An IGBT residual life prediction method, system, medium and device
By combining variational mode decomposition and asymmetric two-stream GRU network with cross-attention mechanism, the problems of aging characteristics being submerged by noise and insufficient physical constraints in IGBT remaining lifetime prediction are solved, achieving accurate and reliable lifetime prediction, which is suitable for industrial field applications.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG UNIV
- Filing Date
- 2026-05-26
- Publication Date
- 2026-08-04
AI Technical Summary
Existing IGBT remaining lifetime prediction technologies suffer from several problems, including the aliasing of physical characteristics across scales, which makes aging characteristics easily submerged by operating noise; the lack of physical causal mapping capability in network topology; and overly soft physical constraints and high deployment costs.
Variational mode decomposition algorithm is used for frequency domain decoupling, and slow-varying aging feature stream and fast-varying operating condition feature stream are constructed. An asymmetric dual-stream GRU network is built for feature extraction and causal verification. Weighted fusion is performed by combining cross-attention mechanism, and physical constraints are applied by parameterless hard coding layer to output predictive results that conform to the rules.
It accurately characterizes the material fatigue evolution trajectory throughout the entire life cycle of a device, solves the problem of high-frequency noise overshadowing aging characteristics, achieves reliable prediction under harsh operating conditions, and eliminates phenomena that violate physical laws, making it suitable for online deployment in industrial settings.
Smart Images
Figure CN122263029B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of IGBT remaining lifetime prediction technology, and particularly relates to an IGBT remaining lifetime prediction method, system, medium and device. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Insulated-gate bipolar transistors (IGBTs), as core components of power converters, are widely used in key areas such as new energy power generation, electric vehicles, and smart grids. Statistics show that IGBT failures account for approximately 21% of all converter failures, with thermal cycling stress being the leading cause, accounting for as much as 55%. Therefore, accurately predicting the remaining lifespan of IGBTs has become a research hotspot in the field of power electronics reliability.
[0004] Current IGBT remaining life prediction technologies mainly fall into two categories. The first category is physical failure models based on degradation mechanisms, such as the Coffin-Manson model and the Darveaux model. These methods predict life by establishing a mathematical relationship between the number of thermal cycles and cumulative damage, and their prediction results have strong physical interpretability. However, these models heavily rely on precisely extracted physical parameters (such as activation energy and material constants) under laboratory conditions, making it difficult to adapt to the complex and variable dynamic load conditions in industrial settings. When actual operating conditions deviate from the modeling conditions, the model's prediction accuracy drops sharply.
[0005] The second category is data-driven deep learning methods, including time-series network models such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and gated recurrent units (GRUs). These methods do not require the establishment of complex physical equations; instead, they predict remaining lifespan by directly mining the evolutionary patterns of historical monitoring data.
[0006] However, existing data-driven methods still have significant shortcomings when dealing with complex industrial field conditions. Firstly, most methods use a single feature (such as...). First, simply concatenating multiple features and inputting them into a single network ignores the fundamental differences in the aging process across time scales, causing weak aging features to be overwhelmed by severe operating noise, leading to prediction lag and misjudgments. Second, the self-attention mechanism in existing methods typically treats operating fluctuations as disturbances and assigns them low weights, failing to model the clear physical causal law of "adverse operating conditions accelerating aging." Third, Physical Information Neural Networks (PINNs) employ a "soft constraint" approach by incorporating physical rules into the loss function, which may still produce unreasonable prediction results that violate physical laws, such as "lifetime rebound," under extreme conditions. Furthermore, the computational cost of complex mechanism models is high, making it difficult to achieve lightweight online deployment on low-computing-power embedded chips. Summary of the Invention
[0007] To overcome the shortcomings of the prior art, the present invention provides a method, system, medium and device for predicting the remaining lifetime of IGBTs, aiming to solve the problems in the prior art where aging characteristics are easily submerged by operating noise due to the aliasing of physical characteristics across scales, the lack of physical causal mapping capability in network topology, and the excessively soft physical constraints and high deployment costs.
[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for predicting the remaining lifetime of an IGBT; A method for predicting the remaining lifetime of an IGBT, comprising: Collect dual-characteristic timing data of saturation on-state voltage drop and collector current during IGBT operation, and perform preprocessing; A variational mode decomposition algorithm is introduced to decouple the preprocessed saturated on-state voltage drop data in the frequency domain, and to construct the slow-varying aging characteristic flow and the fast-varying operating condition characteristic flow. An asymmetric dual-stream GRU network is constructed. The slow-varying GRU branch is used to extract hidden state features that characterize the macroscopic lifetime degradation trend. The fast-varying GRU branch is used to perform physical causality verification on the feature flow of fast-varying operating conditions, and to purify the real transient thermal stress features. A cross-attention mechanism is constructed to weight and fuse the hidden state features and transient thermal stress features that characterize the macroscopic lifetime degradation trend to obtain global aging features; The global aging features are input into the fully connected layer to output the original remaining lifetime prediction value. Then, the monotonically decreasing correction and boundary locking are performed through the parameterless hard-coded layer to output the final remaining lifetime prediction result that conforms to physical laws.
[0009] As a further technical solution, the preprocessing includes: The collected dual-feature time-series data of saturated on-state voltage drop and collector current are first synchronously downsampled to extract the average value within each power cycle; The downsampled dual features were standardized with zero mean and unit variance respectively to eliminate dimensional differences.
[0010] As a further technical solution, a variational mode decomposition algorithm is introduced to decouple the preprocessed saturated on-state voltage drop data in the frequency domain, constructing a slow-varying aging characteristic flow and a fast-varying operating condition characteristic flow, including: Using the preprocessed saturated on-state voltage drop data as input to the variational mode decomposition algorithm, the original signal is adaptively decomposed into K eigenmode components with independent center frequencies by solving the following constrained variational optimization problem:
[0011] in, This represents the k-th modal component obtained from the decomposition; This is the center frequency corresponding to this component. For the Dirac function, Represents convolution operation; Let be the partial derivative with respect to time t; The imaginary unit; Frequency domain analysis is performed on the K intrinsic mode components obtained by decomposition. The mode component with the lowest center frequency is extracted as the low-frequency degradation baseline and constructed as a slow-varying aging characteristic flow. The remaining high-frequency modal components are superimposed and reconstructed to form high-frequency fluctuation noise. The reconstructed high-frequency fluctuation noise is then matrix-joined with the preprocessed collector current data to construct a fast-changing operating condition characteristic flow.
[0012] As a further technical solution, slow-varying GRU branches are used to extract hidden state features characterizing the macroscopic lifetime degradation trend, including: The constructed slow-varying aging feature stream is used as the input sequence of the slow-varying GRU branch, and the long time series information is iteratively modeled through the update gate and reset gate of the gated recurrent unit. At each time step, the slow-varying GRU branch receives the current input and the hidden state of the previous time step. After calculation by the gating mechanism, it outputs the hidden state of the current time step. The hidden state sequence generated after iterating step by step is used as the hidden state feature characterizing the macroscopic lifetime degradation trend.
[0013] As a further technical solution, the rapid-change GRU branch is used to perform physical causality verification on the characteristic flow of rapidly changing operating conditions, and to refine the true transient thermal stress characteristics, including: The constructed fast-changing operating condition feature flow is used as the input sequence of the fast-changing GRU branch, and the conditional probability distribution of high-frequency voltage fluctuation and collector current are jointly learned through the internal gating mechanism. When high-frequency voltage fluctuations are accompanied by synchronous abrupt changes in collector current, they are identified as real transient thermal shocks with a clear physical stress source, and the hidden state transmission corresponding to this feature is enhanced by updating the gate; when high-frequency voltage fluctuations exhibit isolated jumps without the support of abrupt changes in collector current, they are identified as random measurement noise or electromagnetic interference, and the noise is adaptively attenuated and the hidden state transmission is shielded by the forgetting mechanism. The hidden state sequence output after the fast-variable GRU branch iteration is used as the purified true transient thermal stress feature.
[0014] As a further technical solution, a cross-attention mechanism is constructed to weightedly fuse hidden state features and transient thermal stress features characterizing macroscopic lifetime degradation trends, thereby obtaining global aging features, including: The transient thermal stress features output by the fast-varying branch are mapped to a query vector matrix, and the hidden state features representing the macroscopic lifetime degradation trend output by the slow-varying branch are mapped to a key matrix and a value matrix:
[0015]
[0016]
[0017] in, To query the vector matrix; The key matrix; It is a value matrix; , , These are the weight matrices for the query, key, and value, respectively. Characterized by transient thermal stress; Hidden state features characterizing macroscopic lifespan degradation trends; Dynamic modulation weights are calculated using a cross-attention formula, and weighted fusion of cross-scale features is performed.
[0018] in, This is a global aging characteristic. is the matrix dimension.
[0019] As a further technical solution, the monotonically decreasing correction is as follows:
[0020] in, The predicted value is after monotonically decreasing correction; This is the original remaining life prediction value; This is the final revised value from the previous cycle; there are no historical values for the first cycle. The previous cycle correction value is set as the upper limit. If the current original forecast value is higher, it is replaced with the upper limit value to achieve monotonically decreasing. The boundary locking is as follows:
[0021] in, This is the final predicted remaining lifespan of the IGBT.
[0022] A second aspect of the present invention provides an IGBT remaining lifetime prediction system.
[0023] An IGBT remaining lifetime prediction system, comprising: The data acquisition module is configured to acquire dual-characteristic timing data of saturation on-state voltage drop and collector current during IGBT operation, and perform preprocessing. The frequency domain decoupling module is configured to: introduce a variational mode decomposition algorithm to decouple the preprocessed saturated on-state voltage drop data in the frequency domain and construct a slow-varying aging characteristic flow and a fast-varying operating condition characteristic flow; The asymmetric dual-stream GRU module is configured to: build an asymmetric dual-stream GRU network, extract hidden state features that characterize the macroscopic lifetime degradation trend using the slow-varying GRU branch, perform physical causal verification on the fast-varying operating condition feature stream using the fast-varying GRU branch, and purify the real transient thermal stress features. The cross-attention fusion module is configured to: construct a cross-attention mechanism to perform weighted fusion of hidden state features and transient thermal stress features that characterize the macroscopic lifetime degradation trend to obtain global aging features; The parameterless hard-coded constraint module is configured to: input the global aging features into the fully connected layer to output the original remaining lifetime prediction value, and then perform monotonically decreasing correction and boundary locking through the parameterless hard-coded layer to output the final remaining lifetime prediction result that conforms to physical laws.
[0024] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of an IGBT remaining lifetime prediction method as described in the first aspect of the present invention.
[0025] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the IGBT remaining lifetime prediction method as described in the first aspect of the present invention.
[0026] The above one or more technical solutions have the following beneficial effects: (1) This invention introduces the variational mode decomposition (VMD) algorithm into the field of IGBT remaining lifetime prediction. Utilizing the adaptive frequency domain resolution capability of VMD, the mixed collector-emitter saturation voltage drop signal is forcibly separated into two independent components with clear physical meaning: a low-frequency degradation baseline and high-frequency fluctuation noise. The low-frequency component filters out all operating condition interference and accurately characterizes the material fatigue evolution trajectory of the device throughout its entire life cycle; the high-frequency component completely preserves the transient thermal shock information caused by load changes. This solves the core problem of the weak aging baseline being submerged by severe high-frequency noise.
[0027] (2) This invention constructs an asymmetric dual-current GRU architecture. By synchronously inputting high-frequency voltage fluctuations and collector current characteristics, the Fast-GRU network utilizes an internal gate control mechanism to jointly learn the conditional probability distribution of the transient changes of both. When a high-frequency voltage jump is accompanied by a synchronous current jump, the network determines it as a real thermal shock with a clear physical stress source, thus strengthening feature transmission. Conversely, when an isolated voltage jump lacks current support, the network identifies it as random measurement noise or electromagnetic interference, which is adaptively attenuated and shielded by a forgetting mechanism. This fundamentally solves the technical problem of accurately extracting real thermal stress in high-frequency industrial environments. In addition, a cross-attention modulation module oriented towards physical causal reasoning is also constructed. This module uses the purified transient operating condition status identifier as the query vector and the macroscopic aging state as the key and value. Through the dynamic calculation of attention weights, it realizes the algorithm-level mapping of the semiconductor physical law of accelerated aging under harsh operating conditions.
[0028] (3) The present invention applies a forced physical constraint to the prediction result through a parameterless hard-coded layer, transforming the monotonically decreasing characteristic of the remaining lifetime and the [0,1] numerical boundary into a fixed operation logic, thereby fundamentally eliminating phenomena such as "lifetime rebound" and "numerical over-boundary" that violate physical laws.
[0029] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0030] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0031] Figure 1 This is a flowchart of the method in the first embodiment.
[0032] Figure 2 This is a schematic diagram of the overall method structure of the first embodiment.
[0033] Figure 3 This is a schematic diagram of the overall structure of the GRU network in the first embodiment.
[0034] Figure 4 This is a system structure diagram of the second embodiment. Detailed Implementation
[0035] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0036] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.
[0037] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.
[0038] Example 1 This embodiment discloses a method for predicting the remaining life of IGBTs. It achieves frequency domain separation of aging and operating condition characteristics through VMD, employs a dual-stream GRU for feature extraction and thermal stress causality verification, combines cross-attention mapping with physical causality, and applies rigid constraints through parameterless hard coding. This method can eliminate feature aliasing, provides rigorous and reliable predictions, and is suitable for online life prediction in industrial settings.
[0039] like Figure 1 and Figure 2 As shown, a method for predicting the remaining lifetime of an IGBT includes: Step S1: Collect dual-characteristic timing data of saturation on-state voltage drop and collector current during IGBT operation, and perform preprocessing.
[0040] Collect the saturation on-state voltage drop of the IGBT during operation and collector current The time-series data above comes from IGBT thermal overstress experiments, ensuring coverage of the complete aging process from brand new to near failure.
[0041] It directly reflects the long-term aging process of the device, such as gate oxide degradation, solder layer fatigue, and bond wire detachment, and serves as a benchmark signal for lifetime prediction. It directly reflects the degree of transient fluctuations in load conditions and is the core stress source that causes severe thermal cycling and thermal shock inside the device.
[0042] Subsequently, the obtained saturation on-state voltage drop was analyzed. and collector current The timing data is preprocessed by first extracting the timing data of IGBT saturation on-state voltage drop and collector current for each power cycle. and The average value is used to transform the high-frequency raw signal into a power-cycled low-frequency sequence. Then, the downsampled dual features are standardized using zero-mean and unit-variance methods to eliminate the dimensional differences between voltage and current. Subsequently, the standardized data is synchronously smoothed using an exponential moving average (EMA) to filter out short-term noise and fluctuations while preserving the saturation on-state voltage drop. The core aging trend gradually increases, eventually resulting in a stable two-dimensional time-series feature sequence. Each time step , where t is the power cycle and T is the total number of cycles.
[0043] Step S2: Introduce variational mode decomposition algorithm to decouple the preprocessed saturated on-state voltage drop data in the frequency domain and construct the slow-varying aging characteristic flow and the fast-varying operating condition characteristic flow.
[0044] In actual industrial settings, IGBT aging is an extremely slow-cycle physical degradation process, while load fluctuations are transient, high-frequency shocks. Directly inputting raw monitoring features into a deep neural network results in severe aliasing of physical features, easily causing subtle, genuine aging characteristics to be drowned out by intense high-frequency noise. Therefore, introducing the VMD algorithm not only avoids the phase delay introduced by traditional filters but also transforms the mixed signal containing core aging information into a more robust and efficient signal processing mechanism. Forced separation into independent components with clear physical meaning.
[0045] Step S21: The preprocessed saturated on-state voltage drop data is used as input to the variational mode decomposition algorithm. By solving the following constrained variational optimization problem, the original signal is adaptively decomposed into K eigenmode components with independent center frequencies:
[0046] in, This represents the k-th modal component obtained from the decomposition; This is the center frequency corresponding to this component. For the Dirac function, Represents convolution operation; Let be the partial derivative with respect to time t. It is the imaginary unit.
[0047] Step S22: Perform frequency domain analysis on the K intrinsic mode components obtained from the decomposition, assign a real physical meaning to each decomposed mode component, and perform dual-stream reconstruction.
[0048] Specifically, the lowest frequency mode component decomposed by VMD is defined as the low-frequency degradation baseline. This component filters out all operating condition interferences, accurately characterizing the actual material fatigue and long-term physical aging evolution trajectory within the device, spanning its entire lifespan. It is extracted separately and constructed as a slowly varying aging feature stream:
[0049] in, It is a slow-aging characteristic flow.
[0050] The high-frequency modal components decomposed by VMD are reconstructed into high-frequency wave noise. This component characterizes transient thermal shocks caused by sudden load changes and short-term measurement jumps. It is then compared with the current signal representing the system's operating condition. By combining and stitching together, a characteristic flow of rapidly changing operating conditions is constructed:
[0051] in, For rapidly changing operating conditions, the characteristic flow is...
[0052] Through the aforementioned front-end processing, the purification and isolation of long-term decay and short-term impact are achieved at the data source, providing a pure input stream with excellent physical causality for subsequent deep networks.
[0053] Step S3: Construct an asymmetric dual-stream GRU network. Utilize the slow-varying GRU branch to extract hidden state features characterizing macroscopic lifetime degradation trends, and utilize the fast-varying GRU branch to perform physical causality verification on the rapidly changing operating condition feature flow, thereby refining the true transient thermal stress features.
[0054] GRU is a type of recurrent neural network, and its overall structure is shown in the diagram below. Figure 3 As shown, it introduces a gating mechanism, which achieves efficient modeling of long-time-series information through the synergistic effect of update gate and reset gate, and performs excellently in capturing the long-period, slow-time-varying time-series degradation trend during IGBT aging.
[0055] In this embodiment, two structurally independent but parallel-operating GRU network branches are constructed, namely the Slow-GRU branch and the Fast-GRU branch.
[0056] The slow-varying branch (Slow-GRU) receives the slow-varying aging characteristic stream from the front-end input, which is free from any operating condition interference. Through long-term iterative processing of this branch, pure hidden state features characterizing the macroscopic lifetime degradation trend of the device are extracted. .
[0057] Specifically, the slowly varying aging feature stream constructed in step S2... This serves as the input sequence for a slowly varying GRU branch. A slowly varying GRU branch consists of multiple gated recurrent units connected in series, each containing an update gate. With Reset Door This is used for efficient modeling of long-term time-series information. At each time step t, the slowly varying GRU branch receives the input from the current time step. (i.e., the value of the slow-varying aging characteristic current in the current power cycle) and the hidden state passed from the previous time step. .
[0058] First, reset the door. Determine the degree to which the hidden state from the previous time step was ignored, and generate candidate hidden states. .in, For activation function, , For a trainable weight matrix, This indicates element-wise multiplication. This is the hidden state of the previous time step.
[0059] Subsequently, the door was updated. Control the fusion ratio between the hidden state and the candidate hidden states from the previous time step, and output the hidden state at the current time step. .in, This is the candidate hidden state.
[0060] Due to the slow-varying GRU branch receiving input It is a low-frequency degraded baseline reconstructed after variational mode decomposition, which has filtered out all high-frequency operating condition interferences. Therefore, the hidden state sequence iteratively generates is accurate. This method purely characterizes the material fatigue evolution trajectory of the device throughout its entire lifecycle, i.e., the macroscopic lifetime degradation trend. This hidden state sequence will serve as the source of the bond and value matrices in the subsequent cross-attention mechanism.
[0061] The Fast-GRU receives the fast-changing operating condition feature stream from the front-end input. (i.e., includes high-frequency voltage fluctuations) With collector current (The joint matrix). After causality verification and extraction by Fast-GRU, the purified true transient thermal stress characteristics are output. .
[0062] The fast-variable GRU branch consists of multiple gated recurrent units connected in series, and its network structure is independent of the slow-variable GRU branch. At each time step t, the fast-variable GRU branch receives the input at the current time. The hidden state of the previous time step Through the internal update portal and reset door Implement physical causality verification: The formula for resetting the door is: The candidate hidden state is .
[0063] The updated gate calculation formula is as follows The final hidden state at the current moment is .
[0064] In the joint learning process of the aforementioned gating mechanism, the fast-changing GRU branch automatically mines the conditional probability distribution relationship between high-frequency voltage fluctuations and collector current. Specifically: when High-frequency jumps occurred and were accompanied by synchronous changes. When the current changes abruptly, the two show a strong correlation. The network determines that the jump originates from a real physical thermal shock caused by a sudden load change, and then updates the gate. Assigning a larger value strengthens the transmission of the hidden state corresponding to this feature, enabling transient thermal stress information to be effectively written into the current hidden state. Conversely, when An isolated jump occurs. When the state remains stable, there is no physical causal relationship between the two, and the network identifies it as random measurement noise or electromagnetic interference. At this time, the forgetting mechanism (i.e., the decay effect of the smaller update gate value and the reset gate) adaptively weakens the influence of the noise on the hidden state and blocks the transmission of false jumps.
[0065] After all time steps of iteration, the hidden state sequence output by the fast-change GRU branch. This represents the purified true transient thermal stress characteristics. These characteristics successfully filter out random jumps and retain only thermal shock information with clear physical stress sources, characterizing the severity of the absolute thermal shock. This information will serve as the source of the query vector matrix in the subsequent cross-attention mechanism.
[0066] Step S4: Construct a cross-attention mechanism to perform weighted fusion of hidden state features and transient thermal stress features that characterize the macroscopic lifetime degradation trend, and obtain global aging features.
[0067] The core idea of attention mechanisms is to dynamically weight and aggregate information at different positions in an input sequence using learned weight coefficients, automatically focusing on features more critical to the current task while suppressing redundant or irrelevant information. It breaks the limitations of fixed weights in traditional sequence modeling, flexibly capturing long-distance dependencies within a sequence, and is one of the key technologies for improving model feature extraction efficiency and task performance.
[0068] By introducing a cross-attention mechanism, the purified transient thermal stress features are used as query conditions to dynamically evaluate and weight the macroscopic aging state, thus realizing a physical causal mapping of aging induced and accelerated by sudden changes in operating conditions on the network topology. The specific matrix mapping is as follows: Output the fast branch Mapped to the query vector matrix Q, the output of the slow-varying branch... Mapped to a key matrix K and a value matrix V:
[0069]
[0070]
[0071] in, To query the vector matrix; The key matrix; It is a value matrix; , , These are the weight matrices for the query, key, and value, respectively. Characterized by transient thermal stress; Hidden state features that characterize the macroscopic lifespan degradation trend.
[0072] Dynamic modulation weights are calculated using a cross-attention formula, and weighted fusion of cross-scale features is performed.
[0073] in, This is a global aging characteristic. is the matrix dimension.
[0074] In the aforementioned cross-modulation structure, the query matrix Q precisely captures the severity of the current operating condition fluctuations. The physical laws governing semiconductor reliability indicate that operating condition fluctuations (such as...) The more severe the thermal cycling caused by abrupt temperature changes, the more serious the internal damage to the device. Therefore, in dot product operations... In this process, the more drastic the fluctuations in operating conditions, the higher the weight allocation calculated by the attention mechanism. This is equivalent to applying a fatigue acceleration penalty factor that conforms to physical laws to the V dynamic, which represents the slow aging characteristics, using the characteristics of rapidly changing operating conditions.
[0075] Finally, the globally fused aging features are obtained through cross-modulation. This method integrates macroscopic decay trends with microscopic abrupt shocks. This fused feature is mapped through a fully connected layer, ultimately outputting the raw remaining lifetime prediction without subsequent physical rule constraints. And pass it to the backend hard-coded layer.
[0076] Step S5: Input the global aging features into the fully connected layer to output the original remaining lifetime prediction value, and then perform monotonically decreasing correction and boundary locking through the parameterless hard-coded layer to output the final remaining lifetime prediction result that conforms to physical laws.
[0077] A parameterless hard-coded layer is added after the fully connected layer. This coding layer has no trainable parameters, thus not increasing model complexity or training cost. It only contains fixed mathematical operations such as min and max, making the operations simple and computationally insignificant. During online prediction, only the correction results from the previous cycle need to be cached, incurring no additional storage cost. The specific formula is as follows: The monotonically decreasing trend is corrected to:
[0078] in, The predicted value is after monotonically decreasing correction; This is the original remaining life prediction value; This is the final revised value from the previous period; there are no historical values for the first period. .
[0079] The previous cycle correction value is set as the upper limit. If the current original forecast value is higher, it is replaced with the upper limit value to achieve monotonically decreasing.
[0080] The boundary locking is as follows:
[0081] in, This is the final predicted remaining IGBT lifetime. The corrected value is forcibly truncated to the [0,1] interval to ensure the result conforms to the physical definition. Double physical constraints are applied through fixed mathematical operations to strictly prevent unreasonable predictions. The final predicted remaining lifetime, conforming to monotonically decreasing and [0,1] boundary constraints, is obtained. .
[0082] Example 2 This embodiment discloses an IGBT remaining lifetime prediction system; like Figure 4 As shown, an IGBT remaining lifetime prediction system includes: The data acquisition module is configured to acquire dual-characteristic timing data of saturation on-state voltage drop and collector current during IGBT operation, and perform preprocessing. The frequency domain decoupling module is configured to: introduce a variational mode decomposition algorithm to decouple the preprocessed saturated on-state voltage drop data in the frequency domain and construct a slow-varying aging characteristic flow and a fast-varying operating condition characteristic flow; The asymmetric dual-stream GRU module is configured to: build an asymmetric dual-stream GRU network, extract hidden state features that characterize the macroscopic lifetime degradation trend using the slow-varying GRU branch, perform physical causal verification on the fast-varying operating condition feature stream using the fast-varying GRU branch, and purify the real transient thermal stress features. The cross-attention fusion module is configured to: construct a cross-attention mechanism to perform weighted fusion of hidden state features and transient thermal stress features that characterize the macroscopic lifetime degradation trend to obtain global aging features; The parameterless hard-coded constraint module is configured to: input the global aging features into the fully connected layer to output the original remaining lifetime prediction value, and then perform monotonically decreasing correction and boundary locking through the parameterless hard-coded layer to output the final remaining lifetime prediction result that conforms to physical laws.
[0083] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.
[0084] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in an IGBT remaining lifetime prediction method as described in Example 1.
[0085] Example 4 The purpose of this embodiment is to provide an electronic device.
[0086] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in an IGBT remaining lifetime prediction method as described in Embodiment 1.
[0087] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.
[0088] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.
[0089] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A method for predicting the remaining lifetime of an IGBT, characterized in that, include: Collect dual-characteristic timing data of saturation on-state voltage drop and collector current during IGBT operation, and perform preprocessing; A variational mode decomposition algorithm is introduced to decouple the preprocessed saturated on-state voltage drop data in the frequency domain, and to construct the slow-varying aging characteristic flow and the fast-varying operating condition characteristic flow. An asymmetric dual-stream GRU network is constructed. The slow-varying GRU branch extracts hidden state features characterizing macroscopic lifetime degradation trends, while the fast-varying GRU branch performs physical causality verification on the fast-varying operating condition feature stream, refining the true transient thermal stress features. Specifically: the constructed fast-varying operating condition feature stream is used as the input sequence of the fast-varying GRU branch. An internal gating mechanism jointly learns the conditional probability distributions of high-frequency voltage fluctuations and collector current. When a high-frequency voltage fluctuation is accompanied by a synchronous abrupt change in collector current, it is identified as a true transient thermal shock with a clear physical stress source, and the hidden state transmission corresponding to this feature is strengthened by updating the gate. When a high-frequency voltage fluctuation exhibits an isolated jump without supporting abrupt changes in collector current, it is identified as random measurement noise or electromagnetic interference, and the forgetting mechanism adaptively attenuates and shields the hidden state transmission of this noise. The hidden state sequence output after the fast-change GRU branch iteration is used as the purified true transient thermal stress feature; A cross-attention mechanism is constructed to weight and fuse the hidden state features and transient thermal stress features that characterize the macroscopic lifetime degradation trend to obtain global aging features; The global aging features are input into the fully connected layer to output the original remaining lifetime prediction value. This value is then passed through a parameter-free hard-coded layer for monotonically decreasing correction and boundary locking, outputting the final remaining lifetime prediction result that conforms to physical laws. The monotonically decreasing correction is as follows: in, The predicted value is after monotonically decreasing correction; This is the original remaining life prediction value; This is the final revised value from the previous cycle; there are no historical values for the first cycle. The previous cycle correction value is set as the upper limit. If the current original forecast value is higher, it is replaced with the upper limit value to achieve monotonically decreasing. The boundary locking is as follows: in, This is the final predicted remaining lifespan of the IGBT.
2. The IGBT remaining lifetime prediction method as described in claim 1, characterized in that, The preprocessing includes: The collected dual-feature time-series data of saturated on-state voltage drop and collector current are first synchronously downsampled to extract the average value within each power cycle; The downsampled dual features were standardized with zero mean and unit variance respectively to eliminate dimensional differences.
3. The IGBT remaining lifetime prediction method as described in claim 1, characterized in that, A variational mode decomposition algorithm is introduced to decouple the preprocessed saturated on-state voltage drop data in the frequency domain, constructing slow-varying aging characteristic flow and fast-varying operating condition characteristic flow, including: Using the preprocessed saturated on-state voltage drop data as input to the variational mode decomposition algorithm, the original signal is adaptively decomposed into K eigenmode components with independent center frequencies by solving the following constrained variational optimization problem: in, This represents the k-th modal component obtained from the decomposition; This is the center frequency corresponding to this component. For the Dirac function, Represents convolution operation; Let be the partial derivative with respect to time t. The imaginary unit; Frequency domain analysis is performed on the K intrinsic mode components obtained by decomposition. The mode component with the lowest center frequency is extracted as the low-frequency degradation baseline and constructed as a slow-varying aging characteristic flow. The remaining high-frequency modal components are superimposed and reconstructed to form high-frequency fluctuation noise. The reconstructed high-frequency fluctuation noise is then matrix-joined with the preprocessed collector current data to construct a fast-changing operating condition characteristic flow.
4. The IGBT remaining lifetime prediction method as described in claim 1, characterized in that, The slow-varying GRU branching method is used to extract hidden state features that characterize the macroscopic lifetime degradation trend, including: The constructed slow-varying aging feature stream is used as the input sequence of the slow-varying GRU branch, and the long time series information is iteratively modeled through the update gate and reset gate of the gated recurrent unit. At each time step, the slow-varying GRU branch receives the current input and the hidden state of the previous time step. After calculation by the gating mechanism, it outputs the hidden state of the current time step. The hidden state sequence generated after iterating step by step is used as the hidden state feature characterizing the macroscopic lifetime degradation trend.
5. The method for predicting the remaining lifetime of an IGBT as described in claim 1, characterized in that, A cross-attention mechanism is constructed to weightedly fuse hidden state features and transient thermal stress features characterizing macroscopic lifetime degradation trends, resulting in global aging features, including: The transient thermal stress features output by the fast-varying branch are mapped to a query vector matrix, and the hidden state features representing the macroscopic lifetime degradation trend output by the slow-varying branch are mapped to a key matrix and a value matrix: in, To query the vector matrix; The key matrix; It is a value matrix; , , These are the weight matrices for the query, key, and value, respectively. Characterized by transient thermal stress; Hidden state features characterizing macroscopic lifespan degradation trends; Dynamic modulation weights are calculated using a cross-attention formula, and weighted fusion of cross-scale features is performed. in, This is a global aging characteristic. is the matrix dimension.
6. An IGBT remaining lifetime prediction system, characterized in that, include: The data acquisition module is configured to acquire dual-characteristic timing data of saturation on-state voltage drop and collector current during IGBT operation, and perform preprocessing. The frequency domain decoupling module is configured to: introduce a variational mode decomposition algorithm to decouple the preprocessed saturated on-state voltage drop data in the frequency domain and construct a slow-varying aging characteristic flow and a fast-varying operating condition characteristic flow; The asymmetric dual-stream GRU module is configured to: construct an asymmetric dual-stream GRU network; extract hidden state features characterizing macroscopic lifetime degradation trends using a slow-varying GRU branch; and perform physical causality verification on the fast-varying operating condition feature stream using a fast-varying GRU branch to purify the true transient thermal stress features. Specifically, the constructed fast-varying operating condition feature stream is used as the input sequence of the fast-varying GRU branch, and the conditional probability distributions of high-frequency voltage fluctuations and collector currents are jointly learned through an internal gating mechanism. When a high-frequency voltage fluctuation is accompanied by a synchronous abrupt change in collector current, it is identified as a true transient thermal shock with a clear physical stress source, and the hidden state transmission corresponding to this feature is strengthened by updating the gate. When a high-frequency voltage fluctuation exhibits an isolated jump without the support of a sudden change in collector current, it is identified as random measurement noise or electromagnetic interference, and the hidden state transmission of this noise is adaptively attenuated and shielded by a forgetting mechanism. The hidden state sequence output after iterative processing by the fast-varying GRU branch is used as the purified true transient thermal stress feature. The cross-attention fusion module is configured to: construct a cross-attention mechanism to perform weighted fusion of hidden state features and transient thermal stress features that characterize the macroscopic lifetime degradation trend to obtain global aging features; The parameterless hard-coded constraint module is configured to: input the global aging features into the fully connected layer to output the original remaining lifetime prediction value, and then perform monotonically decreasing correction and boundary locking through the parameterless hard-coded layer to output the final remaining lifetime prediction result that conforms to physical laws; wherein, the monotonically decreasing correction is: in, The predicted value is after monotonically decreasing correction; This is the original remaining life prediction value; This is the final revised value from the previous cycle; there are no historical values for the first cycle. The previous cycle correction value is set as the upper limit. If the current original forecast value is higher, it is replaced with the upper limit value to achieve monotonically decreasing. The boundary locking is as follows: in, This is the final predicted remaining lifespan of the IGBT.
7. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps in the IGBT remaining lifetime prediction method as described in any one of claims 1-5.
8. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the IGBT remaining lifetime prediction method as described in any one of claims 1-5.