameter is expressed as follows:

[0052] in This is a dataset of degraded samples from historical periods. It refers to the predicted value output by the deep learning model. It is the input to the model. These are the parameters of the model. It is to predict the true value. The loss function is the first... The gradient of each parameter.

[0053] For the first working stage of the sample playback and knowledge buffer unit, the knowledge buffer is constructed based on historical degraded data. The sample buffer is based on the amount of information in the sample. diversity The process involves screening, and samples that meet the requirements are selected for the sample playback and knowledge buffer unit. among.

[0054] Sample playback and knowledge buffer unit in the first working stage The specific expression is:

[0055] in This indicates that the top scorers are selected from the overall score. Each sample operator is used as a replay sample for the next step. It is the ratio of the number of selected samples to the size of the dataset. ; It is the total number of the initial samples.

[0056] The first-stage evaluation indicator sample information content The expression is:

[0057] In the formula, , These represent the gradient vector of a single sample and the average gradient vector of the entire dataset, respectively, and their specific calculation formulas are as follows:

[0058]

[0059] Evaluation index sample diversity The expression is:

[0060] In the formula, For historical degradation datasets, , These represent the gradient vector of a single sample and the average gradient vector of the entire dataset, respectively.

[0061] For the second working phase of the learning rate dynamic adjustment unit, based on the first... The contribution of parameters calculated after training convergence. To calculate the first Dynamic learning rate of the step Its expression is:

[0062]

[0063] in It is the first The dynamic learning rate calculated step by step. For the newly labeled experimental phase dataset, It refers to the predicted value output by the deep learning model. It is the input to the model. These are the parameters of the model. It is to predict the true value. The loss function is the first... The gradient of each parameter.

[0064] For the The prediction model parameters are updated step by step, with the formula for updating the backpropagation parameters changed to:

[0065] in .

[0066] For the second working stage of the sample playback and knowledge buffer unit, the knowledge buffer is updated based on the input data from the previous stage. The sample buffer is based on the amount of information in the sample. diversity Novelty Three weighted indicators are used for sample selection and replacement: representativeness of the sample to the overall gradient, inter-sample variability, and difference between the new sample and existing buffer samples. Samples that meet the requirements are selected for sample playback and knowledge buffer units. The sample selection process implemented through this operator can ensure that the selected samples have sufficient information content. diversity and novelty The comprehensive optimization of these three aspects improves the coverage of multi-stage degradation features by the knowledge buffer and the robustness of the model's continuous learning. Their expressions are as follows:

[0067] In the formula, This represents a sample selection operator based on a comprehensive score, used to select samples from the dataset at the current stage. Selected from There are 1 optimal samples, among which This is the ratio of the number of selected samples to the size of the dataset. In the first Total number of samples at each update step These are weighting coefficients; Representation phase The sample set, Indicates in The sample buffer during step updates. Its information content... diversity and novelty The expressions are as follows:

[0068]

[0069]

[0070] In the formula, Indicates the first Step in the first The gradient vector of each sample. Indicates the whole The average gradient vector, Represents a knowledge buffer The average gradient vector in the vector.

[0071] During the test monitoring phase, the data acquisition module monitors and records the electrical and thermal data of the power electronic devices under test. Upon receiving the degradation parameters, the data processing module sequentially performs data cleaning, smoothing, normalization, and feature extraction, extracting switching characteristic parameters such as turn-on delay, turn-off delay, rise time, fall time, turn-on time, and turn-off time from the dynamic switching characteristic data. The processed degradation features are then input into the domain adversarial feature alignment module for model training during the offline phase. When new test phase data arrives, the incremental learning module continuously updates the model using the CCLF incremental learning method, thereby improving the accuracy and stability of lifetime status assessment and lifetime trend prediction in subsequent test phases.

[0072] The invention will be further described below with reference to the accompanying drawings and experimental examples. Please refer to... Figure 2 The data acquisition module of the power electronic device life assessment test apparatus of this invention includes an oscilloscope, a current probe, a voltage probe, and a temperature probe, used to acquire information on the operating current, conduction voltage, and package temperature of the power electronic device. During the power cycling test phase, the sampling rate can be set to 10 Hz; during the dual-pulse test phase, the current probe and voltage probe can be used to acquire transient waveforms at high speed. The power electronic device under test is placed in a temperature chamber to ensure it operates under constant temperature and humidity conditions; during the power cycling test, a heat dissipation device can be used to rapidly cool the device under test; the Speedgoat control device is used to automatically switch between power cycling tests and dual-pulse tests.

[0073] Please see Figure 3The data processing module of the power electronic device life assessment test device designed in this invention includes a host computer (PC) for data processing. Unprocessed multi-source degradation parameters are transmitted to the PC, where data cleaning, feature extraction, and multi-dimensional feature stitching are performed. Time-domain feature parameters include mean, standard deviation, root square amplitude, root mean square frequency, peak-to-peak value, and skewness. A Fast Fourier Transform (FFT) is performed on the original voltage, current, and temperature waveforms to extract frequency-domain feature parameters, including centroid frequency, standard deviation frequency, root mean square frequency, frequency kurtosis, frequency skewness, and the mean, variance, skewness index, and kurtosis index of the amplitude. Switching characteristic parameters of the power electronic device are obtained by analyzing the double-pulse waveform acquired through double-pulse testing, including: turn-on delay, turn-off delay, rise time, fall time, turn-on time, and turn-off time. Calculating the turn-on delay and rise time requires attention to the electrical parameter waveform near the rising edge of the second pulse. The turn-on delay is the time interval from when the gate voltage rises to 10% of its peak value to when the turn-on voltage drops to 90% of its peak value, and the rise time is the time interval from when the turn-on voltage drops from 90% to 10% of its peak value. Calculating the turn-off delay and fall time requires attention to the electrical parameter waveform near the falling edge of the first pulse. The turn-off delay is the time interval from when the gate voltage drops to 90% of its peak value to when the turn-on voltage rises to 10% of its peak value, and the fall time is the time interval from when the turn-on voltage rises from 10% to 90% of its peak value. The turn-on time... The calculation formula is Shutdown time The calculation formula is Turn-on voltage based on double-pulse test , conduction current Waveform calculation of energy loss during power device switching, including conduction loss. The calculation formula is: Among them, shutdown losses The calculation formula is: .in The gate voltage corresponds to the moment near the rising edge of the second pulse. The moment when it drops to 90% of its peak value; Near the falling edge of the first pulse, the gate voltage The moment when it rises to 10% of the peak value.

[0074] Please see Figure 4The lifespan analysis and evaluation module in the power electronic device lifespan assessment test device designed in this invention includes a physical information consistency loss module, a domain adversarial feature alignment module, and an incremental learning module. Degraded features after data processing are first input into the domain adversarial feature alignment module for lifespan prediction. Simultaneously, the physical information consistency loss module applies physical consistency constraints to the prediction process, and the incremental learning module continuously updates the model using the CCLF incremental learning method. This improves prediction accuracy while enhancing the model's physical interpretability and long-term stability.

[0075] To verify the effectiveness of the physical information consistency loss module and the domain adversarial feature alignment module proposed in this invention, an ablation experiment was designed to quantitatively analyze the contribution of each module. Four control schemes were set up in the experiment: (1) baseline, i.e., the basic lifetime prediction model without using domain adversarial feature alignment and monotonicity constraints; (2) DA only, using only the domain adversarial feature alignment module; (3) Mon only, using only the monotonicity constraint term in the physical information consistency loss module; (4) DA+Mon, using both the domain adversarial feature alignment module and the monotonicity constraint term in the physical information consistency loss module.

[0076] See Figure 5 The ablation experiments above show that the monotonicity constraint in the physical information consistency loss module can effectively regulate the degradation trend of the prediction output, while the domain adversarial feature alignment module significantly improves the model's generalization ability across different operating conditions. Table 1 presents the quantitative evaluation indicators of the four ablation schemes. The DA+Mon scheme achieves the best results in RMSE (0.0961), MAE (0.0759), and R² (0.8244), significantly outperforming the baseline scheme (RMSE of 0.1438 and R² of only 0.6073). The DA only scheme (RMSE of 0.1116) outperforms the Mon only scheme (RMSE of 0.1310), indicating that domain adversarial feature alignment has a more significant effect on improving prediction accuracy across operating conditions. The combined use of both achieves the best performance in terms of prediction accuracy and physical consistency, verifying the rationality and necessity of the design of each sub-module in the lifetime analysis and evaluation module.

[0077] Table 1 Comparison of Quantitative Evaluation Indicators in Ablation Experiments

[0078] To verify the effectiveness of the incremental learning module proposed in this invention, a multi-stage incremental learning experiment was designed to compare three different continuous learning strategies: (1) the CCLF strategy, namely the Correlated Continual Learning Framework incremental learning method proposed in this invention; (2) the Replay strategy, namely the comparison method using experience replay; and (3) the Naive strategy, namely the comparison method of directly using new stage data to fine-tune the model without using an anti-forgetting mechanism. The experiment divided the full-lifetime degradation data of power electronic devices into four stages (Stage 1 to Stage 4) in chronological order, and the model was incrementally updated in each stage after the initial training was completed.

[0079] Table 2 presents the quantitative comparison results of the three incremental learning strategies across four comprehensive performance metrics. The experimental data shows that the CCLF strategy achieved the best results in Best RMSE (0.1389), Final RMSE (0.1389), and Best MAE (0.1136), followed by the Replay strategy (Best RMSE: 0.1549, Best MAE: 0.1342). Both significantly outperformed the Naive strategy (Best RMSE: 0.2879, Final RMSE: 0.4123, Best MAE: 0.2195). In the Final AF metric, the CCLF strategy achieved the lowest forgetting rate of 0.0142, the Replay strategy 0.0542, while the Naive strategy reached a high of 0.4040, indicating a severe catastrophic forgetting problem in the multi-stage incremental learning process. The results from the combined indicators show that the incremental learning module using the CCLF incremental learning method performs best in terms of prediction accuracy and knowledge retention, thus verifying its effectiveness in the life assessment test device.

[0080] Table 2 Comparison of Comprehensive Performance Indicators of Incremental Learning Strategies

[0081] In summary, this invention addresses the problems of complex degradation parameter acquisition, insufficient automation in degradation testing processes, and susceptibility to catastrophic amnesia in power electronic device lifetime assessment. It proposes a test device for power electronic device lifetime assessment. This device can automatically acquire degradation data throughout the entire lifecycle of power electronic devices and, by combining a physical information consistency loss module, a domain adversarial feature alignment module, and a CCLF-based incremental learning module, assess and predict the lifetime status of power electronic devices based on data collected during the testing process.

[0082] In the foregoing description of this specification, references to terms such as "one embodiment," "another embodiment," or "some embodiments" indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0083] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0084] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A test apparatus for life evaluation of power electronic devices, characterized in that, include: The data acquisition module is used in conjunction with the power electronic device power cycle degradation test bench and dual-pulse test circuit to collect multi-source degradation data throughout the entire life cycle, characterizing the degradation state of the device. The data processing module is used to perform data cleaning, feature extraction, and multi-dimensional feature splicing on the multi-source degraded data to obtain multi-dimensional degraded features; as well as The lifetime analysis and assessment module is used to integrate the physical information consistency loss constraint into the lifetime prediction network based on CNN-BiLSTM, and combine it with the CCLF incremental learning method of the association continuous learning framework to perform phased assessment of the lifetime status and lifetime trend prediction of power electronic devices.

2. The life evaluation test apparatus for power electronic devices according to claim 1, characterized in that, The data acquisition module includes: a current parameter acquisition device, a voltage parameter acquisition device, a thermal parameter acquisition device, a heat dissipation device, a DC power supply, a Speedgoat control device, and a temperature chamber; The device includes a current parameter acquisition unit for real-time acquisition of conduction current; a voltage parameter acquisition unit for real-time acquisition of voltage across the device and gate voltage; a thermal parameter acquisition unit for real-time acquisition of junction temperature and case temperature; a heat dissipation device for rapid cooling of the device during power cycling degradation testing; a DC power supply that serves as a heating current source during power cycling degradation testing and a voltage source during double-pulse testing; a Speedgoat control unit for automatic switching between power cycling testing and double-pulse testing; and a temperature chamber to ensure that the power electronic device under test operates under set temperature and humidity conditions.

3. The life evaluation test apparatus for power electronic devices according to claim 2, characterized in that, The data processing module includes a PC host computer; the Speedgoat control device transmits some degradation parameters during the power cycling degradation test to the PC host computer in real time via an Ethernet interface. The PC host computer is used to perform data cleaning, feature extraction and multi-dimensional feature splicing on the original degradation data, and to construct a training dataset using aging data under different working conditions, dividing it into a source domain dataset and a target domain dataset.

4. The life evaluation test apparatus for power electronic devices according to claim 1, characterized in that, The lifetime analysis and assessment module includes: The physical information consistency loss module is used to construct an objective loss function that satisfies physical constraints during the training of the lifetime prediction network. A domain adversarial feature alignment module is used to extract the spatiotemporal features of degradation parameters and achieve feature alignment between different operating conditions; and The incremental learning module employs the CCLF incremental learning method to continuously update the model and suppress catastrophic forgetting as new experimental data arrives.

5. The test apparatus for life assessment of power electronic devices according to claim 4, characterized in that, The target loss function constructed by the physical information consistency loss module includes a mean square error term, a monotonicity constraint term, and a boundary constraint term. The monotonicity constraint term is used to penalize non-decreasing trends in the remaining lifetime prediction sequence; The boundary constraint term is used to penalize outputs where the normalized remaining lifetime prediction value exceeds the [0,1] interval.

6. The life evaluation test apparatus for power electronic devices according to claim 4, characterized in that, The domain adversarial feature alignment module includes: The feature extractor, using a CNN-BiLSTM structure, is used to extract the spatiotemporal features of the degradation parameters and output a fixed-dimensional intermediate representation. The label predictor, consisting of a fully connected layer and a ReLU activation function, is used to predict remaining lifetime values ​​based on the intermediate output; and The domain discriminator consists of a three-layer fully connected network and uses the Sigmoid function as the output to determine whether the input features belong to the source domain or the target domain.

7. The life evaluation test apparatus for power electronic devices according to claim 4, characterized in that, The incremental learning module employs the CCLF incremental learning method, which includes: a long-term enhancement plasticity adjustment mechanism and an associated playback mechanism. The long-term enhancement plasticity adjustment mechanism calculates the average sensitivity of model parameters to prediction errors as the parameter contribution and adaptively adjusts the learning rate of each parameter in the incremental update stage based on the parameter contribution. The associated playback mechanism includes a sample playback and a knowledge buffer unit, which is used to select representative samples from the current stage data and update the knowledge buffer according to the three criteria of information content, diversity and novelty, so as to realize the playback of historical knowledge and the integration of new knowledge.

8. The life evaluation test apparatus for power electronic devices according to claim 7, characterized in that, The information content criterion is measured based on the degree of difference between the gradient of a single sample and the overall average gradient; the diversity criterion is measured based on the gradient differences between samples; and the novelty criterion is measured based on the difference between the current sample gradient and the gradients of existing samples in the knowledge buffer.

9. The life evaluation test apparatus for power electronic devices according to claim 1, characterized in that, The multi-source degradation data includes: static degradation parameters obtained through power cycling tests and dynamic degradation parameters obtained through double-pulse tests; The static degradation parameters include on-state voltage, gate voltage, on-state current, junction temperature, and case temperature; The dynamic degradation parameters include turn-on delay, turn-off delay, rise time, fall time, turn-on time, turn-off time, and switching energy loss.

10. The test apparatus for life assessment of power electronic devices according to claim 1, characterized in that, The workflow of the life analysis and assessment module includes an offline training phase and an experimental monitoring phase; During the offline training phase, the CNN-BiLSTM-based lifetime prediction network is trained using historical degradation data, and the physical information consistency loss constraint and domain adversarial feature alignment are introduced. During the experimental monitoring phase, when new experimental phase data arrives, the CCLF incremental learning method continuously updates the model and predicts subsequent degradation trends.

Images