Power module aging prediction method and system based on artificial intelligence
By combining distributed sensor networks and neural network models with the Arrhenius equation and Weiber distribution, the high cost and adaptability issues of power equipment aging prediction are solved, enabling real-time, adaptive aging prediction and economic decision-making, and reducing the total life cycle cost.
Patent Information
- Application Number
- CN202511110196.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-08
- Publication Date
- 2025-11-18
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing methods for predicting the aging of power equipment rely on fixed sensor networks, which are costly and difficult to capture high-frequency transient signals. Traditional methods also show significant limitations in dealing with sudden load shocks, small sample training, and cross-model adaptation.
An AI-based power module aging prediction method is adopted. Multi-dimensional aging feature data are collected in real time through a distributed sensor network, a neural network model is constructed, and the model is optimized using transfer learning and PSO algorithm by combining the Arrhenius equation and the Weiber distribution to achieve real-time and adaptive aging prediction.
It enables real-time and adaptive aging prediction in resource-constrained environments, reduces total lifecycle costs, supports economic decision-making, accurately identifies hidden process defects, and guides maintenance.
Smart Images
Figure CN120974110A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power module aging detection, and in particular to a power module aging prediction method and system based on artificial intelligence. Background Technology
[0002] In the field of power equipment aging prediction, existing solutions for data acquisition largely rely on fixed sensor networks, which are not only costly to deploy but also difficult to capture high-frequency transient signals. Furthermore, current methods in power equipment aging prediction include: physical model-driven methods rely on theoretical models such as the Arrhenius equation, but are limited by the assumption of a constant environment and struggle to cope with dynamic temperature changes in actual operating conditions; traditional machine learning methods such as SVM classifiers require massive amounts of failure samples, and feature engineering heavily depends on expert experience; end-to-end deep learning methods, while capable of automatically extracting features, suffer from high computational resource consumption and poor cross-device generalization ability. These methods all exhibit significant limitations when dealing with sudden load shocks, small-sample training, and cross-model adaptation scenarios. Summary of the Invention
[0003] This invention aims to address at least one of the technical problems existing in the prior art. To this end, this invention proposes an artificial intelligence-based method and system for predicting the aging of power modules, which can quantify the risk of random failures and support economic decision-making.
[0004] One embodiment of the present invention provides an artificial intelligence-based method for predicting the aging of power modules, comprising the following steps: S100, real-time acquisition of multi-dimensional aging feature data of power modules based on a distributed sensor network; the aging feature data includes electrical parameters, thermal parameters, and mechanical parameters; and time-frequency domain feature extraction is performed using a sliding window mechanism to obtain the spatiotemporal correlation matrix of aging features; S200, construction of a neural network model, the neural network model comprising: an aging equation constraint layer, an energy conservation layer, and a failure distribution layer; the aging equation constraint layer discretizes the Arrhenius equation as the gating condition of the LSTM unit and couples the electro-thermal-mechanical loss coefficients; the energy conservation layer calculates the residual feedback between the power loss and the measured temperature based on thermodynamic laws; the failure distribution layer uses a three-parameter Weiber distribution including shape parameters, scale parameters, and position parameters to fit the output layer. Remaining lifetime probability density; S300, using the neural network model as the teacher model, adopting the feature space aligned transfer learning method, freezing the relevant parameter layers of the Arrhenius accelerated aging equation, adjusting the last two convolutional modules of the CNN feature extraction layer, to obtain a transfer learning model adapted to the power module; S400, running the transfer learning model on the platform through edge computing nodes, outputting the three-parameter Weiber distribution curve and confidence interval in real time; optimizing the feature extraction network structure of the edge device through the PSO algorithm, including optimizing the particle dimension, the particle dimension including the number of convolutional kernels, the number of LSTM units and the quantization bit width; the fitness function of the PSO algorithm is the first weight coefficient × feature similarity + the second weight coefficient × inference speed + the third weight coefficient × memory usage; when a sudden change in the environment is detected that causes a decrease in prediction confidence, the PSO online parameter readjustment mechanism is automatically triggered.
[0005] According to some embodiments of the present invention, the electrical parameters include at least one of the following: power device switching loss waveform, current harmonic components, and bus voltage fluctuation; the thermal parameters include at least one of the following: heat sink temperature field distribution and thermal resistance change rate of key solder joints; the mechanical parameters include at least one of the following: vibration spectrum characteristics and structural component deformation displacement.
[0006] According to some embodiments of the present invention, step S200 includes: constructing the aging equation constraint layer, including: S210, discretizing the Arrhenius equation into a form containing a temperature acceleration factor and activation energy parameters; S220, extracting electrical loss coefficient, thermal loss coefficient and mechanical loss coefficient; S230, generating constraint conditions through a gating function based on the electrical loss coefficient, thermal loss coefficient and mechanical loss coefficient and the temperature acceleration factor.
[0007] According to some embodiments of the present invention, step S220 includes: obtaining the electrical loss coefficient by fitting the dV / dt and di / dt test data of the switching transient process; calculating the thermal loss coefficient based on the eigenvalues of the thermal resistance matrix of infrared thermal imaging; and determining the mechanical loss coefficient by the correlation coefficient between the vibration spectrum and the FEM simulation results.
[0008] According to some embodiments of the present invention, step S300 includes: S310, initializing the model architecture, loading the pre-trained teacher model and retaining the Arrhenius equation parameter layer, while keeping the first N-2 convolutional modules of the CNN feature extraction layer unchanged; S320, performing feature space alignment operation, calculating the feature distribution difference between the source domain and the target domain, adding a trainable adaptive layer after the last two convolutional modules, including a channel alignment module with a 1x1 convolutional kernel and a feature distribution calibrator based on MMD loss; S330, implementing a parameter optimization strategy, freezing the LSTM gating parameters related to the Arrhenius equation, adjusting the kernel weights of the last two convolutional modules and adopting a learning rate decay strategy, configuring a composite loss function including prediction loss, distribution difference loss and weight decay; S340, validating the model using a test set.
[0009] According to some embodiments of the present invention, the method further includes: step S500, establishing a PSO parameter-process defect mapping library, including: extracting the key parameter offset ΔP of the power device, establishing a process fluctuation baseline in units of standard deviation; associating the key parameter offset ΔP to the physical defect library through the mapping matrix Φ, wherein the mapping matrix Φ contains feature vectors of several types of process defects; using the PSO particle swarm optimization algorithm to solve the defect weight vector Γ, constructing a defect probability matrix Ψ=Φ×Γ, and defining the particle fitness function as 1-‖ΔP-ΦΓ‖2 / ‖ΔP‖2.
[0010] According to some embodiments of the present invention, the method further includes: step S600, generating a production batch-aging mode knowledge graph, including: establishing edge relationships based on the cosine similarity of aging coefficient vectors with production batches as nodes; the aging coefficient vectors include electrical loss coefficients, thermal loss coefficients and mechanical loss coefficients; identifying the dominant aging mode through particle swarm clustering; embedding maintenance schemes in the graph nodes to form a traceable reliability knowledge network.
[0011] According to some embodiments of the present invention, the method further includes: S710, when the location parameter value of the three-parameter Weber distribution is less than the cumulative running time of the module, triggering a first-level warning and generating an immediate maintenance command; S720, calculating the optimal maintenance time window based on the scale parameter: if the scale parameter value falls within the interval [current time + ΔT, current time + 2ΔT], generating a second-level warning and planning preventive maintenance; where ΔT is a preset buffer period; S730, when the shape parameter is less than 1 and the confidence interval width exceeds the threshold, automatically increasing the data acquisition frequency to 3 times the original sampling rate.
[0012] According to some embodiments of the present invention, the method further includes: S810, parsing the defect type corresponding to the maximum value in the defect probability matrix Ψ, and if the defect probability of the defect type is greater than a preset threshold, calling the physical defect library to match the standard maintenance plan; S820, retrieving historical cases of similar defects through the knowledge graph, and prioritizing the maintenance plan with a matching degree greater than the preset probability as the benchmark template; S830, dynamically adjusting the plan in combination with the resource status of the edge device.
[0013] The method of this invention includes at least the following beneficial effects: This invention enables the prediction system to achieve real-time performance and adaptability in resource-constrained environments through Probability of Self-Supporting (PSO), and also avoids cost overruns due to hardware upgrades. It quantifies random failure risks through Weiber distribution and supports economic decision-making, reducing total lifecycle costs. It identifies latent process defects by setting a defect matrix and guides precise maintenance resources; this invention supports decision-making from two dimensions—system-level lifespan prediction and component-level defect diagnosis—through the output of a three-parameter Weiber distribution and the defect probability matrix, respectively corresponding to time-driven maintenance and condition-driven maintenance.
[0014] Another embodiment of the present invention provides an artificial intelligence-based power module aging prediction system, comprising: an aging feature extraction module, used to collect multi-dimensional aging feature data of the power module in real time based on a distributed sensor network; the aging feature data includes electrical parameters, thermal parameters, and mechanical parameters; and a sliding window mechanism is used to extract time-frequency domain features to obtain an aging feature spatiotemporal correlation matrix; and a neural network construction module, used to construct a neural network model, the neural network model including: an aging equation constraint layer, an energy conservation layer, and a failure distribution layer; the aging equation constraint layer discretizes the Arrhenius equation as the gating condition of the LSTM unit and couples the electro-thermal-mechanical loss coefficients; the energy conservation layer calculates the residual feedback between the loss power and the measured temperature based on thermodynamic laws; and the failure distribution layer uses a three-parameter Weiber distribution including shape parameters, scale parameters, and position parameters to fit the residual of the output layer. Lifetime probability density; Transfer learning model module, used to use the neural network model as the teacher model, adopts a feature space aligned transfer learning method, freezes the relevant parameter layers of the Arrhenius accelerated aging equation, and adjusts the last two convolutional modules of the CNN feature extraction layer to obtain a transfer learning model adapted to the power module; Edge computing node module, used to run the transfer learning model on the platform through edge computing nodes, and outputs the three-parameter Weiber distribution curve and confidence interval in real time; Optimizes the feature extraction network structure of the edge device through the PSO algorithm, including optimizing the particle dimension, which includes the number of convolutional kernels, the number of LSTM units, and the quantization bit width; The fitness function of the PSO algorithm is the first weight coefficient × feature similarity + the second weight coefficient × inference speed + the third weight coefficient × memory usage; When a sudden change in the environment is detected that causes a decrease in prediction confidence, the PSO online parameter readjustment mechanism is automatically triggered.
[0015] The system of this invention embodiment has at least the following beneficial effects: This invention embodiment enables the prediction system to achieve real-time performance and adaptability in resource-constrained environments through PSO, and can also avoid cost runaway due to hardware upgrades. It quantifies random failure risk through Weiber distribution and supports economic decision-making, reducing total lifecycle costs. It identifies latent process defects by setting a defect matrix and guides precise maintenance resources; this invention embodiment supports decision-making from two dimensions—system-level lifespan prediction and component-level defect diagnosis—through the output of a three-parameter Weiber distribution and the defect probability matrix, respectively corresponding to time-driven maintenance and condition-driven maintenance.
[0016] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.
[0017] Additional aspects and advantages 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
[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:
[0019] Figure 1 This is a flowchart illustrating the method according to an embodiment of the present invention;
[0020] Figure 2 This is a schematic block diagram of the system modules according to an embodiment of the present invention.
[0021] Figure label:
[0022] The module includes an aging feature extraction module 100, a neural network construction module 200, a transfer learning model module 300, and an edge computing node module 400. Detailed Implementation
[0023] 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.
[0024] 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.
[0025] Reference Figure 1 This invention proposes an artificial intelligence-based method for predicting the aging of power modules, comprising the following steps:
[0026] S100: Real-time acquisition of multi-dimensional aging feature data of power modules based on distributed sensor network; aging feature data includes electrical parameters, thermal parameters and mechanical parameters; and time-frequency domain feature extraction is performed using a sliding window mechanism to obtain the spatiotemporal correlation matrix of aging features.
[0027] S200. Construct a neural network model, which includes: an aging equation constraint layer, an energy conservation layer, and a failure distribution layer. The aging equation constraint layer discretizes the Arrhenius equation as the gating condition of the LSTM unit and couples the electro-thermal-mechanical loss coefficients. The energy conservation layer calculates the residual feedback between the loss power and the measured temperature based on the thermodynamic laws. The failure distribution layer uses a three-parameter Weiber distribution containing shape parameters, scale parameters, and position parameters to fit the remaining lifetime probability density of the output layer.
[0028] S300. Using the neural network model as the teacher model, a feature space aligned transfer learning method is adopted. The relevant parameter layers of the Arrhenius accelerated aging equation are frozen, and the last two convolutional modules of the CNN feature extraction layer are adjusted to obtain a transfer learning model adapted to the power module.
[0029] S400 runs a transfer learning model on the platform through edge computing nodes, outputting the three-parameter Weiber distribution curve and confidence interval in real time; it optimizes the feature extraction network structure of edge devices through the PSO algorithm, including optimizing the particle dimension, which includes the number of convolutional kernels, the number of LSTM units, and the quantization bit width; the fitness function of the PSO algorithm is the first weight coefficient × feature similarity + the second weight coefficient × inference speed + the third weight coefficient × memory usage; when a sudden change in the environment is detected that causes a decrease in prediction confidence, the PSO online parameter readjustment mechanism is automatically triggered.
[0030] The solution of this invention first employs teacher model transfer technology, combined with the MMD loss function, to align the feature space of laboratory data with field data, thereby reducing the training data requirements of the target domain; secondly, it preserves the reliability of the physical model by freezing the Arrhenius parameter layer, while designing an adjustable convolutional module to adaptively learn the features of sudden changes in working conditions; finally, it adopts a parameter selective optimization strategy.
[0031] In some embodiments, the present invention implements online optimization of the feature extraction network for edge devices through a dynamic particle swarm optimization algorithm. Its core implementation process includes three key stages: First, initializing the particle swarm parameters, the number of convolutional kernels (8-64), the number of LSTM units (16-128), and the quantization bit width (4-8 bits) together constitute a three-dimensional search space, with each particle position corresponding to a set of network structure configurations; second, establishing a multi-objective fitness function, using a weighted combination of feature similarity (using MMD distance metric), inference speed (millisecond latency), and memory usage (in MB), where the weight coefficients are dynamically adjusted according to the device type (e.g., for power monitoring equipment, which prioritizes inference speed, the weight is set to 0.6); finally, deploying an online readjustment mechanism. When the environmental sensor detects a temperature change exceeding ±15℃ or a load fluctuation greater than 30%, the system automatically triggers the PSO optimization process, retaining the current optimal particle as the initial population center, and performing 5-10 generations of rapid optimization within the compressed search space (within ±20% of the original space). The PSO embodiment of this invention is an enabling technology that allows prediction systems to achieve real-time performance and adaptability in resource-constrained environments, avoiding cost overruns due to hardware upgrades.
[0032] This invention quantifies random failure risk using the Weiber distribution: when the shape parameter β < 1 (early failure period), monitoring should be strengthened even if there are no high-risk defects in the defect probability matrix; it supports economic decision-making: the optimal maintenance window is calculated through the scale parameter to reduce the total life cycle cost.
[0033] In some embodiments, electrical parameters include at least one of the following: power device switching loss waveform, current harmonic components, and bus voltage fluctuation; thermal parameters include at least one of the following: heat sink temperature field distribution and thermal resistance change rate of key solder joints; mechanical parameters include at least one of the following: vibration spectrum characteristics and structural component deformation displacement.
[0034] In some embodiments, step S200 includes: constructing an aging equation constraint layer, including: S210, discretizing the Arrhenius equation into a form containing a temperature acceleration factor and activation energy parameters; S220, extracting electrical loss coefficient, thermal loss coefficient and mechanical loss coefficient; S230, generating constraint conditions through a gating function based on the electrical loss coefficient, thermal loss coefficient, mechanical loss coefficient and temperature acceleration factor.
[0035] In some embodiments, step S220 includes: obtaining the electrical loss coefficient by fitting the dV / dt and di / dt test data of the switching transient process; calculating the thermal loss coefficient based on the eigenvalues of the thermal resistance matrix of infrared thermal imaging; and determining the mechanical loss coefficient by the correlation coefficient between the vibration spectrum and the FEM simulation results.
[0036] In some embodiments, step S300 includes: S310, initializing the model architecture, loading the pre-trained teacher model and retaining the Arrhenius equation parameter layer, while keeping the first N-2 convolutional modules of the CNN feature extraction layer unchanged; S320, performing feature space alignment operation, calculating the feature distribution difference between the source domain and the target domain, adding a trainable adaptive layer after the last two convolutional modules, including a channel alignment module with a 1x1 convolutional kernel and a feature distribution calibrator based on MMD loss; S330, implementing a parameter optimization strategy, freezing the LSTM gating parameters related to the Arrhenius equation, adjusting the kernel weights of the last two convolutional modules and adopting a learning rate decay strategy, configuring a composite loss function including prediction loss, distribution difference loss and weight decay; S340, validating the model using a test set.
[0037] In some embodiments, the method of the present invention further includes: step S500, establishing a PSO parameter-process defect mapping library, including: extracting the key parameter offset ΔP of the power device, establishing a process fluctuation baseline in units of standard deviation; associating the key parameter offset ΔP to the physical defect library through the mapping matrix Φ, wherein the mapping matrix Φ contains feature vectors of several types of process defects; using the PSO particle swarm optimization algorithm to solve the defect weight vector Γ, constructing a defect probability matrix Ψ=Φ×Γ, and defining the particle fitness function as 1-‖ΔP-ΦΓ‖2 / ‖ΔP‖2.
[0038] In some embodiments, the method of the present invention further includes: step S600, generating a production batch-aging mode knowledge graph, including: establishing edge relationships based on the cosine similarity of aging coefficient vectors with production batches as nodes; the aging coefficient vectors include electrical loss coefficients, thermal loss coefficients and mechanical loss coefficients; identifying the dominant aging mode through particle swarm clustering; and embedding maintenance schemes into the graph nodes to form a traceable reliability knowledge network.
[0039] In some embodiments, the method of the present invention further includes: S710, when the location parameter value of the three-parameter Weber distribution is less than the cumulative running time of the module, triggering a first-level warning and generating an immediate maintenance instruction; S720, calculating the optimal maintenance time window based on the scale parameter: if the scale parameter value falls within the interval [current time + ΔT, current time + 2ΔT], generating a second-level warning and planning preventive maintenance; where ΔT is a preset buffer period; S730, when the shape parameter is less than 1 and the confidence interval width exceeds the threshold, automatically increasing the data acquisition frequency to 3 times the original sampling rate.
[0040] In some embodiments, the method of the present invention further includes: S810, parsing the defect type corresponding to the maximum value in the defect probability matrix Ψ, and if the defect probability of the defect type is greater than a preset threshold, calling the physical defect library to match the standard maintenance plan; S820, retrieving historical cases of similar defects through the knowledge graph, and prioritizing the maintenance plan with a matching degree greater than the preset probability as the benchmark template; S830, dynamically adjusting the plan in combination with the resource status of the edge device.
[0041] In the two embodiments of the present invention above, decision-making can be supported from two dimensions: system-level lifetime prediction and component-level defect diagnosis, respectively, corresponding to time-driven maintenance and state-driven maintenance.
[0042] Corresponding to the foregoing embodiments, the present invention also provides system embodiments. For system embodiments, since they are essentially corresponding to method embodiments, relevant details can be found in the description of the method embodiments.
[0043] Reference Figure 2 This invention proposes an artificial intelligence-based power module aging prediction system, comprising:
[0044] The aging feature extraction module 100 is used to collect multi-dimensional aging feature data of the power module in real time based on a distributed sensor network. The aging feature data includes electrical parameters, thermal parameters and mechanical parameters. The time-frequency domain feature extraction is performed using a sliding window mechanism to obtain the spatiotemporal correlation matrix of aging features.
[0045] The neural network construction module 200 is used to construct a neural network model, which includes an aging equation constraint layer, an energy conservation layer, and a failure distribution layer. The aging equation constraint layer discretizes the Arrhenius equation as the gating condition of the LSTM unit and couples the electro-thermal-mechanical loss coefficients. The energy conservation layer calculates the residual feedback between the loss power and the measured temperature based on the thermodynamic laws. The failure distribution layer uses a three-parameter Weiber distribution containing shape parameters, scale parameters, and position parameters to fit the remaining lifetime probability density of the output layer.
[0046] The transfer learning model module 300 is used to use the neural network model as the teacher model, adopt the feature space aligned transfer learning method, freeze the relevant parameter layers of the Arrhenius accelerated aging equation, and adjust the last two convolutional modules of the CNN feature extraction layer to obtain a transfer learning model adapted to the power module.
[0047] The edge computing node module 400 is used to run transfer learning models on the platform via edge computing nodes, and output three-parameter Weiber distribution curves and confidence intervals in real time. It optimizes the feature extraction network structure of edge devices through the PSO algorithm, including optimizing the particle dimension, which includes the number of convolutional kernels, the number of LSTM units, and the quantization bit width. The fitness function of the PSO algorithm is the first weight coefficient × feature similarity + the second weight coefficient × inference speed + the third weight coefficient × memory usage. When a sudden change in the environment is detected that causes a decrease in prediction confidence, the PSO online parameter readjustment mechanism is automatically triggered.
[0048] Although specific embodiments are described herein, those skilled in the art will recognize that many other modifications or alternative embodiments are also within the scope of this disclosure. For example, any of the functions and / or processing capabilities described in connection with a particular device or component can be performed by any other device or component. Furthermore, while various exemplary embodiments and architectures have been described according to embodiments of this disclosure, those skilled in the art will recognize that many other modifications to the exemplary embodiments and architectures described herein are also within the scope of this disclosure.
[0049] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.
[0050] The above is a detailed description of the preferred embodiments of this application. However, this application 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 this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.
Claims
1. An artificial intelligence-based method for predicting the aging of power modules, characterized in that, Includes the following steps: S100. Real-time acquisition of multi-dimensional aging characteristic data of the power module based on a distributed sensor network; the aging characteristic data includes electrical parameters, thermal parameters, and mechanical parameters; A sliding window mechanism was used to extract time-frequency domain features, resulting in the spatiotemporal correlation matrix of aging features. S200. Construct a neural network model, which includes: an aging equation constraint layer, an energy conservation layer, and a failure distribution layer; The aging equation constraint layer discretizes the Arrhenius equation as the gating condition of the LSTM unit and couples the electro-thermal-mechanical loss coefficient. The energy conservation layer calculates the residual feedback between the loss power and the measured temperature based on the thermodynamic law. The failure distribution layer uses a three-parameter Weiber distribution containing shape parameters, scale parameters and position parameters to fit the remaining lifetime probability density of the output layer. S300. Using the neural network model as the teacher model, a feature space aligned transfer learning method is adopted. The relevant parameter layers of the Arrhenius accelerated aging equation are frozen, and the last two convolutional modules of the CNN feature extraction layer are adjusted to obtain a transfer learning model adapted to the power module. S400: The transfer learning model is run on the platform via edge computing nodes, and the three-parameter Weiber distribution curve and confidence interval are output in real time; the feature extraction network structure of the edge device is optimized through the PSO algorithm, including optimizing the particle dimension, which includes the number of convolutional kernels, the number of LSTM units, and the quantization bit width; the fitness function of the PSO algorithm is the first weight coefficient × feature similarity + the second weight coefficient × inference speed + the third weight coefficient × memory usage; when a sudden change in the environment is detected that causes a decrease in prediction confidence, the PSO online parameter readjustment mechanism is automatically triggered.
2. The artificial intelligence-based power module aging prediction method according to claim 1, characterized in that, The electrical parameters include at least one of the following: power device switching loss waveform, current harmonic components, and bus voltage fluctuation; The thermal parameters include at least one of the following: radiator temperature field distribution and rate of change of thermal resistance of key solder joints; The mechanical parameters include at least one of the following: vibration spectrum characteristics and structural component deformation displacement.
3. The artificial intelligence-based power module aging prediction method according to claim 1, characterized in that, Step S200 includes: constructing the aging equation constraint layer, including: S210. Discretize the Arrhenius equation into a form containing temperature acceleration factor and activation energy parameter; S220, Extract electrical loss coefficient, heat loss coefficient and mechanical loss coefficient; S230. Based on the electrical loss coefficient, thermal loss coefficient, mechanical loss coefficient, and temperature acceleration factor, generate constraint conditions through a gating function.
4. The artificial intelligence-based power module aging prediction method according to claim 3, characterized in that, Step S220 includes: obtaining the electrical loss coefficient by fitting the dV / dt and di / dt test data of the switching transient process; calculating the thermal loss coefficient based on the eigenvalues of the thermal resistance matrix of infrared thermal imaging; and determining the mechanical loss coefficient by the correlation coefficient between the vibration spectrum and the FEM simulation results.
5. The artificial intelligence-based power module aging prediction method according to claim 1, characterized in that, Step S300 includes: S310. Initialize the model architecture, load the pre-trained teacher model and retain the Arrhenius equation parameter layer, while keeping the first N-2 convolutional modules of the CNN feature extraction layer unchanged. S320: Perform feature space alignment operation, calculate the difference in feature distribution between the source domain and the target domain, and add a trainable adaptive layer after the last two convolutional modules, which includes a channel alignment module with a 1x1 convolutional kernel and a feature distribution calibrator based on MMD loss. S330. Implement parameter optimization strategy, freeze the LSTM gating parameters related to the Arrhenius equation, adjust the kernel weights of the last two convolutional modules and adopt the learning rate decay strategy, and configure a composite loss function that includes prediction loss, distribution difference loss and weight decay. S340. Validate the model using the test set.
6. The artificial intelligence-based power module aging prediction method according to claim 1, characterized in that, The method further includes: Step S500: Establish a PSO parameter-process defect mapping library, including: extracting the key parameter offset ΔP of power devices and establishing a process fluctuation baseline in units of standard deviation; associating the key parameter offset ΔP with the physical defect library through the mapping matrix Φ, where the mapping matrix Φ contains feature vectors of several types of process defects; using the PSO particle swarm optimization algorithm to solve for the defect weight vector Γ, constructing a defect probability matrix Ψ=Φ×Γ, and defining the particle fitness function as 1-‖ΔP-ΦΓ‖2 / ‖ΔP‖2.
7. The artificial intelligence-based power module aging prediction method according to claim 6, characterized in that, The method further includes: Step S600: Generate a production batch-aging mode knowledge graph, including: establishing edge relationships based on the cosine similarity of aging coefficient vectors with production batches as nodes; the aging coefficient vectors include electrical loss coefficients, thermal loss coefficients, and mechanical loss coefficients; identifying the dominant aging mode through particle swarm clustering; and embedding maintenance schemes into the graph nodes to form a traceable reliability knowledge network.
8. The artificial intelligence-based power module aging prediction method according to claim 7, characterized in that, The method further includes: S710. When the location parameter value of the three-parameter Weber distribution is less than the cumulative running time of the module, a first-level early warning is triggered and an immediate maintenance command is generated. S720. Calculate the optimal maintenance time window based on the scale parameter: If the scale parameter value falls within the interval [current time + ΔT, current time + 2ΔT], then generate a level 2 early warning and plan preventive maintenance; where ΔT is the preset buffer period. S730: When the shape parameter is less than 1 and the confidence interval width exceeds the threshold, automatically increase the data acquisition frequency to 3 times the original sampling rate.
9. The artificial intelligence-based power module aging prediction method according to claim 8, characterized in that, The method further includes: S810. Analyze the defect type corresponding to the maximum value in the defect probability matrix Ψ. If the defect probability of the defect type is greater than the preset threshold, call the physical defect library to match the standard maintenance plan. S820. Search for historical cases of similar defects through knowledge graphs, and prioritize the maintenance solutions with a matching degree greater than the preset probability as the benchmark template. S830, combined with a dynamic adjustment scheme based on the resource status of edge devices.
10. An artificial intelligence-based power module aging prediction system, used to perform the method as described in claims 1 to 9, characterized in that, include: An aging feature extraction module is used to collect multi-dimensional aging feature data of power modules in real time based on a distributed sensor network. The aging characteristic data includes electrical parameters, thermal parameters, and mechanical parameters; A sliding window mechanism was used to extract time-frequency domain features, resulting in the spatiotemporal correlation matrix of aging features. A neural network construction module is used to construct a neural network model, which includes an aging equation constraint layer, an energy conservation layer, and a failure distribution layer. The aging equation constraint layer discretizes the Arrhenius equation as the gating condition of the LSTM unit and couples it with the electro-thermal-mechanical loss coefficient. The energy conservation layer calculates the residual feedback between the loss power and the measured temperature based on the thermodynamic laws. The failure distribution layer uses a three-parameter Weiber distribution containing shape parameters, scale parameters, and position parameters to fit the remaining lifetime probability density of the output layer. The transfer learning model module is used to take the neural network model as the teacher model, adopt the feature space aligned transfer learning method, freeze the relevant parameter layer of the Arrhenius accelerated aging equation, adjust the last two convolutional modules of the CNN feature extraction layer, and obtain the transfer learning model adapted to the power module. The edge computing node module is used to run the transfer learning model on the platform via edge computing nodes, and output the three-parameter Weiber distribution curve and confidence interval in real time; it optimizes the feature extraction network structure of the edge device through the PSO algorithm, including optimizing the particle dimension, which includes the number of convolutional kernels, the number of LSTM units, and the quantization bit width; the fitness function of the PSO algorithm is the first weight coefficient × feature similarity + the second weight coefficient × inference speed + the third weight coefficient × memory usage; when a sudden change in the environment is detected that causes a decrease in prediction confidence, the PSO online parameter readjustment mechanism is automatically triggered.