A method, device, and dielectric for predicting the remaining lifetime of a power semiconductor module.

Abstract

This invention discloses a method, device, and medium for predicting the remaining lifetime of power semiconductor modules, relating to the field of semiconductor lifetime prediction technology. The method includes: collecting material lattice parameters and thermodynamic operating condition data to generate a transport dataset using molecular dynamics simulation; generating a reciprocal space group velocity tensor field based on the transport dataset; generating a thermodynamic entropy change index based on the reciprocal space group velocity tensor field; performing time-series cumulative analysis based on the thermodynamic entropy change index to output a phonon entropy change feature sequence; extracting a statistical feature subset and a dynamic gradient subset from the phonon entropy change feature sequence; constructing a spiking neural network and inputting the phonon entropy change feature sequence to generate an operation function; and calling the transport equation and initializing it based on the statistical feature subset to generate an initial transport equation. This invention achieves accurate modeling of phonon scattering nonlinear effects by dynamically reconstructing the collision terms of the transport equation using a spiking neural network, overcoming the problem of thermal transport model mismatch under varying operating conditions.

Description

Technical Field

[0001] This invention relates to the field of semiconductor lifetime prediction technology, and in particular to a method, device and medium for predicting the remaining lifetime of a power semiconductor module. Background Technology

[0002] Recent advancements in remaining lifetime prediction technology for power semiconductor modules have primarily focused on thermodynamic parameter monitoring and failure physics models. With the development of machine learning, data-driven methods are increasingly being applied to anomaly detection, using data collected on operating parameters such as temperature, current, and stress to train regression models for lifetime prediction. Some studies have incorporated finite element analysis to simulate thermomechanical stress distribution and combined it with the Coffin-Manson equation to calculate fatigue life. Furthermore, online monitoring technologies utilize real-time data acquisition of module on-state voltage drop and thermal resistance changes, employing Kalman filtering or particle filtering algorithms to estimate lifetime status.

[0003] Existing methods for predicting remaining lifetime have shortcomings. First, the lack of microscopic dynamic characterization makes it impossible to capture the nonlinear effects of phonon scattering and damage evolution mechanisms at the lattice scale, resulting in a disconnect between macroscopic lifetime models and microscopic failure mechanisms. Second, prediction models based on fixed physical equations are difficult to adapt to the dynamic response of entropy rate change caused by abrupt changes in phonon transport relaxation under varying operating conditions. Furthermore, data-driven methods rely on training with massive amounts of fault samples, and the real-time mapping relationship between pulse signals and damage evolution has not been established, resulting in a significant delay in early warning of sudden thermal failures. The fragmentation of spatial coverage and the empirical nature of lifetime prediction jointly constrain the timeliness of decision-making in high-reliability scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for predicting the remaining lifetime of power semiconductor modules to address the problems of missing phonon dynamics characterization and insufficient adaptive capability under varying operating conditions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for predicting the remaining lifetime of a power semiconductor module, comprising: generating a transport dataset from collected material lattice parameters and thermodynamic operating condition data using molecular dynamics simulation methods; generating a reciprocal space group velocity tensor field based on the transport dataset; generating a thermodynamic entropy change index based on the reciprocal space group velocity tensor field; performing time-series cumulative analysis based on the thermodynamic entropy change index to output a phonon entropy change feature sequence; extracting a statistical feature subset and a dynamic gradient subset from the phonon entropy change feature sequence; constructing a spiking neural network and inputting the phonon entropy change feature sequence to generate an operation function; calling the transport equation and initializing it based on the statistical feature subset to generate an initial transport equation; reconstructing the initial transport equation based on the operation function to generate a reconstructed transport equation; generating a thermoacoustic pulse signal from the dynamic gradient subset using Hilbert transform; solving the reconstructed transport equation in real time when the amplitude of the thermoacoustic pulse signal exceeds a preset amplitude threshold to output a lifetime decay gradient function; performing confidence interval density gradient analysis on the lifetime decay gradient function to generate a damage risk density cloud map and extracting high-density regions; and generating a remaining lifetime estimate and thermal damage location information when the high-density region exceeds a preset warning threshold.

[0008] As a preferred embodiment of the power semiconductor module remaining lifetime prediction method of the present invention, the specific steps for generating the reciprocal space group velocity tensor field are as follows:

[0009] Extract material lattice parameters from the product specification document of the target power semiconductor module; extract thermodynamic condition data from the operating history database;

[0010] Based on material lattice parameters and thermodynamic conditions, molecular dynamics simulations are run and atomic trajectory data are output. The transport dataset is then generated by Fourier transform.

[0011] A three-dimensional grid range is defined in a preset reciprocal space based on the transport dataset, and the reciprocal space group velocity tensor field is generated by interpolation calculation.

[0012] As a preferred embodiment of the power semiconductor module remaining lifetime prediction method of the present invention, the specific steps for extracting the statistical feature subset and the dynamic gradient subset are as follows:

[0013] Frequency-weighted energy distribution calculations are performed on the inverted space group velocity tensor field to generate a thermodynamic entropy change index.

[0014] A sliding window mean accumulation is performed on the thermodynamic entropy change index sequence to generate a phonon entropy change characteristic sequence.

[0015] Extract the mean and variance of the phonon entropy change feature sequence to generate a statistical feature subset; perform first-order difference on the phonon entropy change feature sequence to generate a dynamic gradient subset.

[0016] In a preferred embodiment of the power semiconductor module remaining lifetime prediction method of the present invention, the specific steps for generating the operation function are as follows:

[0017] The input layer is constructed based on the phonon entropy change feature sequence dimension, the neuron layer is constructed based on the leakage integral ignition characteristic, and the output layer is constructed based on the pulse-continuous value conversion.

[0018] Construct a spiking neural network based on an input layer, a neuron layer, and an output layer;

[0019] The phonon entropy change feature sequence is input into a spiking neural network, and the spiking neural network is solidified using the time backpropagation algorithm. The solidified spiking neural network is then defined as an operation function.

[0020] In a preferred embodiment of the power semiconductor module remaining lifetime prediction method of the present invention, the specific steps for generating the reconstructed transport equation are as follows:

[0021] The transport equation is invoked, and the thermal conductivity and specific heat capacity parameters of the transport equation are initialized based on a subset of statistical features to generate the initial transport equation;

[0022] The operation function is substituted into the initial transport equation to replace the collision term, generating the reconstructed transport equation.

[0023] As a preferred embodiment of the power semiconductor module remaining lifetime prediction method of the present invention, the specific steps of the output lifetime decay gradient function are as follows:

[0024] A Hilbert transform is performed on a subset of dynamic gradients to generate an analytic signal, and a thermoacoustic pulse signal is generated based on the analytic signal.

[0025] The amplitude of the thermoacoustic pulse signal is monitored in real time. When the amplitude of the thermoacoustic pulse signal exceeds the preset amplitude threshold, the thermodynamic working condition data is input into the reconstructed transport equation, and the reconstructed transport equation is solved in real time through the event-driven mechanism to generate the lattice damage accumulation rate field.

[0026] The lifetime decay gradient function is generated by differentiating the lattice damage accumulation rate field along the time dimension.

[0027] In a preferred embodiment of the power semiconductor module remaining lifetime prediction method of the present invention, the specific steps for generating the damage risk density cloud map are as follows:

[0028] The kernel density of the lifetime decay gradient function is estimated to generate the probability density distribution, and the confidence interval boundary of the probability density distribution is calculated.

[0029] The probability density distribution is mapped to the time-probability space, contour maps are drawn and confidence interval boundaries are marked, and damage risk density contour maps are output.

[0030] As a preferred embodiment of the power semiconductor module remaining lifetime prediction method of the present invention, the specific steps for generating the remaining lifetime estimate and thermal damage location information are as follows.

[0031] In the damage risk density cloud map, identify continuous regions with probability density greater than a preset high density threshold, calculate the center coordinates and area ratio, and generate a set of high density regions.

[0032] When the area ratio exceeds the preset warning threshold, the thermal damage location information is retrieved based on the center coordinates; the remaining lifetime estimate is generated by integrating the lifetime decay gradient function.

[0033] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the power semiconductor module remaining lifetime prediction method as described in the first aspect of the present invention.

[0034] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the power semiconductor module remaining lifetime prediction method as described in the first aspect of the present invention.

[0035] The beneficial effects of this invention are as follows: By dynamically reconstructing the collision terms of the transport equation through a pulsed neural network, accurate modeling of the nonlinear effect of phonon scattering is achieved, overcoming the problem of thermal transport model mismatch under varying operating conditions; thermoacoustic pulse signals are generated by analyzing dynamic gradient subsets using Hilbert transform, and combined with amplitude threshold-triggered event-driven solutions, achieving microsecond-level capture of the critical point of damage acceleration, breaking through the delay limitation of sudden thermal failure early warning; a four-dimensional spatiotemporal probability cloud map is constructed through confidence interval density gradient analysis, integrating high-density region dynamic tracking and gradient integral calculation, and simultaneously outputting micrometer-level positioning coordinates of thermal damage and remaining life estimation, eliminating monitoring blind spots and empirical prediction biases. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart of a method for predicting the remaining lifetime of power semiconductor modules.

[0038] Figure 2 The flowchart for generating the group velocity tensor field.

[0039] Figure 3 A flowchart for generating the reconstructed transport equations.

[0040] Figure 4 A flowchart for generating remaining lifetime estimates and thermal damage location information. Detailed Implementation

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0042] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0043] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0044] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a method for predicting the remaining lifetime of a power semiconductor module, comprising the following steps:

[0045] S1. The collected material lattice parameters and thermodynamic conditions data are used to generate a transport dataset through molecular dynamics simulation, and the reciprocal space group velocity tensor field is generated based on the transport dataset.

[0046] S1.1 Extract material lattice parameters from the product specification document of the target power semiconductor module; extract thermodynamic condition data from the operating history database;

[0047] It should be noted that the semiconductor chip material type (such as SiC or Si) declared in the product specification document is read, and parameters such as lattice constant, atomic mass and bonding length are extracted; for example, the lattice constant of SiC is 3.073 Å on the a-axis and 10.053 Å on the c-axis.

[0048] The system calls the historical database and retrieves historical operating data of the target power semiconductor module. It extracts the highest junction temperature, temperature rise rate, and power cycle count as thermodynamic operating data. The example junction temperature range is -40°C to 175°C.

[0049] S1.2 Based on material lattice parameters and thermodynamic conditions data, run molecular dynamics simulation and output atomic trajectory data, and generate transport dataset through Fourier transform;

[0050] It should be noted that the initial coordinate arrangement of atoms in three-dimensional space is determined based on the lattice constant in the material's lattice parameters, forming a periodic lattice structure; the atomic mass is assigned to the corresponding particle as the mass parameter in Newton's equations of motion; the equilibrium distance between atoms is calculated based on the bond length, and the force is defined using the mathematical relationship of the harmonic oscillator potential energy. Specifically, when the equilibrium distance between atoms deviates from the bond length, a linear restoring force is generated, and the magnitude of the linear restoring force is proportional to the offset distance. For example, in silicon carbide material, the silicon-carbon atom bond length is 1.89 angstroms. When the interatomic spacing becomes 1.95 angstroms, an attractive force is generated, the magnitude of which is the product of the proportionality constant (based on the definition of the material's elastic constant, the value of which reflects the physical properties of the atomic bond resisting deformation) and 0.06 angstroms.

[0051] The highest junction temperature from the thermodynamic operating condition data is used to set the temperature constraint value; the temperature increment for each time step is determined based on the temperature rise rate, specifically, the product of the temperature rise rate and the time step size is used as the single-step temperature change value, and the temperature change value is gradually accumulated according to the time step size over the total simulation time to achieve a linear heating process; the total simulation time is determined with reference to the number of power cycles; Newton's equations of motion are solved, and the atomic positions and velocities at each time step are updated through iterative calculations, outputting the recorded data of atomic positions and velocities changing over time, which is the atomic trajectory data; the expression for Newton's equations of motion is...

[0052] ;

[0053] in, Indicates other atoms to atoms The sum of the forces acting on the vector; Indicates the atom index number; Represents atoms The quality; Represents the differential operator; Represents atoms Position vector; Indicates a specific moment in time;

[0054] The atomic acceleration is obtained by the ratio of the sum of the forces acting on the atom to the atom's mass; the expression for calculating the atom's position at each time step is:

[0055] ;

[0056] in, Represents atoms exist Position vector at any given moment; Indicates time Time Atom position vector ； Represents atoms exist Position vector at any given moment; Indicates the time step; Indicates time Time Atom The atomic acceleration values;

[0057] Atom velocities at each time step are generated using the central difference method based on adjacent position vector data; velocity sequence data of all atoms changing over time is extracted from the atomic trajectory data; autocorrelation function processing is performed on the velocity sequence data of each atom to obtain the time-domain correlation curve showing the decay of velocity correlation strength over time. The expression for the velocity autocorrelation function is as follows:

[0058] ;

[0059] in, Indicates time delay The velocity correlation strength at that time; Indicates a time delay; Indicates the total simulation duration; Indicates time Time Atom The speed.

[0060] The time-domain correlation curve data sequence is subjected to spectral decomposition using the Fast Fourier Transform algorithm, converting the time-delay sequence into a frequency component sequence. Each frequency component outputs a complex amplitude containing both real and imaginary parts, representing the combined amplitude and phase information of the frequency component. The squared modulus of the complex amplitude is taken as the frequency-domain energy intensity, generating a numerical sequence of energy distribution values ​​for all frequency points within a frequency range from zero to the Nyquist limit (determined by the time step of the time-domain correlation curve, with the value being half the time step). This forms a frequency-domain energy distribution spectrum with frequency on the horizontal axis and energy intensity on the vertical axis. Local maxima with significantly higher energy intensities than neighboring frequencies are identified in the frequency-domain energy distribution spectrum, determining the peak frequency value of the peak frequency point. Based on the lattice constant and atomic mass, the peak frequency value is substituted into the mathematical expression of the phonon dispersion relation in the inherent properties of lattice dynamics to solve for the corresponding wave vector value. The group velocity value of the phonon energy propagation rate at the frequency point is obtained based on the wave vector value and the angular frequency of the phonon vibration mode, expressed as follows:

[0061] ;

[0062] ;

[0063] in, Represents the group velocity value; The angular frequency representing the phonon vibration mode; Indicates the peak frequency value; Indicates the wave vector value;

[0064] The peak frequency values ​​and group velocity values ​​of all peak frequency points are integrated to form a transport dataset containing the frequency-group velocity mapping relationship.

[0065] S1.3. Define a three-dimensional grid range in the preset reciprocal space based on the transport dataset, and generate the reciprocal space group velocity tensor field through interpolation calculation.

[0066] It should be noted that, with the origin of the coordinate system (based on the standard convention of solid-state physics, corresponding to the center of symmetry of the crystal lattice in real space) as the center, positive and negative boundary ranges are symmetrically set along each coordinate axis in the three-dimensional wave vector space to form a reciprocal three-dimensional grid. The product of the reciprocal of the lattice constant and pi is used as the maximum boundary value of the reciprocal three-dimensional grid in the wave vector direction to determine the boundary range of the reciprocal three-dimensional grid.

[0067] Within the boundary of the reciprocal three-dimensional mesh, uniformly spaced coordinate point sequences are generated along each coordinate axis; the coordinate point sequences of the three coordinate axes are combined to form a three-dimensional mesh point array;

[0068] Locate the coordinates of each grid point in the 3D grid array; construct a continuous smooth 3D surface passing through all reference nodes using cubic spline interpolation, with the wave vector point as the reference node; extend the continuous smooth 3D surface to cover the coordinates of each grid point in the 3D grid array, read the velocity vector values ​​of the continuous smooth 3D surface at the corresponding coordinate positions, and complete the mapping and transfer of group velocity values ​​from the wave vector point to the 3D grid array; integrate the velocity vector values ​​of all 3D grid arrays to form a reciprocal space group velocity tensor field covering the entire reciprocal space 3D grid.

[0069] S2. Generate a thermodynamic entropy change index based on the reciprocal group velocity tensor field; perform time-series cumulative analysis based on the thermodynamic entropy change index to output a phonon entropy change feature sequence; extract a statistical feature subset and a dynamic gradient subset from the phonon entropy change feature sequence.

[0070] S2.1 Calculate the frequency-weighted energy distribution of the velocity tensor field of the reciprocal space group to generate a thermodynamic entropy change index;

[0071] It should be noted that, based on the group velocity components and peak frequency values ​​in the reciprocal group velocity tensor field, the group velocity value corresponding to the peak frequency value is extracted at each location point in the reciprocal group velocity tensor field; the phonon occupancy probability weight is calculated for each peak frequency value, expressed as follows:

[0072] ;

[0073] in, This indicates the probability weight of phonon occupancy; Represents the reduced Planck constant; Represents the Boltzmann constant; This represents the real-time junction temperature value;

[0074] The product of the group velocity value and the phonon occupancy probability weight at the corresponding frequency is used as the frequency-weighted group velocity value; the frequency-weighted group velocity value at each location point is integrated and accumulated along the frequency dimension to generate a thermodynamic entropy change index sequence.

[0075] S2.2. Accumulate the sliding window mean of the thermodynamic entropy change index sequence to generate the phonon entropy change characteristic sequence;

[0076] It should be noted that, based on the thermodynamic entropy change index sequence, a fixed-length time window is set. Starting from the beginning of the thermodynamic entropy change index sequence, the time window is slid forward by a fixed step size along the time axis. Within each time window position, the arithmetic mean of the thermodynamic entropy change indexes of the time points covered by each time window is calculated, and the arithmetic mean is used as the entropy change characteristic value of the current time window center point. The time windows are continuously slid until the end of the thermodynamic entropy change index sequence, and the entropy change characteristic values ​​of all time windows are connected to generate the phonon entropy change characteristic sequence.

[0077] S2.3 Extract the mean and variance of the phonon entropy change feature sequence to generate a statistical feature subset; perform first-order difference on the phonon entropy change feature sequence to generate a dynamic gradient subset.

[0078] It should be noted that the mean and variance of all data points in the phonon entropy change feature sequence are calculated using the mean formula and variance formula. The mean and variance are then integrated into a statistical feature subset. The expressions for calculating the mean and variance are as follows:

[0079] ;

[0080] ;

[0081] in, This represents the mean of the phonon entropy change characteristic sequence; The variance of the phonon entropy change characteristic sequence is represented; Indicates the first The deviation of the phonon entropy change eigenvalue of a data point from the mean is the deviation of a single data point. The total number of data points representing the phonon entropy change feature sequence; This represents the data point index, with a value of 1- ; Indicates the first Phonon entropy change eigenvalues ​​of each data point;

[0082] The first-order difference calculation of adjacent point values ​​is performed on the data points of the phonon entropy change feature sequence arranged in chronological order. That is, the difference between the value of each time point and the value of the previous time point is the rate of change between adjacent time points. The rate of change of all adjacent time points is integrated into a rate of change sequence, which is the dynamic gradient subset.

[0083] S3. Construct a spiking neural network and input the phonon entropy change feature sequence to generate an operation function; call the transport equation and initialize it based on a subset of statistical features to generate an initial transport equation; reconstruct the initial transport equation according to the operation function to generate a reconstructed transport equation.

[0084] S3.1 Construct an input layer based on the phonon entropy change feature sequence dimension, construct a neuron layer based on the leakage integral ignition characteristic, and construct an output layer based on the pulse-continuous value conversion; construct a spiking neural network based on the input layer, neuron layer, and output layer;

[0085] It should be noted that the total number of nodes required for the input layer is determined based on the total number of data points in the phonon entropy change feature sequence. The corresponding number of input layer nodes are arranged according to the input order of the data points in the phonon entropy change feature sequence, forming a node array structure that strictly matches the dimension of the phonon entropy change feature sequence, thus completing the hierarchical construction of the input layer.

[0086] Based on the membrane resistance, membrane capacitance, and threshold potential required for leakage integral ignition characteristics, the basic electrical parameters of the neuron nodes are set. A neural connection weight value is assigned to the connection between each neuron node and the input layer node. The neural connection weight value is a real scalar value, initially generated using a uniformly distributed random sampling method within a preset value range (based on physical limitations and mathematical optimization constraints of the neuron circuit, e.g., [-0.5, 0.5]), and stored in the parameter register, thus establishing the static signal transmission path from the input layer to the neuron layer. Based on the manufacturing constraints of the neuron circuit (derived from the physical process limits of semiconductor integrated circuits, specifically including wiring congestion limitations, clock synchronization requirements, and thermal density limits) and computational efficiency optimization requirements, a two-dimensional or three-dimensional mesh array is used as the basic topology template. Fixed spatial coordinate positions of each neuron node are allocated according to row and column order, forming a non-random and non-dynamically adjustable node position distribution structure, determining the static spatial architecture configuration of the neuron layer, and completing the construction of the neuron layer.

[0087] It should also be noted that membrane resistance, membrane capacitance, and threshold potential are all inherent properties of neuronal circuits, and are defined according to the neuronal circuit definition specification. Membrane resistance controls the rate of potential decay, membrane capacitance controls the slope of potential rise, and threshold potential determines the pulse triggering critical point. The leakage integral ignition characteristic describes the electrical behavior of neuronal membrane potential changes over time, specifically: when the neuron receives input current, the membrane potential rises; when there is no input current, the membrane potential decays exponentially; when the membrane potential exceeds the threshold potential, a pulse discharge is triggered, generating a discrete pulse event sequence, and the potential is reset to the baseline value.

[0088] Based on the pulse-to-continuous value conversion rule, the number of output layer nodes is set to a single node; an output connection weight value is assigned to the connection path between each neuron node and the single node of the output layer. The output connection weight value is generated by uniformly distributed random sampling within a preset value range (such as [-0.5, 0.5]) and stored in the parameter register to complete the establishment of the static signal transmission path from the neuron layer to the output layer.

[0089] The pulse-to-continuous value conversion rule requires integrating the discrete pulse event sequence output by the neuron layer into a continuous numerical signal. The pulse-to-continuous value conversion rule forcibly limits the number of output layer nodes to a single node in order to achieve the weighted aggregation function of all discrete pulse events. Specifically, during the output layer construction stage, it is directly declared that the number of output layer nodes is fixed at 1, and the single node of the output layer is configured as a linear weighted summer. The input connection covers all nodes of the neuron layer, and the output value is the sum of the product of the total number of pulse events in the discrete pulse event sequence and the output connection weight value, thereby mapping the discrete pulse event sequence into a continuous scalar output and completing the hierarchical construction of the output layer.

[0090] A spiking neural network is constructed based on the input layer, neuron layer, and output layer. During the training phase, the number of iterations is preset based on the training data scale and hardware computing power. The mean squared error or cross-entropy loss function is selected according to the task requirement type (lifetime prediction, fault classification) and the structure optimization objective is preset, keeping the structure of input layer nodes, neuron nodes, and single nodes in the output layer unchanged. The values ​​of neural connection weights and output connection weights are adjusted. The membrane resistance, membrane capacitance, and threshold potential values ​​of neuron nodes are locked. The static signal transmission paths and hierarchical structure between all layers are kept fixed. The mean squared error or cross-entropy loss function is minimized or the minimum number of iterations is reached only by reconfiguring the values ​​of neural connection weights and output connection weights, thus completing the training of the spiking neural network.

[0091] S3.2 Input the phonon entropy change feature sequence into the spiking neural network, solidify the spiking neural network through the time backpropagation algorithm, and define the solidified spiking neural network as the operation function;

[0092] It should be noted that the phonon entropy change feature sequence is input into the trained spiking neural network to perform forward computation; the neural connection weights and output connection weights of the spiking neural network are locked in a non-updateable state, so that the spiking neural network's pulse-to-continuous value conversion rule is completely solidified.

[0093] After the spiking neural network is solidified, the complete computational path from the input layer nodes to the neuron nodes and then to the single output layer node, representing the phonon entropy change feature sequence, is declared as a deterministic mathematical mapping rule from input to output. Specifically, each data point in the phonon entropy change feature sequence activates the corresponding input layer node in the input order; the neuron node generates a discrete pulse event sequence; and the single output layer node sums the discrete pulse events to output a continuous value. The rule combination of all data transformation links in the computational path is refined into a one-to-one mathematical function relationship between input and output values, generating independently callable operation function entities. The expression of the operation function is...

[0094] ;

[0095] ;

[0096] in, This represents the intensity value of the reconstructed collision term, which is essentially the energy dissipation rate caused by phonon scattering. This indicates the output connection weight values; Neuron layer impulse triggering function; This represents the numerical value of neural connection weights; Represents the phonon entropy change characteristic sequence; Indicates the neuronal membrane potential; Indicates the threshold potential; This indicates that a pulse discharge is triggered, generating a discrete pulse event sequence. Indicates a resting state.

[0097] S3.3 Call the transport equation, initialize the thermal conductivity and specific heat capacity parameters of the transport equation based on the statistical feature subset, and generate the initial transport equation;

[0098] It should be noted that, based on a statistical feature subset (mean and variance), the thermal conductivity and specific heat capacity parameters of the transport equation are initialized as follows: the product of the mean and the lattice density (calculated based on the lattice constant and atomic mass) and the group velocity value is assigned to the thermal conductivity parameter; the ratio of the product of the variance and the atomic mass to the Boltzmann constant (a fundamental constant of statistical thermodynamics, axiomatic benchmark value in physics) is assigned to the specific heat capacity parameter; the thermal conductivity and specific heat capacity parameters are then written into the transport equation to generate the initial transport equation.

[0099] The formula for calculating the thermal conductivity parameter is as follows:

[0100] ;

[0101] ;

[0102] in, This represents the thermal conductivity parameter; Indicates lattice density; Represents the group velocity value; Indicates the number of atoms in the crystal lattice; Represents the lattice volume;

[0103] The formula for calculating the specific heat capacity parameter is as follows:

[0104] ;

[0105] in, This represents the specific heat capacity parameter;

[0106] The initial transport equation is expressed as follows:

[0107] ;

[0108] in, Indicates the temperature field distribution; This represents the relaxation time, which is a fixed value of 1 ns. Represents the vector differential operator; This represents the rate of temperature change.

[0109] S3.4 Substitute the operation function into the replacement collision term in the initial transport equation to generate the reconstructed transport equation.

[0110] It should be noted that the collision terms in the initial transport equations Directly replace it with the operation function to generate the reconstructed transport equation, the expression of which is:

[0111] ;

[0112] Existing technologies construct thermal transport equations using fixed material parameters and linear collision terms. Their advantages lie in high computational efficiency (fast solution of analytical equations) and ease of engineering deployment (few parameters and simple logic). However, their fundamental drawback is that static parameters cannot respond to the dynamic evolution of microstructures, and linear collision terms cannot characterize the nonlinear effects of phonon scattering, leading to distorted thermal field predictions under varying operating conditions. This new approach, on the other hand, dynamically drives the initialization of thermal property parameters (mapping thermal conductivity and specific heat capacity parameters to mean and variance) and reconstructs collision terms using a spiking neural network (operation functions replace analytical terms). This achieves adaptive thermal transport modeling for material degradation and accurate characterization of the nonlinear mechanism of phonon scattering. It overcomes the two major bottlenecks of operating condition mismatch and excessive physical simplification in existing technologies, achieving a leap in accuracy and a breakthrough in universality across complex scenarios while maintaining the same computational efficiency. This provides an industrial-grade, reliable tool for power semiconductor lifetime prediction.

[0113] S4. The dynamic gradient subset generates a thermoacoustic pulse signal through Hilbert transform. When the amplitude of the thermoacoustic pulse signal exceeds the preset amplitude threshold, the transport equation is solved and reconstructed in real time, and the lifetime decay gradient function is output.

[0114] S4.1 Perform Hilbert transform on the dynamic gradient subset to generate an analytic signal, and generate a thermoacoustic pulse signal based on the analytic signal;

[0115] It should be noted that the Hilbert transform is performed on the dynamic gradient subset sequence to generate an analytic signal complex sequence: the dynamic gradient subset sequence is taken as the real part sequence; a fast Fourier transform is performed on the dynamic gradient subset sequence to transform it to the frequency domain; in the frequency domain, all positive frequency components are multiplied by the imaginary unit, and all negative frequency components are multiplied by the negative imaginary unit to achieve a 90-degree phase shift of the components across the entire frequency domain; an inverse Fourier transform is performed on the frequency domain after the phase shift to transform it back to the time domain, generating the orthogonal signal component sequence corresponding to the dynamic gradient subset sequence as the imaginary part sequence; the real part sequence and the imaginary part sequence are combined one-to-one according to the time points to form a complex sequence, generating an analytic signal whose real part is the value of the dynamic gradient subset sequence and whose imaginary part is the orthogonal signal component sequence after a 90-degree phase shift; the analytic signal is directly defined as a thermoacoustic pulse signal.

[0116] S4.2 Real-time monitoring of thermoacoustic pulse signal amplitude. When the thermoacoustic pulse signal amplitude exceeds the preset amplitude threshold, thermodynamic operating condition data is input into the reconstructed transport equation, and the reconstructed transport equation is solved in real time through the event-driven mechanism to generate the lattice damage accumulation rate field.

[0117] It should be noted that the amplitude threshold is preset based on the material fatigue limit theory and industrial accelerated aging test data. Specifically, the median amplitude of the thermoacoustic pulse signal corresponding to the critical point of material thermal damage is measured by power cycling test, and the product of the median amplitude and the safety margin coefficient (based on the risk level definition of the application scenario) is used as the amplitude threshold value; the amplitude threshold value is stored in the threshold register.

[0118] The amplitude of the thermoacoustic pulse signal at each data point in the thermoacoustic pulse signal is monitored in real time. When the amplitude of the thermoacoustic pulse signal at any time exceeds the amplitude threshold, the current real-time junction temperature value is used as the initial temperature field, and an event-driven solution instruction is immediately triggered. The temperature change rate is obtained by solving and reconstructing the transport equation. The temperature change rate is integrated along the time dimension to generate the spatiotemporal distribution of the temperature field. The temperature gradient field is calculated based on the spatiotemporal distribution of the temperature field. The lattice damage accumulation rate field is generated according to the product of the temperature gradient field and the damage coefficient.

[0119] The expression for the spatiotemporal distribution of the generated temperature field is:

[0120] ;

[0121] in, The spatial coordinate vector is (x, y, z) and the time is... The spatiotemporal distribution of the temperature field at time refers to the temperature value of each data point at a certain moment; This represents the initial temperature field, and its value is equal to the current real-time junction temperature value. Indicates the start time of integration; Indicates the time when integration stops; This represents a spatial coordinate vector, i.e., (x, y, z);

[0122] The expression for calculating the temperature gradient field is:

[0123] ;

[0124] in, Represents the temperature gradient field; express directional temperature change rate; express directional temperature change rate; express directional temperature change rate;

[0125] The expression for the lattice damage accumulation rate field is:

[0126] ;

[0127] in, This represents the lattice damage accumulation rate field; This represents the damage coefficient, defined based on the material's fatigue characteristics.

[0128] S4.3. Differentiate the lattice damage accumulation rate field along the time dimension to generate the lifetime decay gradient function.

[0129] It should be noted that numerical difference calculation is performed on the lattice damage accumulation rate value at each location point in the lattice damage accumulation rate field: the lattice damage accumulation rate values ​​of adjacent time steps are extracted in chronological order, and the difference between the lattice damage accumulation rate value at the next time step and the lattice damage accumulation rate value at the previous time step is taken as the rate change. The ratio of the rate change to the time interval between adjacent time steps is taken as the lifetime decay gradient function value of the current location point in the current time interval (i.e., the time rate of change of the lattice damage accumulation rate). This numerical difference calculation is repeated for all locations to generate a three-dimensional data field of lifetime decay gradient function values, which is essentially the spatial distribution of lattice damage accumulation acceleration. The expression for the lifetime decay gradient function is...

[0130] ;

[0131] in, The lifetime decay gradient function refers to the function whose spatial coordinate vector is (x, y, z) and whose time is... Spatial distribution of lattice damage accumulation acceleration at time; The rate of change of the accumulation rate of lattice damage over time.

[0132] Existing technologies achieve coarse-grained prediction of lifespan trends by continuously monitoring thermal signals and relying on analytical models to calculate damage rates, solving the basic aging assessment problem with low computational overhead. However, their fundamental drawback lies in the delayed signal capture, leading to delayed early warnings and the inability to quantify the dynamic evolution acceleration of damage, resulting in a higher rate of missed detections for sudden thermal failures. In contrast, this solution integrates thermoacoustic pulse event-driven and damage acceleration analysis: it extracts the amplitude of thermoacoustic pulses in real time through Hilbert transform, and triggers instantaneous solution when the amplitude of the thermoacoustic pulse signal exceeds the amplitude threshold, transforming the damage rate field into a damage acceleration field. This not only significantly improves the early warning response speed but also reveals the spatiotemporal accelerated evolution law of lattice damage, solving the problem of early warning of synergistic failure caused by transient thermal shock and fatigue accumulation.

[0133] S5. Perform confidence interval density gradient analysis on the lifetime decay gradient function to generate a damage risk density cloud map and extract high-density areas; when the high-density area exceeds the preset warning threshold, generate the remaining lifetime estimate and thermal damage location information.

[0134] S5.1 Perform kernel density estimation on the lifetime decay gradient function to generate the probability density distribution, and calculate the confidence interval boundary of the probability density distribution;

[0135] It should be noted that, based on the three-dimensional data field of the lifetime decay gradient function value, Gaussian kernel density estimation is performed on all data points to extract the three-dimensional spatial coordinates (x, y, z) of each data point as the first three-dimensional components of the coordinate system; the data acquisition time point corresponding to each data point is extracted as the fourth-dimensional component of the coordinate system, and combined with the three-dimensional spatial coordinates (x, y, z) to form a four-dimensional coordinate (x, y, z, t) to represent a single spatiotemporal point; the corresponding lifetime decay gradient function value is then assigned to this single spatiotemporal point. This forms coordinate points with attribute associations; a combination operation is performed on all data points to generate a four-dimensional spatiotemporal coordinate point set consisting of multiple (x,y,z,t) quadruples;

[0136] For each data point (x, y, z, t) in the four-dimensional spatiotemporal coordinate set, Gaussian kernel density estimation is performed based on the bandwidth parameter: The distance between the target location point and all points in the four-dimensional spatiotemporal coordinate set is calculated; the distance values ​​are input into the Gaussian kernel function to calculate the weight values; the lifetime decay gradient function values ​​of all points are multiplied and summed with their corresponding weight values; the ratio of the summation result to the total number of points in the four-dimensional spatiotemporal coordinate set is used as the probability density estimate of the target location point; the Gaussian kernel density smoothing estimation operation is repeated for all data in the four-dimensional spatiotemporal coordinate set, outputting a set of probability density estimates to form a probability density distribution; the expression for calculating the weight values ​​is...

[0137] ;

[0138] The expression for calculating the probability density estimate is as follows:

[0139] ;

[0140] in, Indicates the first The weight values ​​of each data point; Indicates the target point The probability density estimate; This represents the index of a four-dimensional spacetime coordinate point set, with values ​​ranging from 1 to... ; This represents the bandwidth parameter value, defined based on the Silverman optimal bandwidth criterion, for example, 0.25; Indicates the target point and the first The distance values ​​between each data point; This represents the total number of points in the four-dimensional spacetime coordinate set; Indicates the coordinates of the target point; Indicates the first The coordinates of each data point; Indicates the first The lifetime decay gradient function values ​​of the coordinates of each data point;

[0141] Bootstrap resampling statistical calculations are performed on the probability density distribution: A sample subset of the same length as the probability density estimate set is randomly selected from the probability density estimate set (repeated sampling of the same value is allowed; for example, the probability density estimate set is [0.1, 0.2, 0.3], and the sample subset is [0.1, 0.1, 0.2]; the probability density estimate set and the sample subset are only the same length). The 95th quantile value of the sample subset is calculated as the confidence boundary value for a single resampling. Specifically, all values ​​in the sample subset are completely sorted in ascending order to form an ordered numerical sequence. The total length of the ordered numerical sequence is multiplied by 0.95. If the product result is an integer, the ordered value corresponding to the integer index position is taken as the 95th quantile. If the product result is not an integer, the two ordered values ​​corresponding to the floor index position and the floor index position of the product result are taken, and the weighted average of the two ordered values ​​is calculated by linear interpolation as the 95th quantile.

[0142] Repeat the random sampling and 95th percentile calculation operations 1000 times to obtain 1000 confidence boundary values; sort all confidence boundary values ​​in ascending order, and take the 950th value after sorting as the final 95% confidence level confidence interval boundary value output; the 95% confidence level is defined according to the statistical test risk control requirements, and it is stipulated that the parameter confidence interval must cover 95% probability to ensure a balance between detection sensitivity and false alarm rate.

[0143] S5.2 Map the probability density distribution to the time-probability space, draw contour maps and mark the confidence interval boundaries, and output the damage risk density map;

[0144] It should be noted that, based on the probability density distribution, the time series and probability density estimate of each spatial location point are extracted; a time-probability two-dimensional coordinate system is constructed with the time series as the x-axis and the probability density estimate as the y-axis; in the time-probability two-dimensional coordinate system, slices are taken at each fixed time point, and the magnitude of the probability density estimate of each spatial location point in the xyz direction is calculated as a scalar value. Specifically, the three components of the probability density estimate are squared, summed, and the square root is taken to generate the magnitude value of the corresponding spatial location point; the ratio of the magnitude value to the maximum magnitude value in the density change rate field is used as a normalization process to obtain a relative intensity coefficient between 0 and 1; the relative intensity coefficient is multiplied by 255 (which is the maximum value of 8-bit RGB color channels, based on the definition of digital image processing standards and RGB color space specifications) and rounded to obtain the RGB color value of the red channel (the green and blue channels are fixed at 0), generating the color attribute of the corresponding spatial location point; the magnitude value calculation and color attribute calculation are repeated to generate a continuous color distribution surface;

[0145] Plot the contour map of probability density estimates on the color distribution surface. Specifically, connect spatial points with the same probability density estimates to form closed contour lines. Mark the boundary values ​​of the 95% confidence interval as red dashed lines and overlay them onto the contour map. Repeat the plotting and labeling operation for each time slice, integrate the contour map sequence of all time slices, and output the damage risk density contour map.

[0146] S5.3 Identify continuous regions with probability density greater than the preset high density threshold in the damage risk density cloud map, calculate the center coordinates and area ratio, and generate a set of high density regions.

[0147] It should be noted that the probability density estimate of each time slice in the damage risk density cloud map is scanned point by point. When the probability density estimate of a certain location point is greater than the high density threshold (based on the damage risk level preset in the industrial reliability standard), it is marked as a high density point. High density areas are formed by connecting adjacent high density points by traversing adjacent points.

[0148] Calculate the geometric center coordinates and area proportion of high-density regions: Extract the three-dimensional coordinate data of all locations within the high-density region, sum the xyz coordinate values ​​of all locations within the high-density region, and use the ratio of the summation result to the total number of locations within the high-density region as the xyz center coordinates; use the ratio of the total number of locations within the high-density region to the total number of locations in the damage risk density cloud map as the area proportion; integrate the high-density regions of all time slices to generate a high-density region set.

[0149] S5.4 When the area ratio exceeds the preset warning threshold, the thermal damage location information is inverted based on the center coordinates; the remaining lifetime estimate is generated by integrating the lifetime decay gradient function.

[0150] It should be noted that when the area proportion of the high-density region set exceeds the warning threshold (defined based on the industrial reliability and safety level requirements of the target field), the center coordinates of the corresponding high-density region are extracted as the thermal damage location information output; the discrete sequence of the lifetime decay gradient function in the high-density region over a continuous time period is obtained, and numerical integration is performed on the discrete sequence of the lifetime decay gradient function: the product of the lifetime decay gradient function value and the time step size at all time steps is summed, and the ratio of the summation result to the normalization factor (defined based on the standard value of the maximum allowable damage rate of the material) is used as the cumulative damage value; the remaining lifetime estimate is generated based on the ratio of the preset lifetime failure threshold (defined based on the standard value of material fatigue fracture characteristics) to the cumulative damage value.

[0151] It should also be noted that existing technologies achieve thermal damage early warning through local sensor networks and statistical lifetime models, solving the basic risk perception problem and reducing the probability of sudden failure. However, the fundamental flaw lies in the fragmentation of spatial coverage and the empirical nature of lifetime prediction, leading to both excessive maintenance and sudden failures. In contrast, this solution constructs a dynamic tracking system for the evolution of damage across the entire domain. It achieves the location of the core damage area by fusing spatiotemporal probability distribution with damage risk density cloud maps. Combined with gradient integral normalization and critical value inversion, it breaks through the constraints of empirical lifetime prediction models and overcomes the dual bottlenecks of spatial blind spots and temporal distortion in existing technologies.

[0152] This embodiment also provides a computer device applicable to the power semiconductor module remaining lifetime prediction method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the power semiconductor module remaining lifetime prediction method proposed in the above embodiment.

[0153] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0154] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the method for predicting the remaining lifetime of a power semiconductor module as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0155] In summary, this invention achieves accurate modeling of phonon scattering nonlinear effects by dynamically reconstructing collision terms in the transport equation using a pulsed neural network, overcoming the challenge of thermal transport model mismatch under varying operating conditions; it generates thermoacoustic pulse signals by analyzing dynamic gradient subsets using Hilbert transform, combined with amplitude threshold-triggered event-driven solutions, achieving microsecond-level capture of the critical point of damage acceleration, breaking through the delay limitation of sudden thermal failure early warning; and it constructs a four-dimensional spatiotemporal probability cloud map through confidence interval density gradient analysis, integrating high-density region dynamic tracking and gradient integration calculations to simultaneously output micrometer-level positioning coordinates of thermal damage and remaining lifetime estimates, eliminating monitoring blind spots and empirical prediction biases.

[0156] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the remaining lifetime of a power semiconductor module, characterized in that: include, The collected material lattice parameters and thermodynamic conditions data are used to generate transport datasets through molecular dynamics simulations, and the reciprocal space group velocity tensor field is generated based on the transport datasets. Thermodynamic entropy change index is generated based on the reciprocal group velocity tensor field; time-series cumulative analysis is performed based on the thermodynamic entropy change index to output the phonon entropy change characteristic sequence. Extract statistical feature subsets and dynamic gradient subsets from phonon entropy change feature sequences; A spiking neural network is constructed and the phonon entropy change feature sequence is input to generate an operation function; the transport equation is called and initialized based on a subset of statistical features to generate an initial transport equation; the initial transport equation is reconstructed according to the operation function to generate a reconstructed transport equation; The dynamic gradient subset generates a thermoacoustic pulse signal through Hilbert transform. When the amplitude of the thermoacoustic pulse signal exceeds the preset amplitude threshold, the transport equation is solved and reconstructed in real time, and the lifetime decay gradient function is output. A confidence interval density gradient analysis is performed on the lifetime decay gradient function to generate a damage risk density cloud map and extract high-density areas. When the high-density area exceeds the preset warning threshold, the remaining lifetime estimate and thermal damage location information are generated.

2. The method for predicting the remaining lifetime of a power semiconductor module as described in claim 1, characterized in that: The specific steps for generating the reciprocal space group velocity tensor field are as follows. Extract material lattice parameters from the product specification document of the target power semiconductor module; extract thermodynamic condition data from the operating history database; Based on material lattice parameters and thermodynamic conditions, molecular dynamics simulations are run and atomic trajectory data are output. The transport dataset is then generated by Fourier transform. A three-dimensional grid range is defined in a preset reciprocal space based on the transport dataset, and the reciprocal space group velocity tensor field is generated by interpolation calculation.

3. The method for predicting the remaining lifetime of a power semiconductor module as described in claim 2, characterized in that: The specific steps for extracting the statistical feature subset and the dynamic gradient subset are as follows. Frequency-weighted energy distribution calculations are performed on the inverted space group velocity tensor field to generate a thermodynamic entropy change index. A sliding window mean accumulation is performed on the thermodynamic entropy change index sequence to generate a phonon entropy change characteristic sequence. Extract the mean and variance of the phonon entropy change feature sequence to generate a statistical feature subset; perform first-order difference on the phonon entropy change feature sequence to generate a dynamic gradient subset.

4. The method for predicting the remaining lifetime of a power semiconductor module as described in claim 3, characterized in that: The generation operation function involves the following specific steps. The input layer is constructed based on the phonon entropy change feature sequence dimension, the neuron layer is constructed based on the leakage integral ignition characteristic, and the output layer is constructed based on the pulse-continuous value conversion. Construct a spiking neural network based on an input layer, a neuron layer, and an output layer; The phonon entropy change feature sequence is input into a spiking neural network, and the spiking neural network is solidified using the time backpropagation algorithm. The solidified spiking neural network is then defined as an operation function.

5. The method for predicting the remaining lifetime of a power semiconductor module as described in claim 4, characterized in that: The specific steps for generating the reconstructed transport equations are as follows. The transport equation is invoked, and the thermal conductivity and specific heat capacity parameters of the transport equation are initialized based on a subset of statistical features to generate the initial transport equation; The operation function is substituted into the initial transport equation to replace the collision term, generating the reconstructed transport equation.

6. The method for predicting the remaining lifetime of a power semiconductor module as described in claim 5, characterized in that: The specific steps for defining the output lifetime decay gradient function are as follows: A Hilbert transform is performed on a subset of dynamic gradients to generate an analytic signal, and a thermoacoustic pulse signal is generated based on the analytic signal. The amplitude of the thermoacoustic pulse signal is monitored in real time. When the amplitude of the thermoacoustic pulse signal exceeds the preset amplitude threshold, the thermodynamic condition data is input into the reconstructed transport equation, and the reconstructed transport equation is solved in real time through the event-driven mechanism to generate the lattice damage accumulation rate field. The lifetime decay gradient function is generated by differentiating the lattice damage accumulation rate field along the time dimension.

7. The method for predicting the remaining lifetime of a power semiconductor module as described in claim 6, characterized in that: The specific steps for generating the damage risk density cloud map are as follows: The kernel density of the lifetime decay gradient function is estimated to generate the probability density distribution, and the confidence interval boundary of the probability density distribution is calculated. The probability density distribution is mapped to the time-probability space, contour maps are drawn and confidence interval boundaries are marked, and damage risk density contour maps are output.

8. The method for predicting the remaining lifetime of a power semiconductor module as described in claim 7, characterized in that: The specific steps for generating the remaining lifetime estimate and thermal damage location information are as follows. In the damage risk density cloud map, identify continuous regions with probability density greater than a preset high density threshold, calculate the center coordinates and area ratio, and generate a set of high density regions. When the area ratio exceeds the preset warning threshold, the thermal damage location information is retrieved based on the center coordinates; the remaining lifetime estimate is generated by integrating the lifetime decay gradient function.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the power semiconductor module remaining lifetime prediction method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the power semiconductor module remaining lifetime prediction method according to any one of claims 1 to 8.

Images