Power semiconductor life prediction method and system based on rain flow method

Abstract

The present invention relates to the technical field of equipment predictive maintenance, and in particular to a method and system for predicting the life of a power semiconductor based on the rain flow method. The method comprises the following steps: collecting semiconductor operation data, and performing temperature change calculation based on the semiconductor operation data to obtain semiconductor temperature change data; performing rain flow counting analysis based on the semiconductor temperature change data to obtain effective temperature cycle data; performing single cycle life evaluation based on the effective temperature cycle data to obtain single cycle life evaluation data; performing cumulative life calculation on the single cycle life evaluation data to obtain power semiconductor total cycle life data. The present invention not only improves the accuracy of life prediction, but also can effectively optimize maintenance plans, improve the reliability and safety of equipment, and has a wide range of engineering application value.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment predictive maintenance, and in particular to a method and system for predicting the life of a power semiconductor based on a rain flow method. Background Art

[0002] Power semiconductors (such as IGBT, MOSFET, SiC, GaN devices) have been widely used in new energy vehicles, power grid conversion equipment, industrial automation, power electronics systems and other fields. These devices work in high voltage, high current and high temperature environments. Long-term operation will lead to thermal stress, material aging, fatigue cracking of welding layers and other problems, which seriously affect the reliability and safety of the equipment. At present, the failure of power semiconductors is mainly caused by thermal-mechanical stress, among which thermal fatigue failure caused by temperature cycling is the key factor affecting the life. Therefore, how to accurately predict the remaining life of power semiconductors has become a problem. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes a power semiconductor life prediction method and system based on the rain flow method to solve at least one of the above technical problems.

[0004] The present application provides a method for predicting the life of a power semiconductor based on a rain flow method, comprising the following steps:

[0005] Step S1: collecting semiconductor operation data, and performing temperature change calculation based on the semiconductor operation data to obtain semiconductor temperature change data;

[0006] Step S2: Perform rain flow counting analysis based on the semiconductor temperature change data to obtain effective temperature cycle data;

[0007] Step S3: performing a single cycle life assessment according to the effective temperature cycle data to obtain single cycle life assessment data;

[0008] Step S4: Calculate the cumulative life of the single cycle life evaluation data to obtain the total cycle life data of the power semiconductor.

[0009] In the present invention, by combining temperature change calculation with rain flow counting analysis, the effective temperature cycle experienced by the semiconductor during operation can be accurately identified, the interference of invalid small fluctuations can be effectively eliminated, and the accuracy of life prediction can be improved. Using temperature change data as the core input parameter for life assessment can truly reflect the thermal stress of power semiconductors and make the prediction results more in line with the actual operating environment. The temperature cycles are classified and summarized by the rain flow counting method, and the cycles with the greatest impact on the life of the semiconductor are effectively screened out to cope with complex thermal loads and improve the efficiency of life analysis. The combination of single-cycle life assessment and cumulative life calculation can not only evaluate the impact of a single temperature cycle, but also comprehensively calculate the cumulative damage over the entire life cycle, thereby improving the reliability of the prediction.

[0010] Preferably, step S1 specifically includes:

[0011] Deploy low-power sensors at the semiconductor equipment end to collect semiconductor operation data, where the semiconductor operation data includes semiconductor operation current data, semiconductor operation voltage data, semiconductor operation temperature data, and semiconductor operation power loss data;

[0012] Extracting features based on semiconductor operation data to obtain semiconductor operation feature data;

[0013] Power loss modeling is performed based on semiconductor operation characteristic data to obtain semiconductor power loss curve data;

[0014] Perform thermal impedance network calculation based on semiconductor power loss curve data to obtain equivalent thermal impedance data;

[0015] The junction temperature change is calculated based on the semiconductor power loss curve data and the equivalent thermal impedance data to obtain the transient junction temperature curve data;

[0016] Timing prediction and temperature change extraction are performed based on transient junction temperature curve data to obtain semiconductor temperature change data.

[0017] In the present invention, more accurate power loss curve data can be obtained through power loss modeling. Compared with the traditional analysis method based on a fixed working point, this method can more realistically reflect the power loss of semiconductors under dynamic load conditions. The use of thermal impedance network calculation can comprehensively consider factors such as packaging structure, heat dissipation conditions, and dynamic changes in power loss to obtain accurate equivalent thermal impedance data and improve the modeling accuracy of junction temperature changes. The use of transient junction temperature curve calculation can capture rapid temperature changes in a short period of time and reflect the actual temperature fluctuations of semiconductors under dynamic working conditions.

[0018] Preferably, step S2 specifically includes:

[0019] Discretize the temperature waveform according to the semiconductor temperature change data to obtain discrete temperature waveform data;

[0020] Perform threshold segmentation analysis on discrete temperature waveform data to obtain temperature cycle marker data;

[0021] Rain flow count extraction is performed based on the temperature cycle marker data to obtain effective temperature cycle data.

[0022] The temperature change data is segmented in the present invention, which can effectively identify temperature cycles with significant effects and exclude small temperature fluctuations that do not affect the life of the device. The rain flow counting method is used to analyze the temperature cycle, which can identify the key temperature cycle mode and extract the temperature change data that has the greatest impact on fatigue life. It can effectively simulate the actual thermal fatigue damage accumulation process and is more accurate than traditional statistical methods. Due to the complex working mode of semiconductor devices, temperature fluctuations include random fluctuations, short-term changes, long-term trends and other characteristics. The temperature cycle marking and rain flow counting analysis of this method can effectively cope with complex working conditions and improve the robustness of the model. Through temperature waveform discretization and threshold segmentation, unnecessary data points are effectively reduced and calculation efficiency is improved. The rain flow counting method avoids redundant cyclic calculations, only retains temperature cycles that affect fatigue damage, improves analysis speed, and is suitable for large-scale semiconductor device monitoring data processing.

[0023] Preferably, the temperature waveform is discretized as follows:

[0024] Extract temperature variation characteristics according to semiconductor temperature variation data to obtain semiconductor temperature variation characteristic data;

[0025] Sampling semiconductor temperature change data according to semiconductor temperature change characteristic data to obtain semiconductor temperature change sampling data;

[0026] Perform multi-scale curvature analysis on the semiconductor temperature change sampling data to obtain the semiconductor temperature change trend matrix data;

[0027] Perform short-time waveform segmentation on the semiconductor temperature change sampling data to obtain short-time temperature waveform segment data;

[0028] Perform variational autoencoding on the semiconductor temperature change sampling data to obtain long-term temperature waveform characteristic data;

[0029] According to the semiconductor temperature change trend matrix data, short-term temperature waveform segment data and long-term temperature waveform feature data are subjected to Bayesian correlation to obtain inter-segment correlation data;

[0030] The waveform of the correlation data between the segments is discretized to obtain discrete temperature waveform data.

[0031] In the present invention, the key temperature change patterns that affect the life of the semiconductor are screened out by extracting temperature change features, reducing irrelevant data and improving computing efficiency. Temperature change data sampling is adopted to avoid redundant information, improve data processing speed, and retain key temperature change features, so that computing resources are used more efficiently. Through multi-scale curvature analysis, the temperature change trend under different time scales can be captured to ensure that both local short-term fluctuations and long-term trends can be identified. Short-term waveform segmentation can extract small temperature change patterns in a short time, which is helpful for analyzing the thermal shock effect of semiconductors under high-frequency switching. Variational autoencoder (VAE) can learn the long-term characteristics of semiconductor temperature waveforms and avoid prediction deviations caused by modeling based solely on instantaneous data. Combining short-term and long-term features makes the temperature waveform discretization more globally adaptable and improves the generalization ability of the prediction model. Through Bayesian association, the short-term waveform segment data and long-term feature data can be combined to dynamically adjust the temperature waveform correlation on different time scales to improve the matching accuracy of the temperature change pattern. The waveform discretization of the correlation data between segments can automatically screen out the most representative temperature change pattern, rather than simply dividing the time interval, to improve the discretization accuracy.

[0032] Preferably, the threshold segmentation analysis is specifically:

[0033] Perform Bayesian optimal segmentation on discrete temperature waveform data to obtain a preliminary segmentation data set;

[0034] Identify the temperature extreme points of the preliminary segmented data set to obtain segmented temperature extreme point data;

[0035] Perform local maximum-minimum contrast analysis based on style temperature extreme point data to obtain peak-valley matching data;

[0036] Perform segmented linear regression on the preliminary segmented data set based on the peak-valley matching data to obtain segmented temperature change interval data;

[0037] The temperature cycle marking data is obtained by classifying and marking the divided temperature change interval data.

[0038] In the present invention, the optimal segmentation point of the temperature waveform can be automatically determined through Bayesian optimal segmentation, avoiding the limitations brought by artificially setting fixed thresholds, being able to dynamically adapt to changes in the temperature waveform under different working conditions, and improving the flexibility and intelligence level of temperature segmentation. The temperature extreme point recognition is used to accurately locate the peak and valley points in the temperature cycle to ensure that the segmented temperature cycle can truly reflect the thermal load of the semiconductor. By matching the peak and valley data, it is ensured that each temperature cycle consists of a complete heating-cooling process to avoid erroneous counting caused by a single local extreme value. The piecewise linear regression method is used to fit and analyze the preliminary segmentation data, making the temperature change interval more reasonable, ensuring that the extracted temperature cycle accurately reflects the thermal load of the device, and being able to effectively identify the characteristics of different temperature change rates, which is particularly suitable for thermal cycle analysis of power semiconductors under different working conditions.

[0039] Preferably, the rain flow count extraction is specifically as follows:

[0040] Recursive interval merging is performed based on the temperature cycle marker data to obtain peak-valley pair processing data;

[0041] Perform sequence matching analysis on the peak-to-valley pair processed data to obtain half-cycle cycle data;

[0042] Merge the cycle intervals according to the half-cycle cycle data to obtain the merged temperature cycle data;

[0043] Calculate the cycle amplitude based on the combined temperature cycle data to obtain the temperature cycle amplitude matrix data;

[0044] Performing threshold adjustment on the temperature cycle amplitude matrix data to obtain temperature threshold data;

[0045] Perform effective cycle screening according to the temperature cycle amplitude matrix data and the temperature threshold data to obtain first effective temperature cycle data;

[0046] Count the cycle intensity stratified rain flow according to the half-cycle cycle data to obtain the second effective temperature cycle data;

[0047] The first effective temperature cycle data and the second effective temperature cycle data are sorted by correlation to obtain effective temperature cycle data.

[0048] In the present invention, by recursive interval merging, peak-valley pair processing is optimized, redundant or meaningless small fluctuations are eliminated, and it is ensured that the identified temperature cycle is the cycle that truly affects the fatigue life of the semiconductor. By using sequence matching analysis, half-cycle cycles can be automatically identified to ensure that the basic data of rain flow counting is more accurate. By merging cycle intervals, local small cycles that appear in a short period of time can be eliminated, and erroneous life calculations caused by noise or short-term fluctuations can be avoided, thereby improving data quality. By using cycle amplitude calculation, a temperature cycle amplitude matrix is ​​generated to accurately reflect the temperature change amplitude of different cycles, which is helpful to evaluate the magnitude of thermal stress on semiconductors. It can more finely characterize the relative influence of large cycles and small cycles than traditional methods, and improve the credibility of life assessment. By using threshold adjustment, the calculation threshold can be dynamically adjusted according to the actual application scenario, the adaptability of the algorithm can be improved, and misjudgment or missed judgment caused by fixed thresholds can be avoided. Cycle intensity layered rain flow counting can perform hierarchical analysis on temperature cycles of different intensities, and improve the hierarchical accuracy of life prediction. By sorting the correlation, the correlation between each temperature cycle can be calculated, and the most representative cycle can be further screened to improve the reliability of life assessment.

[0049] Preferably, the circulation intensity layered rain flow count is specifically:

[0050] Calculate the cycle intensity based on the half-cycle cycle data to obtain the temperature cycle intensity data;

[0051] Perform hierarchical clustering according to the temperature cycle intensity data to obtain the temperature cycle intensity stratified data;

[0052] Perform key cycle screening based on temperature cycle intensity stratification data to obtain strength stratification key cycle data;

[0053] According to the intensity stratified key cycle data, intensity stratified rain flow counting is performed to obtain the second effective temperature cycle number.

[0054] The cycle intensity calculation in the present invention can quantify the intensity of each temperature cycle, reflecting its contribution to the fatigue damage of power semiconductor materials. It pays more attention to high-intensity cycles than the traditional rain flow counting method, avoids the interference of small-amplitude cycles on life prediction, and improves the prediction accuracy. Through hierarchical clustering, temperature cycles of different intensities are classified to improve the hierarchical analysis capability of temperature cycle data. Different intensity levels can be automatically divided according to semiconductor material characteristics and workload patterns to ensure that high-intensity temperature cycles are not masked by small cycle noise. By using key cycle screening, the temperature cycles with the greatest impact on life can be automatically screened out, and irrelevant or low-impact cycles can be excluded, reducing the calculation cost, so that the life assessment is more focused on the cycle patterns that truly affect the fatigue life of power semiconductors. Through intensity-layered rain flow counting, the rain flow counting results of different intensity levels can be calculated separately, avoiding the defect of traditional methods that treat all cycles equally.

[0055] Preferably, step S3 is specifically:

[0056] Perform temperature cycle distribution processing according to effective temperature cycle data to obtain temperature cycle frequency distribution data;

[0057] Matching the temperature cycle frequency distribution data with a preset semiconductor lifetime model library to obtain a semiconductor lifetime matching model;

[0058] The single cycle lifetime calculation is performed according to the semiconductor lifetime matching model to obtain the single cycle lifetime evaluation data.

[0059] The present invention adopts temperature cycle distribution processing, performs frequency statistics on temperature cycles of different intensities, generates temperature cycle frequency distribution data, and ensures that the life calculation takes into account the influence of various cycle modes. By using life model library matching, the most suitable life calculation model can be selected according to different power semiconductor materials (Si, SiC, GaN, etc.) and different application environments (steady state / dynamic load). The single-cycle life calculation can calculate the impact of each effective temperature cycle on the device life according to the matching life model, thereby quantifying the fatigue damage of a single cycle, ensuring that the life assessment not only considers the number of temperature cycles, but also calculates the specific damage impact of a single cycle, avoiding overestimation or underestimation of the life.

[0060] Preferably, step S4 is specifically:

[0061] Perform damage accumulation calculation on single cycle life assessment data to obtain semiconductor cumulative damage data;

[0062] The total cycle life of the power semiconductor is calculated based on the semiconductor cumulative damage data to obtain the total cycle life data of the power semiconductor.

[0063] In the present invention, damage accumulation calculation is performed through single cycle life assessment data, and the contribution of each temperature cycle to the semiconductor life can be accurately evaluated, which conforms to the law of thermal fatigue damage accumulation. The total cycle life calculation is performed through semiconductor cumulative damage data, and the remaining useful life of the device can be directly obtained, which is convenient for intelligent maintenance management.

[0064] Preferably, the present application also provides a power semiconductor life prediction system based on the rain flow method, which is used to execute the power semiconductor life prediction method based on the rain flow method as described above. The power semiconductor life prediction system based on the rain flow method includes:

[0065] A semiconductor temperature change calculation module is used to collect semiconductor operation data and perform temperature change calculation based on the semiconductor operation data to obtain semiconductor temperature change data;

[0066] Rain flow counting analysis module, used to perform rain flow counting analysis based on semiconductor temperature change data to obtain effective temperature cycle data;

[0067] A single cycle life assessment module is used to perform a single cycle life assessment based on effective temperature cycle data to obtain single cycle life assessment data;

[0068] The cumulative life calculation module is used to perform cumulative life calculation on the single cycle life assessment data to obtain the total cycle life of the power semiconductor.

[0069] The beneficial effects of the present invention are: using low-power sensors to collect data such as current, voltage, temperature, power loss, etc. in real time, and combining with thermal impedance modeling to accurately calculate transient temperature changes and improve the accuracy of temperature monitoring. The optimized rain flow counting method ensures the scientific nature of fatigue damage calculation, accurately extracts effective temperature cycles, reduces the interference of invalid or noise cycles on life prediction, and improves calculation accuracy. High-precision single-cycle life calculation and optimized life modeling ensure that the most suitable life calculation method is provided for different materials (Si, SiC, GaN) and different application scenarios (steady-state load / dynamic load), improving the applicability and accuracy of the prediction. Intelligent cumulative life calculation improves the accuracy of life prediction, gradually accumulates the damage effects of different temperature cycles, calculates the total cycle life of power semiconductors, and provides more accurate remaining life estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting implementations made with reference to the following drawings:

[0071] Figure 1 A flowchart of a method for predicting the life of a power semiconductor based on the rain flow method according to an embodiment is shown;

[0072] Figure 2 A flowchart showing a method for calculating semiconductor temperature change according to an embodiment of the present invention is provided;

[0073] Figure 3 A flow chart showing the steps of a rain flow counting analysis method according to an embodiment is shown;

[0074] Figure 4 A flow chart showing the steps of a single cycle life assessment method according to an embodiment is shown. DETAILED DESCRIPTION

[0075] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0076] In addition, the drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0077] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0078] See also Figures 1 to 4 The present application provides a method for predicting the life of a power semiconductor based on the rain flow method, comprising the following steps:

[0079] Step S1: collecting semiconductor operation data, and performing temperature change calculation based on the semiconductor operation data to obtain semiconductor temperature change data;

[0080] Specifically, the operating data of power semiconductors under different working conditions are collected, including current data , voltage data , Ambient temperature data , using the equivalent thermal resistance and thermal capacitance model of power semiconductors to calculate the junction temperature (chip temperature) : ,in is the instantaneous power, The thermal resistance from junction to environment (i.e., the thermal impedance of the power semiconductor device from the chip junction to the environment, indicating that the device transfers heat to the external environment through the package and natural convection without additional heat sink or forced heat dissipation. The thermal resistance from junction to environment indicates the relationship between temperature change and power dissipation). The calculation formula is: .

[0081] Specifically, in an inverter module, multiple low-power sensors are deployed to collect key parameters of power semiconductors during actual operation. These sensors measure the operating current and operating voltage to calculate instantaneous power; ambient temperature and package temperature to monitor the impact of the external working environment; junction temperature sensors directly measure the internal temperature changes of power semiconductors. Based on the collected data, the system calculates the instantaneous junction temperature curve of the power semiconductor in combination with the thermal impedance model, and further analyzes the temperature change trend using a timing prediction algorithm or temperature change calculation to obtain temperature change data containing multiple time series.

[0082] Step S2: Perform rain flow counting analysis based on the semiconductor temperature change data to obtain effective temperature cycle data;

[0083] Specifically, find the local peak (maximum point) and local valley (minimum point) according to the temperature curve. According to the rules of rain flow counting, pair the temperature rise and fall intervals to form a temperature cycle. If some cycles are not closed (i.e., a complete rise-fall cycle is not formed), they are merged into adjacent cycles to obtain complete temperature cycle data.

[0084] Specifically, the data segmentation method is used to convert the continuous temperature change sequence into discrete temperature waveform data. By setting the minimum threshold of temperature change, small fluctuations are removed to screen out the main temperature peaks and valleys. The rain flow counting method is used to pair the temperature peaks and valleys, and extract complete temperature cycles, each of which contains the process of temperature rise and fall. The amplitude and frequency of each temperature cycle are counted to generate a temperature cycle data set.

[0085] Step S3: performing a single cycle life assessment according to the effective temperature cycle data to obtain single cycle life assessment data;

[0086] Specifically, for each temperature cycle, the fatigue life model of semiconductor devices is used to calculate the single cycle life. Based on the Coffin-Manson relationship: ,in is the life of a single cycle (i.e. the number of cycles it can withstand), is the material constant (determined by experimental data or information), is an exponential factor (values ​​range from 3 to 6), is the temperature change amplitude.

[0087] Specifically, since different power semiconductor materials and packaging processes will affect fatigue life, this embodiment selects the most matching life model from the life model library, such as a fatigue damage model based on the temperature cycle amplitude. According to the matching life model, the number of cycles that the semiconductor can withstand under each temperature cycle mode is calculated. Larger temperature changes will result in a shorter life, while small temperature cycles have little effect on life. After calculation, the system generates a set of single-cycle life data, in which each temperature cycle mode corresponds to a maximum number of cycles that can be tolerated.

[0088] Step S4: Calculate the cumulative life of the single cycle life evaluation data to obtain the total cycle life data of the power semiconductor.

[0089] Specifically, according to Miner's damage accumulation theory: ,in , the life of the semiconductor device ends. Calculate the total life: ,The total life is the reciprocal of the sum of the reciprocals of the damage of each temperature cycle.

[0090] Specifically, the linear damage accumulation method is used to comprehensively calculate the impact of all temperature cycles. The system calculates the total damage value based on the number of temperature cycles and its single cycle life, and determines the health status of the semiconductor. The total cycle life of the power semiconductor is calculated by the inverse of the cumulative damage. If the cumulative damage value is close to the preset failure threshold, it means that the semiconductor is about to reach the end of its service life and needs to be replaced or maintained. The system finally outputs the estimated total cycle life of the power semiconductor and generates a detailed life assessment report, including key parameters, failure trend prediction, etc., to provide a basis for equipment maintenance.

[0091] Preferably, step S1 specifically includes:

[0092] Step S11: deploying a low-power sensor at the semiconductor device end to collect semiconductor operation data, wherein the semiconductor operation data includes semiconductor operation current data, semiconductor operation voltage data, semiconductor operation temperature data, and semiconductor operation power loss data;

[0093] Specifically, a variety of low-power sensors are deployed inside the power module of the inverter, including current sensors, which measure the operating current of the semiconductor for calculating instantaneous power; voltage sensors, which measure the operating voltage of the semiconductor to assist in calculating power loss; temperature sensors, which are installed in the environment, package surface and chip nodes to collect external temperature and internal junction temperature data; and power sensors, which directly measure the power loss of the semiconductor for thermal calculation analysis.

[0094] Step S12: extracting features based on the semiconductor operation data to obtain semiconductor operation feature data;

[0095] Specifically, the instantaneous power calculation formula is: . Calculate the power rate of change: ,in, For the preset time window data, the characteristic parameters such as the mean, variance, maximum value, minimum value, etc. of the power are extracted respectively to form a characteristic vector.

[0096] Specifically, a sliding window filtering method is used to remove high-frequency interference and ensure data stability. Instantaneous power is calculated through current and voltage data, and the maximum, minimum, mean and change rate are counted. The temperature gradient of ambient temperature, package temperature and junction temperature is calculated to identify overheating risks. The power change rate in a short period of time is recorded to extract the power loss pattern.

[0097] Step S13: Perform power loss modeling according to semiconductor operation characteristic data to obtain semiconductor power loss curve data;

[0098] Specifically, power loss mainly comes from conduction loss and switching loss. Conduction loss: , is the conduction loss, is the conduction current, is the drain-source on-resistance of MOSFET or IGBT in the on-state, switching loss: , is the switching loss, is the drain-source voltage of the MOSFET or IGBT, is the switching current, is the switching frequency, is the switching time (s), including the on-time and off-time. Based on these power loss calculation formulas and the collected data, a power loss model is constructed: ,in is the total power loss (W), which varies with time, is the conduction loss weight coefficient, is the switching loss weight coefficient, is the compensation term for other losses (W), such as leakage loss or external environmental influence.

[0099] Specifically, obtaining conduction loss includes calculating the energy loss during conduction based on current and on-resistance; switching loss includes analyzing the power mutation of semiconductors at the moment of switching; leakage loss includes calculating the impact of leakage current on overall power consumption in a high temperature environment. Combining operating current, switching frequency and temperature data, a power loss curve is constructed and updated in real time.

[0100] Step S14: performing thermal impedance network calculation according to semiconductor power loss curve data to obtain equivalent thermal impedance data;

[0101] Specifically, a thermal impedance network model is established, and the thermal impedance of the device It can be expressed as:

[0102] ,in is the Laplace transform of the junction temperature, is the ambient temperature, is the Laplace transform of power loss. Use RC equivalent thermal network model, such as: ,in is the ambient temperature, is the total power loss (W), which varies with time, is the thermal resistance from junction to ambient (°C / W).

[0103] Specifically, obtain the thermal impedance parameters in the production line setting parameters, including junction-case thermal resistance, junction-ambient thermal resistance, package thermal capacitance, etc. Combined with the power loss curve, calculate the internal heat transfer path of the chip and simulate the heat conduction process. Adjust the thermal impedance parameters based on real-time data to adapt to different workloads and ambient temperature changes.

[0104] Step S15: Calculate the junction temperature change according to the semiconductor power loss curve data and the equivalent thermal impedance data to obtain transient junction temperature curve data;

[0105] Specifically, based on the RC thermal model, the transient junction temperature curve is calculated: ,in The junction temperature at time The junction temperature is the temperature inside the semiconductor, which is affected by power loss and changes over time. The ambient temperature is the temperature outside the semiconductor device, which is the reference value of the junction temperature change. The semiconductor junction temperature will change relative to the ambient temperature. is the power loss at time instant. The power loss is determined by the working state of the semiconductor device (such as conduction loss and switching loss). It is the thermal impedance of semiconductor devices, which is the thermal resistance between the junction temperature of semiconductor devices and the ambient temperature. It indicates the efficiency of heat transfer from the junction temperature to the environment. The unit is °С / W. is a natural constant, As time goes by, the semiconductor junction temperature gradually changes with the change of power loss. It is the thermal time constant, which reflects the time required for a semiconductor device to change from one state to another.

[0106] Specifically, the thermal impedance network is used to calculate how power loss affects the junction temperature and obtain the transient junction temperature curve. The time delay characteristics of temperature rise and fall are extracted to identify the temperature change pattern. Combined with load fluctuations, the maximum junction temperature reached under extreme conditions is inferred.

[0107] Step S16: Perform timing prediction and temperature change extraction based on the transient junction temperature curve data to obtain semiconductor temperature change data.

[0108] Specifically, Divide into segments and calculate the temperature change trend in a short period of time: ,in The transient junction temperature curve data Junction temperature at the moment (°C), The transient junction temperature curve data Junction temperature (°C) at time instant. Use polynomial regression or exponential smoothing to fit the temperature curve: ,in is the initial temperature (°C), i.e. the junction temperature at t=0, is the linear coefficient, which indicates the linear rate of change of temperature over time (℃ / s), that is, the initial trend of temperature change. is time (s), representing the independent variable of temperature change, is the quadratic term coefficient, which represents the acceleration effect of temperature change, that is, the nonlinear growth trend. is the higher-order coefficient, indicating the higher-order temperature change effect, is the polynomial order, which determines the complexity of the fitting function. Based on the temperature curve, the temperature change rate is calculated, peak and valley detection is performed, the local temperature maximum and minimum values ​​are found, and the difference calculation is performed to obtain the semiconductor temperature change data.

[0109] Specifically, the junction temperature curve is segmented and the temperature change pattern in a short period of time is extracted. The historical temperature data and environmental factors are combined to predict the future temperature change trend. The periodic pattern in the temperature change is identified, and key features such as the temperature cycle amplitude and duration are extracted.

[0110] Preferably, step S2 specifically includes:

[0111] Step S21: discretizing the temperature waveform according to the semiconductor temperature change data to obtain discrete temperature waveform data;

[0112] Specifically, the continuous temperature change curve is uniformly sampled, the time step is set, and the temperature curve is sampled at equal intervals to convert it into a set of discrete temperature point sequences. According to the sampling interval, the temperature data is interpolated to make the data points evenly distributed. Further extract key temperature features, including maximum temperature, minimum temperature, temperature change rate, etc., to form discrete temperature waveform data.

[0113] Step S22: performing threshold segmentation analysis on the discrete temperature waveform data to obtain temperature cycle mark data;

[0114] Specifically, the minimum temperature change threshold of the temperature cycle is set to filter invalid cycles. The local extreme points are calculated, and the temperature change amplitude between adjacent peaks and valleys is calculated. If the temperature change amplitude is less than the minimum temperature change threshold of the temperature cycle, the peak-valley pair is ignored and is not considered as part of the temperature cycle. Otherwise, the peak-valley pair is retained and marked as a valid temperature cycle.

[0115] Specifically, by scanning discrete temperature data, local maxima (peaks) and local minima (valleys) are identified. Minor fluctuations are excluded, and only significant peak and valley points that meet the set threshold are retained. A minimum temperature change threshold is set to filter small temperature changes to avoid misidentifying small fluctuations as valid cycles. Only temperature fluctuations that exceed the set threshold are recorded as part of a valid cycle. In the screened temperature data, adjacent peaks and valleys are matched to form a preliminary temperature cycle. The temperature change amplitude between each pair of peaks and valleys is calculated, and the duration of each cycle is recorded.

[0116] Step S23: extracting rain flow counts according to the temperature cycle mark data to obtain effective temperature cycle data.

[0117] Specifically, the rainflow counting method identifies all half-cycle patterns (i.e., a rising or falling part) by scanning the temperature change sequence. For each half-cycle, the temperature change amplitude and duration are recorded, and it is checked whether a complete cycle can be formed. By analyzing adjacent half-cycles, they are merged to form a complete temperature cycle. If a cycle has not been closed, it is temporarily retained and continued to be matched to form a complete cycle. The number of occurrences of cycles with different temperature amplitudes is calculated to form cycle frequency distribution data. Further screen high-frequency cycle patterns to determine the main temperature cycles that have a greater impact on semiconductor life.

[0118] Preferably, the temperature waveform is discretized as follows:

[0119] Extract temperature variation characteristics according to semiconductor temperature variation data to obtain semiconductor temperature variation characteristic data;

[0120] Specifically, by analyzing time series data, the rate of temperature change over time is calculated to identify rapid temperature rise or drop. The peak value, valley value and fluctuation range of temperature are recorded, key temperature change events that affect the life span are marked, and the characteristic data of semiconductor temperature change is obtained.

[0121] Sampling semiconductor temperature change data according to semiconductor temperature change characteristic data to obtain semiconductor temperature change sampling data;

[0122] Specifically, fixed-interval sampling is performed, and an initial sampling interval is set, such as selecting 1 millisecond or 10 milliseconds as the initial sampling point. Based on this setting, uniform sampling is performed on the time axis to obtain a basic temperature change sequence. The temperature change rate is calculated to capture areas with faster temperature changes. The temperature change rate is obtained by calculating the derivative of the temperature over time. When the temperature change rate is large, it means that the semiconductor is in a thermal shock state, and the sampling points need to be increased at this time; when the temperature change is close to zero, it means that the semiconductor temperature changes slowly, and the device enters a steady-state operation stage, and the sampling points can be reduced. Analyze the key features of the temperature curve and perform multi-scale curvature analysis. Curvature is the second-order derivative of the temperature change curve, which reveals the severity of the temperature change. If the curvature value is large, it means that the temperature has changed dramatically, such as in the switching transient or thermal equilibrium stage, and the data needs to be sampled with high density; if the curvature value is small, it means that the temperature changes slowly, and the sampling points can be reduced. During the sampling process, key temperature points are extracted, which include local extreme points, that is, the peak and valley values ​​of the temperature. Turning points are points where the direction of temperature change changes. These turning points can be derived through curvature analysis. Set the threshold change point. If the temperature change exceeds a certain threshold (for example, 5°C), a sampling point is automatically added. In order to further optimize the sampling efficiency, a dynamic time window adjustment strategy is adopted. Specifically, for different temperature changes, the sampling interval is dynamically adjusted, such as in the mutation area. When the temperature suddenly changes (for example, it suddenly rises from 25°C to 85°C), the sampling interval is reduced, for example, sampling once every 1 millisecond. In the steady-state area, if the temperature changes slowly (for example, from 60°C to 62°C), the sampling interval is increased, for example, sampling once every 100 milliseconds. Special working condition adjustment: if the power loss increases suddenly, the sampling density is forced to increase; if the external ambient temperature changes slowly, the sampling density is reduced.

[0123] Perform multi-scale curvature analysis on the semiconductor temperature change sampling data to obtain the semiconductor temperature change trend matrix data;

[0124] Specifically, the curvature calculation formula is adopted: ,in For the The curvature value at a time point is used to measure the curvature of the temperature curve at that point. A larger curvature indicates that the point is an area with a sudden change in temperature or an obvious change trend. For the The second-order derivative of temperature at a time point, that is, the acceleration of temperature change over time, when the value is large, it means that the temperature change rate changes dramatically at this point. For the The first derivative of the temperature at a time point, that is, the rate of change of temperature over time. If the first derivative is large, it means that the temperature is rising or falling rapidly. Use different scale parameters to calculate the curvature: ,in For the scale parameter Next, The curvature value at each time point. The curvature calculation reflects the short-term and long-term temperature trends. For time point The temperature value at The temperature of the time step, For time point The temperature value at, that is, the current temperature, For time point The temperature value at The temperature of the time step, Take different values ​​(such as 1, 2, 5, 10) and calculate the curvature at different time scales.

[0125] Specifically, the local change trend of the temperature change curve is analyzed to identify the inflection point position to determine the critical point of temperature rise and fall. The temperature curvature is calculated using different time scales to identify the relationship between short-term fluctuations and long-term trends.

[0126] Perform short-time waveform segmentation on the semiconductor temperature change sampling data to obtain short-time temperature waveform segment data;

[0127] Specifically, referring to the thermal equivalent circuit model of power semiconductors, two parameters, thermal resistance and thermal capacitance, are extracted. Thermal resistance is the resistance to heat transfer from the inside of the semiconductor to the external environment, and thermal capacitance refers to the amount of heat that the semiconductor needs to absorb or release when the unit temperature changes. The thermal time constant is calculated based on thermal resistance and thermal capacitance. The time constant reflects the time required for the semiconductor to reach a stable temperature after being heated. The thermal time constant calculation formula is: the thermal time constant is equal to the product of thermal resistance and thermal capacitance. For example, if the thermal resistance of a MOSFET is 0.5℃ / W and the thermal capacitance is 0.02J / ℃, the thermal time constant is 0.5 times 0.02, which is 0.01 seconds, indicating that the temperature of the semiconductor will change significantly within 0.01 seconds. According to the working state of the semiconductor, the basic time window is set. The basic time window refers to the time interval when collecting data. For example, in a stable operating state (low power consumption or constant power), the time window is set to 5 times the thermal time constant, and the window is larger, aiming to reduce the amount of calculation. For power fluctuations or switching states (high-frequency temperature changes), the time window is set to the length of the thermal time constant, and the window is smaller, which can improve the accuracy of the data. For sudden thermal events (such as instantaneous high power changes), the time window is set to be less than the length of the thermal time constant, and the window is very small to ensure that every detail of the change is captured. Through the set time window, the temperature data will be divided into multiple short-term segments. Based on the set time window starting point as the benchmark, the temperature data is traversed backward, and the size of the time window is dynamically adjusted according to different states. Calculate the temperature change rate in the current window, that is, the ratio of the temperature change amount to the time interval between the current time point and the next time point. When the temperature change rate is high (exceeding the set threshold), it means that the temperature changes violently. At this time, reduce the time window (for example, reduce the time window to half of the thermal time constant) to improve the accuracy of the data. When the temperature change rate is low (below the set threshold), it means that the temperature changes smoothly. At this time, the time window can be increased (for example, increased to 10 times the thermal time constant) to reduce the amount of data. In each adjusted window, the corresponding temperature data is intercepted as a short-term segment.

[0128] Specifically, according to the thermal inertia of power semiconductors, a time window is set to divide the temperature data into multiple short-time segments. In each short-time segment, the average temperature, temperature change rate and fluctuation range are calculated to characterize the local temperature characteristics.

[0129] Perform variational autoencoding on the semiconductor temperature change sampling data to obtain long-term temperature waveform characteristic data;

[0130] Specifically, in the encoding stage, a neural network encoder is used to extract the key features of the input short-term temperature waveform fragment data. The role of the encoder is to compress the input temperature data from a high dimension to a low-dimensional feature space, thereby extracting the core information of the temperature change. The time series data of the temperature change is converted into a low-dimensional feature vector, called a long-term temperature feature vector, which contains the global information of the temperature change. In the decoding stage, the original temperature data is reconstructed from the low-dimensional feature vector by the decoder. By comparing the reconstructed temperature data with the actual input data, the error between them is calculated and the model is optimized. The goal is to enable the decoder to accurately restore the original temperature waveform from the compressed feature vector, thereby ensuring that the extracted feature vector can retain the complete temperature feature information. After training and optimization, the low-dimensional feature vector extracted by the encoder becomes the long-term temperature waveform feature data. The data set contains the global temperature change features in each time period. Each feature vector represents the overall information of the temperature change in a specific time period. These feature vectors are stored as a set, called the long-term temperature feature set, which contains the temperature change features in each time period.

[0131] Specifically, the long-term variation pattern of temperature data is learned through the autoencoder method to obtain a high-dimensional representation of the temperature trend. The encoder reduces the data dimension, removes unimportant information, and improves the stability of data analysis.

[0132] According to the semiconductor temperature change trend matrix data, short-term temperature waveform segment data and long-term temperature waveform feature data are subjected to Bayesian correlation to obtain inter-segment correlation data;

[0133] Specifically, ,in For a given long-term temperature characteristic In the case of short-term temperature waveform The probability distribution of is the probability of the short-term waveform affecting the long-term temperature trend, is the prior probability of the short-term temperature segment, is the prior probability of the long-term temperature feature. The Kullback-Leibler (KL) divergence is calculated to measure the difference between two probability distributions P and Q: , The Kullback-Leibler divergence (KL divergence for short) is used to measure the difference between two probability distributions P and Q, and measures the information loss when using distribution Q to approximate distribution P. The value of KL divergence is always non-negative, and when P and Q are exactly the same, KL divergence is zero. is the probability distribution In Events The probability density function value or probability mass function value on the random variable In distribution The probability of the following happening, is the probability distribution In Events The probability density function value or probability mass function value on the random variable In distribution The probability of the following happening, It corresponds to the relationship between the short-term temperature waveform segment data and the long-term temperature waveform feature data, such as Represents the temperature value of a temperature waveform segment at a certain moment. Through the calculation of the above Bayesian association and KL divergence, the relationship between the short-term temperature waveform segment data and the long-term temperature waveform feature data is evaluated, thereby obtaining the correlation data between the segments. Calculate the conditional probability between the short-term waveform and the long-term temperature feature, and use the Bayesian theorem to calculate the conditional probability between the short-term waveform segment data and the long-term temperature waveform feature data, that is, . The difference between the short-term temperature waveform data and the long-term temperature feature data is calculated using the KL divergence formula. The smaller the KL divergence, the stronger the correlation between the short-term waveform and the long-term temperature feature. According to the KL divergence value, the correlation between different short-term temperature waveform segments and the long-term temperature waveform features is ranked. The smaller the KL divergence, the stronger the correlation between the segment and the long-term temperature feature, indicating that they have a higher consistency in the trend of temperature change.

[0134] Specifically, through probability distribution analysis, we evaluate whether the temperature change patterns at different time scales are related. We identify short-term fluctuations that have a greater impact on long-term trends to extract temperature change patterns that have a greater impact on lifespan.

[0135] The waveform of the correlation data between the segments is discretized to obtain discrete temperature waveform data.

[0136] Specifically, the temperature data is clustered according to the correlation data between the segments, similar waveforms are merged, and the main waveform patterns are extracted to obtain discrete temperature waveform data.

[0137] Specifically, based on the correlation between the segments, the temperature segments with similar patterns are merged to reduce data redundancy. By marking different types of temperature changes, a discrete temperature sequence is formed.

[0138] Preferably, the threshold segmentation analysis is specifically:

[0139] Perform Bayesian optimal segmentation on discrete temperature waveform data to obtain a preliminary segmentation data set;

[0140] Specifically, the input discrete temperature waveform data includes multiple temperature change points, each of which represents the temperature change at a specific moment and the time at that moment. For example, the temperature change and time data are recorded in a one-to-one manner, which can be expressed as a combination of several temperature changes and corresponding times. In the Bayesian optimal segmentation method, the temperature curve is composed of multiple piecewise linear models. The goal is to find the optimal set of segmentation points so that the temperature change error in each segment is minimized. The likelihood function is used for fitting, which reflects the degree of fit between the temperature data and the segmentation point. By fitting the data of each temperature segment, the matching degree of the temperature data of each segment with the segmentation point is calculated. In order to ensure the rationality of the segmentation point, the number of segmentation points is modeled using the Poisson process. The Poisson process has a smoothing factor that represents the distribution of the segmentation points and ensures that the number of segmentation points is appropriate. Through these two calculation steps, a posterior probability can be obtained, which represents the possibility of various segmentation points under given temperature data. By maximizing the posterior probability, the optimal set of segmentation points is found, and the error of each segment of the temperature curve is minimized by optimization, so as to obtain a set of segmentation points that best conforms to the actual situation.

[0141] Specifically, based on the working characteristics of power semiconductors, the temperature curve is segmented according to the change trend to avoid errors caused by artificially setting fixed segmentation points. The data probability model is used to calculate the optimal segmentation point to ensure that the temperature change trend in each segment remains stable. By dynamically optimizing and adjusting the segmentation point, the temperature change pattern in the same segment is similar, while the patterns between different segments are significantly different.

[0142] Identify the temperature extreme points of the preliminary segmented data set to obtain segmented temperature extreme point data;

[0143] Specifically, the segmented segments are scanned to find the area where the temperature rises to the highest point and then drops, and the point is marked as the peak. If the temperature difference between adjacent extreme points is less than the set minimum threshold, the fluctuation is considered to be noise and is not included in the extreme point list.

[0144] Perform local maximum-minimum contrast analysis based on style temperature extreme point data to obtain peak-valley matching data;

[0145] Specifically, the temperature variation between each pair of adjacent peak-valley pairs is calculated, and the significance of the temperature fluctuation is evaluated. If the temperature difference between some peak-valley points is less than the set threshold, it is considered that the fluctuation is not enough to affect the semiconductor life, and the pair of peak-valley points can be ignored. Only the peak-valley pairs with large temperature variation are retained to ensure that the calculation is based on the key temperature cycle mode.

[0146] Perform segmented linear regression on the preliminary segmented data set based on the peak-valley matching data to obtain segmented temperature change interval data;

[0147] Specifically, in each temperature cycle interval, a linear trend is fitted to describe the rate of temperature rise or fall. If the regression result shows that the temperature change rate is close to zero, then the interval is a stable stage of semiconductor temperature. If the regression result shows that the temperature change rate is large, then the interval belongs to a rapid heating or cooling stage, which has a greater impact on the semiconductor life. If the temperature change trend in some intervals does not conform to the typical pattern, the split point needs to be re-optimized to ensure accuracy.

[0148] The temperature cycle marking data is obtained by classifying and marking the divided temperature change interval data.

[0149] Specifically, the change trend of the temperature change interval is classified. Temperature changes are classified by analyzing their change rate (i.e., the rate at which the temperature changes over time). According to the direction and amplitude of the temperature change, it is divided into the following stages: the heating stage, when the temperature change rate is greater than zero, it means that the temperature is rising, and the temperature change is an upward trend. The cooling stage, when the temperature change rate is less than zero, it means that the temperature is falling, and the temperature change is a downward trend. The stable stage, when the temperature change rate is close to zero, it means that the temperature remains almost unchanged, which belongs to the stage of small temperature change. Based on these classifications, the next step is to perform heating → cooling pattern recognition, and mark each temperature cycle according to the law of temperature change. Mark the temperature cycle, such as marking each heating and cooling process in the temperature change interval as a temperature cycle. For example, the first heating stage and the subsequent cooling stage form a temperature cycle. For each temperature cycle, calculate its corresponding temperature amplitude, duration and other characteristics, and record this information. The label of each temperature cycle will include the amplitude of the temperature change and the position of the cycle in the temperature change (such as heating stage, cooling stage, etc.). All temperature cycle tag data are output and form a data set, which includes the type of each temperature cycle (heating, cooling, stable stage, etc.) and its related temperature characteristic data (such as temperature change amplitude, duration, number of occurrences, etc.).

[0150] Specifically, based on the temperature change rate and peak-to-valley matching relationship, the temperature cycle is divided into the following categories: rising cycle (significant temperature increase), falling cycle (significant temperature decrease), and stable stage (small temperature change). If the change amplitude of a cycle is small but the duration is long, it belongs to the progressive aging mode and needs to be marked separately. Assign the corresponding category mark to each temperature cycle, and record the temperature amplitude, duration, and number of occurrences of the cycle.

[0151] Preferably, the rain flow count extraction is specifically as follows:

[0152] Recursive interval merging is performed based on the temperature cycle marker data to obtain peak-valley pair processing data;

[0153] Specifically, the temperature change amplitude and its occurrence frequency are extracted from the temperature cycle marker data, and the temperature cycles in the data set are traversed in sequence to check whether two adjacent cycles can be merged. It is determined whether the adjacent cycles are too close. If the time difference between adjacent cycles is less than a certain time threshold (such as 10s), they are considered as part of the same cycle and further determined whether to merge. It is determined whether the temperature change amplitudes of the two cycles are close, and a merging threshold (such as 5%) is set. If it is less than the merging threshold, it means that the temperature changes of the two cycles are similar and belong to the same trend, and should be merged.

[0154] Specifically, in the temperature cycle marker data, all peaks and valleys are extracted and arranged in chronological order. If two adjacent temperature cycles have small amplitudes or short time intervals, they are merged into a larger cycle to reduce data redundancy.

[0155] Perform sequence matching analysis on the peak-to-valley pair processed data to obtain half-cycle cycle data;

[0156] Specifically, the temperature change pattern is based on the peak-valley pair as the basic structure, and the half-cycle is defined as a rise-fall pair: the temperature first rises to the peak value and then falls to the valley value; the fall-rise pair: the temperature first falls to the valley value and then rises to the peak value. Each half-cycle consists of a peak point and a valley point. Traverse the temperature data and find all the peak points and valley points. The peak point is the local maximum temperature point, that is, the temperature of the point is higher than its adjacent data points; the valley point is the local minimum temperature point, that is, the temperature of the point is lower than its adjacent data points. For adjacent peak points and valley points, check their time sequence. If the peak point appears before the valley point, it is matched into a rise-fall half-cycle. If the valley point appears before the peak point, it is matched into a fall-rise half-cycle. For each matched half-cycle, calculate the following parameters: the half-cycle duration is such as calculating the duration of the half-cycle, that is, the time interval between the peak point and the valley point. The temperature change amplitude, calculate the temperature change amplitude of the half-cycle, that is, the difference between the peak temperature and the valley temperature. If the temperature change trend is rising first and then falling, it is marked as a rise-fall half-cycle. If the temperature change trend is to first decrease and then increase, it is marked as a decrease-increase half cycle. Traverse all the peak and valley points in the entire temperature data and match them according to the above rules to obtain a list containing half-cycle cycle data.

[0157] Specifically, the temperature change curve is traversed in sequence, and adjacent peak and valley points are matched to determine whether they constitute a half-cycle cycle. The temperature change amplitude and duration of each half cycle are calculated and stored.

[0158] Merge the cycle intervals according to the half-cycle cycle data to obtain the merged temperature cycle data;

[0159] Specifically, all half-cycle data are traversed to find the up-down ascending sequence that can be merged. If two adjacent half-cycle cycles have similar duration and amplitude, they are merged into a complete temperature cycle.

[0160] Calculate the cycle amplitude based on the combined temperature cycle data to obtain the temperature cycle amplitude matrix data;

[0161] Specifically, the maximum temperature difference of the temperature cycle is calculated as the amplitude of the cycle.

[0162] Performing threshold adjustment on the temperature cycle amplitude matrix data to obtain temperature threshold data;

[0163] Specifically, a minimum threshold for temperature change is set, and cycles below this threshold will be ignored. The threshold is adjusted according to different operating conditions to ensure that only cycles with a greater impact on the semiconductor life are retained.

[0164] Perform effective cycle screening according to the temperature cycle amplitude matrix data and the temperature threshold data to obtain first effective temperature cycle data;

[0165] Specifically, for each cycle, its normalized weight is calculated, and if the weight is less than five percent, it is ignored.

[0166] Specifically, temperature cycles with large amplitudes are retained and cycles with low amplitudes are eliminated. High-frequency cycle patterns are calculated to determine the main cycles that affect the semiconductor lifetime.

[0167] Count the cycle intensity stratified rain flow according to the half-cycle cycle data to obtain the second effective temperature cycle data;

[0168] Specifically, according to the change amplitude of the temperature cycle (clustering calculation is performed, divided into two clusters, the cluster center is used as the standard for threshold judgment, the value greater than the larger value in the cluster center is a large cycle, the value between the larger value in the cluster center and the smaller value in the cluster center is a medium cycle, and the value less than the smaller value in the cluster center is a small cycle), it is divided into large cycles, medium cycles and small cycles to reflect the degree of its impact on the life of power semiconductors. The large cycle (main cycle) refers to a cycle with a large temperature change amplitude, which has a high stress effect and has the most significant impact on the fatigue life of the device. If the change amplitude of a certain temperature cycle is greater than the set high intensity threshold, it is considered to belong to the large cycle. The medium cycle refers to a cycle with a temperature change amplitude in a medium range, which has a secondary impact on the life of the device than the large cycle, but still causes a certain degree of fatigue damage. If the change amplitude of a certain temperature cycle is between the medium intensity threshold and the high intensity threshold, it is considered to belong to the medium cycle. The small cycle refers to a cycle with a small temperature change amplitude, which has a small impact on the life, but if the frequency is high, it still causes cumulative damage. If the change amplitude of a certain temperature cycle is less than the medium intensity threshold, it is considered to belong to the small cycle. Calculate the hierarchical weight of each temperature cycle, and for each temperature cycle, calculate its intensity ratio weight to measure the impact of the cycle relative to the overall temperature change. Calculate the relative intensity of each temperature cycle, that is, the normalized ratio of its change amplitude to the overall temperature cycle change range. This ratio can be used to measure the impact of the temperature cycle in the entire life assessment. After completing the intensity stratification, it is necessary to count the number of temperature cycles at each level in order to evaluate the contribution of different levels to the life of power semiconductors. At the large cycle level, calculate the number of occurrences of high-intensity temperature cycles to identify the cycle mode that causes rapid fatigue failure of the device. At the medium cycle level, calculate the number of medium-intensity temperature cycles and analyze their impact on long-term life. At the small cycle level, calculate the number of occurrences of small-amplitude temperature cycles to determine their cumulative damage effects.

[0169] Specifically, according to the temperature cycle amplitude, the cycles are divided into three levels: high intensity, medium intensity, and low intensity. The frequency of temperature cycles of different intensities is calculated to identify the cycle with the greatest impact on fatigue life.

[0170] The first effective temperature cycle data and the second effective temperature cycle data are sorted by correlation to obtain effective temperature cycle data.

[0171] Specifically, the characteristics of each temperature cycle are determined by its temperature change amplitude. A larger amplitude means higher thermal stress and has a more significant impact on the semiconductor life. By comparing the amplitude differences of different temperature cycles, we can measure whether two temperature cycles are similar. If the amplitudes of two temperature cycles are slightly different, they are considered to have a high correlation and belong to the same thermal load mode. By normalizing the amplitudes of all temperature cycles, we ensure that there is no deviation in the correlation calculation due to different data scales. After calculating the correlation of each temperature cycle, we sort them according to their influence to screen out the most representative effective cycles. We mainly sort them according to the temperature change amplitude and frequency of occurrence: the larger the change amplitude, the higher the sorting priority, and the larger amplitude corresponds to more severe temperature stress. The higher the frequency, the higher the sorting priority, because frequently occurring temperature cycles cause greater cumulative damage to the device. First, sort them in descending order according to the temperature change amplitude, so that the most drastic temperature change is in the front. In the case of the same amplitude, sort them in descending order according to the number of cycles, so that the most frequently occurring cycles are retained first.

[0172] Specifically, the matching degree between the first effective temperature cycle and the second effective temperature cycle is calculated to find similar cycle patterns. The cycles are sorted according to the frequency of occurrence and the intensity of influence of the cycles, and the temperature cycle with the greatest influence is analyzed first.

[0173] Preferably, the circulation intensity layered rain flow count is specifically:

[0174] Calculate the cycle intensity based on the half-cycle cycle data to obtain the temperature cycle intensity data;

[0175] Specifically, the cycle intensity is mainly determined by the temperature amplitude, time scale and material properties: ,in For the The intensity of a temperature cycle describes the impact of the cycle on the power semiconductor. is a material constant, determined experimentally, For the The temperature amplitude of a temperature cycle. That is, the maximum temperature change in the cycle, which reflects the magnitude of temperature fluctuation. is the exponential coefficient of the temperature amplitude, which describes the relationship between the temperature amplitude and the cycle intensity. It is obtained by fitting the experimental data and indicates the degree of influence of temperature change on the cycle intensity. For the The duration (time scale) of a temperature cycle. That is, the time interval between temperature fluctuations, which reflects the length of time the temperature change lasts. It is the exponential coefficient of the time scale, which expresses the influence of the duration (time scale) of temperature change on the cycle intensity, and is usually obtained by fitting experimental data.

[0176] Specifically, the amplitude, duration, and temperature change rate of each temperature cycle are extracted from the previous half-cycle data set. Based on the temperature amplitude and cycle duration, the thermal shock intensity of each cycle is calculated, and combined with the material properties of the power semiconductor, the contribution of each cycle to fatigue damage is determined.

[0177] Perform hierarchical clustering according to the temperature cycle intensity data to obtain the temperature cycle intensity stratified data;

[0178] Specifically, the temperature amplitude characteristics and the temperature change rate characteristics are extracted according to the temperature cycle intensity data. The initial cluster center is set according to the temperature amplitude characteristics and the temperature change rate characteristics. For example, three types of clusters (high intensity, medium intensity, and low intensity) are selected, and three points are randomly selected from the data as the initial cluster centers. For each temperature cycle data, its distance from the three cluster centers is calculated, that is, the distance between each data point and the three cluster centers is calculated, and the cluster center with the smallest distance is found, and the data point is assigned to the category. Each data point is assigned to the nearest cluster center according to the calculated distance, thereby forming a preliminary category. According to the currently assigned category, the cluster center of each category is recalculated. The new cluster center is the mean of all data points in the category. For each category, the new cluster center is the characteristic mean of all data points in the category. Compare the cluster centers before and after adjustment. If the change is small, it means that the clustering process has converged and the clustering is ended. If the cluster center changes greatly, return to step three to continue adjustment. During the clustering process, the cluster center will be gradually adjusted according to the distribution of the data to ensure that the data points within each cluster are as close as possible and the differences between different clusters are large. The maximum inter-class variance method is used to calculate the optimal classification boundary for each category. The inter-class variance formula is used to measure the differences between different categories and is defined as the sum of the weight of each category multiplied by its variance. Inter-class variance = the sum of the weight of each category multiplied by the category variance, where the weight represents the relative importance of each category and the variance reflects the degree of discreteness of the data within the category. A threshold that maximizes the inter-class variance is selected to distinguish the data into different categories. The thresholds include high intensity boundary value (maximum value), medium intensity boundary value, and low intensity boundary value. Based on the calculated classification boundaries, the temperature cycle data is divided into three categories. Each temperature cycle will be labeled as one of the following three categories according to its intensity: high intensity cycle (intensity greater than the high intensity boundary value), medium intensity cycle (intensity between medium and high intensity), and low intensity cycle (intensity less than the medium intensity boundary value).

[0179] Specifically, according to the fatigue characteristics of semiconductor materials, the temperature cycle intensity data is divided into three categories: high intensity, medium intensity, and low intensity. The temperature cycles are automatically stratified based on the cycle intensity, amplitude, duration and other characteristics, and similar cycles are classified into the same category. The frequency of temperature cycles in different categories is counted, and the main characteristics of each layer are recorded.

[0180] Perform key cycle screening based on temperature cycle intensity stratification data to obtain strength stratification key cycle data;

[0181] Specifically, the cumulative fatigue effect of temperature cycles refers to the contribution of different temperature cycle modes to the overall life of the semiconductor. The key cycles can be screened by calculating the cumulative fatigue contribution of all temperature cycles. Calculate the single fatigue contribution value of all temperature cycles, which reflects the degree of influence of the temperature cycle on the overall life. Calculate the total fatigue contribution value of all temperature cycles, that is, the sum of all single fatigue contribution values. Calculate the cumulative contribution percentage of each temperature cycle and sort them according to the contribution. Set the screening criteria for key cycles, retain the temperature cycles with the first 90% of the cumulative contribution, that is, the cycles with the largest contribution are retained first, and the cycles with low impact are screened out. After screening out the most influential cycles, further remove unnecessary low-frequency or low-contribution cycles to optimize data quality. If the contribution of a certain type of temperature cycle is less than 10% of the total fatigue contribution, it is considered that it has little impact on the overall life and can be merged into adjacent strength categories. If the number of occurrences of a certain type of temperature cycle is less than five times, it is considered that it has little impact on the long-term life and is merged into adjacent strength categories to reduce data redundancy.

[0182] Specifically, the contribution of each level of temperature cycle to the total fatigue damage is counted, and the cycle mode with the largest contribution is selected. If the contribution of a cycle to the total damage is lower than the set threshold, it is considered that the cycle has little impact on the life and is not included in the subsequent calculation. Only the temperature cycles with high impact are retained for subsequent life calculation.

[0183] According to the intensity stratified key cycle data, intensity stratified rain flow counting is performed to obtain the second effective temperature cycle number.

[0184] Specifically, the number of occurrences of high-intensity, medium-intensity, and low-intensity temperature cycles is calculated, and their distribution is analyzed. Based on the intensity of the temperature cycle, the contribution of each cycle level to the total life loss is calculated and weighted statistics are performed. The number of temperature cycles at each level and its damage impact are summarized to provide standardized input for life calculation.

[0185] Preferably, step S3 is specifically:

[0186] Step S31: performing temperature cycle distribution processing according to the effective temperature cycle data to obtain temperature cycle frequency distribution data;

[0187] Specifically, the data is partitioned and counted, that is, all temperature cycles are classified into different intervals according to the size of the temperature change amplitude. Set a minimum and maximum value of the amplitude change as the amplitude range of the overall temperature cycle. The amplitude range is divided into multiple fixed intervals, and the interval size of each interval is determined according to the actual application scenario, for example, five degrees Celsius is an interval to ensure the rationality of data statistics. Each temperature cycle data is classified into its amplitude interval for frequency statistics. For each divided amplitude interval, the cumulative number of occurrences in the interval is calculated, that is, all temperature cycle data are traversed, the interval to which its amplitude belongs is determined, and the number of occurrences is accumulated to the interval to obtain the total number of cycles in each amplitude interval, that is, the cumulative number of temperature cycles occurring within the temperature change amplitude range. In order to further measure the relative importance of each amplitude interval, it is necessary to calculate the relative frequency, that is, calculate the proportion of the cumulative number of occurrences of each amplitude interval in all temperature cycles, and obtain the relative contribution of different amplitude cycles in the overall temperature change pattern. After the previous calculation, the temperature cycle frequency distribution data can be output, including the amplitude interval of each temperature cycle. The number of cycles within this amplitude range. The relative frequency of this amplitude range, that is, its proportion in the overall temperature cycle data.

[0188] Specifically, based on the effective temperature cycle data extracted by rain flow counting in the previous step, the temperature change amplitude and the number of occurrences of each cycle are extracted. The amplitude range of all temperature cycles is counted, and they are classified according to different amplitude intervals to form a temperature cycle distribution table. The number of cycles in each temperature change amplitude interval is counted, and the cycle frequency distribution data is obtained after normalization.

[0189] Step S32: matching the temperature cycle frequency distribution data with a preset semiconductor lifetime model library to obtain a semiconductor lifetime matching model;

[0190] Specifically, the life models include the Coffin-Manson low-cycle fatigue model: the fatigue life of the semiconductor is calculated by the temperature change amplitude and the fatigue constant. It describes the effect of temperature change on the fatigue life of the semiconductor and assumes a power relationship between fatigue damage and the temperature change amplitude. The Arrhenius failure model, which takes into account temperature dependence, calculates the semiconductor life by the activation energy of temperature change and the ambient temperature. The Miner damage accumulation model: This model estimates the semiconductor life by accumulating damage. Each temperature cycle will cause a certain amount of damage, and the accumulation of damage determines the life of the semiconductor. The mean and standard deviation of the temperature cycle amplitude are calculated. The data is normalized according to the maximum and minimum values ​​of the temperature cycle amplitude and the difference in frequency distribution. The similarity between the normalized temperature data and each model in the life model library is calculated. The similarity calculation is based on the dot product of the vector, that is, the matching degree between the temperature cycle data and the life model data in the multidimensional space is calculated. According to the similarity calculation result, the life model most similar to the temperature cycle data is selected. By selecting the life model with the maximum similarity, the matching life model is obtained. By calculating the similarity between the normalized data and each life model, the life prediction model that best matches the current temperature cycle data is found. The life model with the greatest similarity is selected as the semiconductor life matching model.

[0191] Specifically, a variety of power semiconductor life models are pre-stored, which are established based on experimental data of different materials, packaging structures and working conditions. According to the temperature cycle frequency distribution data, multiple models in the life model library are matched to select the model that best matches the current temperature cycle characteristics. If there is a deviation between the matched life model and the actual operating conditions, the model parameters are adjusted to make it more suitable for the current application environment.

[0192] Step S33: performing single cycle lifetime calculation according to the semiconductor lifetime matching model to obtain single cycle lifetime evaluation data.

[0193] Specifically, if the Coffin-Manson model is selected, the single cycle life is calculated as the life of a single temperature cycle is inversely proportional to the power of the temperature change amplitude, and the formula contains a constant and an experimentally determined exponent. The corresponding temperature cycle life can be calculated based on the temperature change amplitude. If the Arrhenius failure model is selected, the life is determined by both temperature and activation energy. The activation energy is a material constant, and the temperature is determined by the average value of each temperature cycle. In this model, the effect of temperature on life is an exponential relationship, so the life of the cycle will increase or decrease exponentially with increasing temperature. If the Miner linear damage accumulation model is selected, the model compares the life contribution of each temperature cycle with the accumulated damage of other cycles. The damage contribution of each temperature cycle is calculated based on the proportional relationship between its life and the total life.

[0194] Specifically, the maximum number of cycles that the power semiconductor can withstand in each temperature cycle mode is calculated based on the life model. The contribution of each temperature cycle to the overall fatigue life is calculated to obtain single cycle life assessment data.

[0195] Preferably, step S4 is specifically:

[0196] Perform damage accumulation calculation on single cycle life assessment data to obtain semiconductor cumulative damage data;

[0197] Specifically, based on the single-cycle life assessment data calculated in the previous stage, the tolerable number of each temperature cycle and its occurrence in actual operation are extracted. By comparing the actual number of occurrences of each temperature cycle with the theoretical tolerable number, the damage contribution of the cycle to the overall life is calculated. Since there are many types of temperature cycles for semiconductors, each cycle mode has a different impact on the life, so it is necessary to accumulate the damage contribution of each cycle to evaluate the overall fatigue level.

[0198] The total cycle life of the power semiconductor is calculated based on the semiconductor cumulative damage data to obtain the total cycle life data of the power semiconductor.

[0199] Specifically, the total cycle life data of power semiconductors: , ,in The total cycle life of the power semiconductor (unit: number of cycles / time), which indicates the total number of cycles or working time of the power semiconductor in its entire life cycle. is the damage accumulation value (unit: dimensionless). This value represents the damage accumulation of the power semiconductor within a certain period of time, which is obtained through the accumulation of multiple cycles of damage. Reflects the durability and service life of power semiconductors, A_T is a temperature-related factor (unit: dimensionless), which affects the life of semiconductor materials through temperature and changes with temperature. A_T is calculated by the Arrhenius equation, which expresses the relationship between temperature and material failure. Activation energy (unit: electron volt, eV) is the energy barrier that a material needs to overcome when undergoing a physical or chemical change. It describes the effect of temperature on the life of power semiconductors and is obtained through experimental data. is the Boltzmann constant (unit: electron volt / Kelvin, eV / K), which is used to quantify the effect of thermal energy on particle systems. is the junction temperature (unit: Kelvin, K). The junction temperature is the temperature of the chip inside the semiconductor, which directly affects the thermal performance and life of the semiconductor. Acquired by sensors or calculated by thermal impedance models.

[0200] Specifically, based on the cumulative damage data, the current damage ratio of the semiconductor is calculated, and its state relative to the failure threshold is determined. Based on the cumulative damage of the power semiconductor, its remaining life under the current load conditions is predicted, that is, the number of temperature cycles that it is expected to withstand. If the cumulative damage is close to the failure threshold, it means that the semiconductor is about to reach the end of its life and needs to be replaced or maintained. If the cumulative damage is far below the failure threshold, it means that the semiconductor can continue to operate stably.

[0201] Preferably, the present application also provides a power semiconductor life prediction system based on the rain flow method, which is used to execute the power semiconductor life prediction method based on the rain flow method as described above. The power semiconductor life prediction system based on the rain flow method includes:

[0202] A semiconductor temperature change calculation module is used to collect semiconductor operation data and perform temperature change calculation based on the semiconductor operation data to obtain semiconductor temperature change data;

[0203] Rain flow counting analysis module, used to perform rain flow counting analysis based on semiconductor temperature change data to obtain effective temperature cycle data;

[0204] A single cycle life assessment module is used to perform a single cycle life assessment based on effective temperature cycle data to obtain single cycle life assessment data;

[0205] The cumulative life calculation module is used to perform cumulative life calculation on the single cycle life assessment data to obtain the total cycle life of the power semiconductor.

[0206] Therefore, from any point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present invention is limited by the attached application documents rather than the above description, and it is intended that all changes falling within the meaning and scope of equivalent elements of the application documents are included in the present invention.

[0207] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A power semiconductor life prediction method based on the rain flow method, characterized in that: The following steps are involved: Step S1: collecting semiconductor operation data, and performing temperature change calculation based on the semiconductor operation data to obtain semiconductor temperature change data; Step S2: Perform rain flow counting analysis based on the semiconductor temperature change data to obtain effective temperature cycle data; Step S3: performing a single cycle life assessment according to the effective temperature cycle data to obtain single cycle life assessment data; Step S4: performing cumulative life calculation on the single cycle life evaluation data to obtain total cycle life data of the power semiconductor; Step S2 is specifically as follows: Discretize the temperature waveform according to the semiconductor temperature change data to obtain discrete temperature waveform data; Perform threshold segmentation analysis on discrete temperature waveform data to obtain temperature cycle marker data; Extract rain flow counts based on temperature cycle marker data to obtain effective temperature cycle data; The specific extraction of rain flow count is as follows: Recursive interval merging is performed based on the temperature cycle marker data to obtain peak-valley pair processing data; Perform sequence matching analysis on the peak-to-valley pair processed data to obtain half-cycle cycle data; Merge the cycle intervals according to the half-cycle cycle data to obtain the merged temperature cycle data; Calculate the cycle amplitude based on the combined temperature cycle data to obtain the temperature cycle amplitude matrix data; Performing threshold adjustment on the temperature cycle amplitude matrix data to obtain temperature threshold data; Perform effective cycle screening according to the temperature cycle amplitude matrix data and the temperature threshold data to obtain first effective temperature cycle data; Count the cycle intensity stratified rain flow according to the half-cycle cycle data to obtain the second effective temperature cycle data; The first effective temperature cycle data and the second effective temperature cycle data are sorted by correlation to obtain effective temperature cycle data.

2. The method according to claim 1, characterized in that Step S1 is specifically as follows: Deploy low-power sensors at the semiconductor equipment end to collect semiconductor operation data, where the semiconductor operation data includes semiconductor operation current data, semiconductor operation voltage data, semiconductor operation temperature data, and semiconductor operation power loss data; Extracting features based on semiconductor operation data to obtain semiconductor operation feature data; Power loss modeling is performed based on semiconductor operation characteristic data to obtain semiconductor power loss curve data; Perform thermal impedance network calculation based on semiconductor power loss curve data to obtain equivalent thermal impedance data; The junction temperature change is calculated based on the semiconductor power loss curve data and the equivalent thermal impedance data to obtain the transient junction temperature curve data; Timing prediction and temperature change extraction are performed based on transient junction temperature curve data to obtain semiconductor temperature change data.

3. The method according to claim 1, characterized in that The temperature waveform is discretized as follows: Extract temperature variation characteristics according to semiconductor temperature variation data to obtain semiconductor temperature variation characteristic data; Sampling semiconductor temperature change data according to semiconductor temperature change characteristic data to obtain semiconductor temperature change sampling data; Perform multi-scale curvature analysis on the semiconductor temperature change sampling data to obtain the semiconductor temperature change trend matrix data; Perform short-time waveform segmentation on the semiconductor temperature change sampling data to obtain short-time temperature waveform segment data; Perform variational autoencoding on the semiconductor temperature change sampling data to obtain long-term temperature waveform characteristic data; According to the semiconductor temperature change trend matrix data, short-term temperature waveform segment data and long-term temperature waveform feature data are subjected to Bayesian correlation to obtain inter-segment correlation data; The waveform of the correlation data between the segments is discretized to obtain discrete temperature waveform data.

4. The method according to claim 1, characterized in that: The specific analysis of threshold segmentation is as follows: Perform Bayesian optimal segmentation on discrete temperature waveform data to obtain a preliminary segmentation data set; Identify the temperature extreme points of the preliminary segmented data set to obtain segmented temperature extreme point data; Perform local maximum-minimum contrast analysis based on style temperature extreme point data to obtain peak-valley matching data; Perform segmented linear regression on the preliminary segmented data set based on the peak-valley matching data to obtain segmented temperature change interval data; The temperature cycle marking data is obtained by classifying and marking the divided temperature change interval data.

5. The method according to claim 1, characterized in that The specific cycle intensity layered rain flow count is: Calculate the cycle intensity based on the half-cycle cycle data to obtain the temperature cycle intensity data; Perform hierarchical clustering according to the temperature cycle intensity data to obtain the temperature cycle intensity stratified data; Perform key cycle screening based on temperature cycle intensity stratification data to obtain strength stratification key cycle data; According to the intensity stratified key cycle data, intensity stratified rain flow counting is performed to obtain the second effective temperature cycle number.

6. The method according to claim 1, characterized in that Step S3 is specifically as follows: Perform temperature cycle distribution processing according to effective temperature cycle data to obtain temperature cycle frequency distribution data; Matching the temperature cycle frequency distribution data with a preset semiconductor lifetime model library to obtain a semiconductor lifetime matching model; The single cycle lifetime calculation is performed according to the semiconductor lifetime matching model to obtain the single cycle lifetime evaluation data.

7. The method according to claim 1, characterized in that Step S4 is specifically as follows: Perform damage accumulation calculation on single cycle life assessment data to obtain semiconductor cumulative damage data; The total cycle life of the power semiconductor is calculated based on the semiconductor cumulative damage data to obtain the total cycle life data of the power semiconductor.

8. A power semiconductor life prediction system based on the rain flow method, characterized in that: For executing the power semiconductor life prediction method based on the rain flow method as claimed in claim 1, the power semiconductor life prediction system based on the rain flow method comprises: A semiconductor temperature change calculation module is used to collect semiconductor operation data and perform temperature change calculation based on the semiconductor operation data to obtain semiconductor temperature change data; Rain flow counting analysis module, used to perform rain flow counting analysis based on semiconductor temperature change data to obtain effective temperature cycle data; A single cycle life assessment module is used to perform a single cycle life assessment based on effective temperature cycle data to obtain single cycle life assessment data; The cumulative life calculation module is used to perform cumulative life calculation on the single cycle life assessment data to obtain the total cycle life of the power semiconductor.

Images