Power device life assessment method based on multi-operating condition state clustering

Abstract

This invention relates to the field of life assessment technology, and more particularly to a power device life assessment method based on multi-condition operating state clustering. The method includes the following steps: collecting multi-dimensional raw operating condition data of the power device during its continuous operation throughout its entire life cycle, and constructing a multi-dimensional operating condition feature dataset; performing condition state differentiation mining on the multi-dimensional operating condition feature dataset, and constructing an adaptive density clustering model to complete clustering, generating various operating condition clusters; generating a dynamic feature sequence of device losses based on the characteristic distribution patterns of various operating condition clusters; coupling the device's fatigue damage accumulation mechanism based on the dynamic feature sequence of device losses, and completing real-time fatigue damage quantification calculation of the power device to generate dynamic damage accumulation data; and combining the dynamic damage accumulation data to infer the remaining service life of the power device, thereby completing the full-cycle life assessment of the power device. This invention can significantly align with the actual service damage patterns of power devices.

Description

Technical Field

[0001] This invention relates to the field of life assessment technology, and in particular to a method for assessing the life of power devices based on clustering of multiple operating conditions. Background Technology

[0002] Power devices, as core components in power electronic equipment, new energy converters, industrial control devices, and energy storage, directly determine the service life and operational safety of the entire equipment through their operational reliability and remaining service life. In practical engineering applications, power devices operate under complex and ever-changing dynamic conditions, frequently experiencing different working states such as steady-state operation, instantaneous overload, voltage fluctuations, load abrupt changes, and temperature transients. Under different operating conditions, the internal loss mechanisms, thermal stress distribution, and fatigue accumulation rates of the devices vary significantly, and fatigue damage exhibits strong condition-dependent and spatiotemporal cumulative characteristics. Therefore, accurately identifying the operating conditions of power devices throughout their entire lifecycle and quantifying the fatigue damage evolution patterns under multiple coupled operating conditions are core prerequisites for achieving high-precision life assessment of power devices, predicting failure risks in advance, and ensuring stable equipment operation.

[0003] Currently, most traditional power device life assessment methods in the industry employ fixed operating condition parameter modeling, empirical formula derivation, or single feature data fitting. These methods typically select electrical and temperature parameters from rated operating conditions and typical steady-state operating conditions to establish damage models, neglecting the dynamic differences in operating conditions, the coupling and correlation of multiple parameters, and the sudden fatigue damage caused by transient and fluctuating conditions during actual power device operation. In practical applications, power devices frequently experience complex operating conditions that are not rated or steady-state. The loss evolution mechanisms of steady-state and transient conditions are completely different. Traditional unified models cannot accurately distinguish the damage contribution of various operating conditions, easily leading to inaccurate damage quantification and large deviations in lifespan estimation, making it difficult to meet the dynamic life assessment requirements of the entire device lifecycle. Summary of the Invention

[0004] Therefore, it is necessary for the present invention to provide a power device lifetime assessment method based on multi-operating-condition clustering to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, a power device lifetime assessment method based on multi-condition operating state clustering includes the following steps: Step S1: Collect multi-dimensional raw operating condition data of the power device during continuous operation throughout its entire life cycle, filter core operating condition characteristic parameters based on the device operating mechanism, and construct a multi-dimensional operating condition characteristic dataset; Step S2: Perform differential mining of operating conditions on the multi-dimensional operating condition feature dataset, extract the dynamic evolution correlation features of each feature parameter, and generate a multi-dimensional operating condition correlation feature set corresponding to the power device. Step S3: Construct an adaptive density clustering model based on the multi-dimensional operating condition association feature set, complete the clustering of the power device's full operating condition status, and generate various operating condition clusters, including steady-state operating condition clusters and transient operating condition clusters. Step S4: Based on the characteristic distribution patterns of various operating condition clusters, match the internal loss evolution mechanism of power devices under different operating conditions, and generate the device loss dynamic characteristic sequence corresponding to each operating condition cluster. Step S5: Based on the dynamic characteristic sequence of device loss of each operating condition cluster coupled with the device fatigue damage accumulation mechanism, construct a device fatigue damage evolution model coupled with multiple operating conditions, and complete the real-time fatigue damage quantification calculation of power devices based on the device fatigue damage evolution model to generate device dynamic damage accumulation data. Step S6: Combining the device dynamic damage accumulation data with the runtime proportion of various operating condition clusters, iteratively correct the operating condition coupling coefficient of the device fatigue damage evolution model, and extrapolate the remaining service life of the power device based on the corrected device fatigue damage evolution model, thereby completing the full life cycle assessment of the power device.

[0006] The beneficial effects of this invention are: The power device lifetime assessment method based on multi-condition operating state clustering proposed in this invention has the following advantages compared with the prior art: It collects multi-dimensional raw operating condition data of the power device throughout its entire life cycle, combines the device's operating mechanism to screen core operating condition characteristic parameters, and constructs a multi-dimensional operating condition characteristic dataset. This step focuses on the entire life cycle operation of the device, comprehensively collects multi-dimensional raw operating condition data, and completes feature screening based on the device's physical operating mechanism. Invalid and redundant data is eliminated, and core parameters reflecting loss changes and fatigue accumulation are retained, constructing a comprehensive and targeted operating condition characteristic dataset. This provides a real and complete data foundation for subsequent operating condition difference mining, feature correlation analysis, and damage modeling. By conducting operating condition difference mining on the multi-dimensional operating condition characteristic dataset, the dynamic evolution correlation features of each parameter are extracted, and a multi-dimensional operating condition correlation feature set of the power device is constructed. This step deeply explores the dynamic correlation relationship of different operating condition parameters over time, distinguishes the characteristic differences between steady-state operating conditions and dynamic fluctuation operating conditions, captures the parameter evolution law during operating condition switching, and integrates them to form a multi-dimensional correlation feature set, which can fully reflect the linkage change characteristics of the device's operating state under complex and variable operating conditions. An adaptive density clustering model is built based on a multi-dimensional set of associated features to complete the clustering of all operating conditions, distinguishing between steady-state and transient operating condition clusters. This step utilizes an adaptive clustering algorithm to autonomously classify massive dynamic operating condition features, dividing different operating condition clusters based on differences in feature distribution. This effectively distinguishes between steady-state routine operating conditions and transient abrupt changes, clarifying the distribution characteristics and operating patterns of various operating conditions, laying a classification foundation for subsequent targeted matching of loss evolution mechanisms and quantification of fatigue damage under different operating conditions. Secondly, based on the feature distribution patterns of various operating condition clusters, differentiated internal loss evolution mechanisms of power devices are matched, generating dynamic feature sequences of device losses for the corresponding operating condition clusters. This step matches the internal loss generation mechanism of each segmented operating condition cluster, combining the dynamic evolution characteristics of the operating conditions to generate time-series loss feature sequences, fully reconstructing the dynamic accumulation process of device losses under various operating conditions throughout the entire lifecycle, achieving differentiated characterization of loss mechanisms under different operating conditions, and providing accurate loss data support for fatigue damage quantification. Then, by combining the dynamic characteristic sequence of multi-condition loss with the fatigue damage accumulation mechanism of the device, a multi-condition coupled fatigue damage evolution model is constructed to complete the real-time fatigue damage quantification calculation of the device and generate dynamic damage accumulation data. This step deeply couples the loss characteristics of different operating conditions with the physical fatigue accumulation mechanism of the device, establishes a dynamic evolution model that can be adapted to all operating conditions, and can simultaneously characterize the accumulation process of steady-state slow damage and transient sudden damage, quantify the fatigue damage increment of the device under complex alternating operating conditions in real time, and greatly conform to the actual service damage law of power devices.Finally, by combining dynamic damage accumulation data with the operating time ratio of each operating condition, the model operating condition coupling coefficient is iteratively corrected. Based on the corrected evolution model, the remaining service life is extrapolated, and the full-cycle life assessment is completed. This step relies on real-time operating condition data and damage data to continuously optimize the model coupling coefficient, realize the dynamic adaptive update of model parameters, conform to the evolution law of the device's full life cycle operating conditions, weaken the extrapolation error caused by the fixed model, restore the damage accumulation trend under complex operating conditions, and thus realize a dynamic and realistic assessment of the remaining service life of power devices. Attached Figure Description

[0007] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the steps in the power device lifetime assessment method based on multi-condition operating state clustering of the present invention. Figure 2 for Figure 1 A detailed flowchart of step S1. Detailed Implementation

[0008] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0009] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0010] 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 merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0011] To achieve the above objectives, please refer to Figure 1 This invention provides a power device lifetime assessment method based on multi-condition operating state clustering, comprising the following steps: Step S1: Collect multi-dimensional raw operating condition data of the power device during continuous operation throughout its entire life cycle, filter core operating condition characteristic parameters based on the device operating mechanism, and construct a multi-dimensional operating condition characteristic dataset; In this embodiment, during the continuous operation of the power device throughout its entire lifecycle, multi-dimensional raw operating condition data is acquired through four parallel data acquisition links. The first link acquires instantaneous voltage and current values ​​at the input and output terminals of the power device at a rate of 10,000 sampling points per second, while simultaneously recording derived electrical parameters such as power factor and reactive power. The second link acquires data on the power device's casing temperature, estimated junction temperature, and heat sink surface temperature distribution at a rate of 100 sampling points per second. The third link acquires ambient temperature, ambient humidity, and atmospheric pressure data at a rate of one sampling point per second. The fourth link acquires time-domain waveform data of vibration acceleration in three axes at the power device's mounting base and terminal connections at a rate of 5,000 sampling points per second. The four data acquisition links are synchronized via a GPS timing module to ensure strict alignment of all data in the time dimension. After continuously collecting all operating data of power devices from commissioning to the current moment, the validity of features is evaluated based on the semiconductor conduction loss, switching loss, and thermal fatigue failure mechanism of the power devices. Redundant data dimensions uncoupled from device lifespan degradation are eliminated, retaining ten core operating condition feature parameters: RMS current value, on-state resistance value, maximum junction temperature, minimum junction temperature, switching frequency, voltage rise rate, current fall rate, case temperature, heat sink temperature, and RMS vibration acceleration. Min-max normalization mapping is performed on the ten retained core feature parameters to eliminate dimensional differences between different feature dimensions, generating a set of ten rows of equal-dimensional time-series feature sequences multiplied by the total number of time segments. This set of sequences constitutes the multidimensional operating condition feature dataset.

[0012] Step S2: Perform differential mining of operating conditions on the multi-dimensional operating condition feature dataset, extract the dynamic evolution correlation features of each feature parameter, and generate a multi-dimensional operating condition correlation feature set corresponding to the power device. In this embodiment, based on a multi-dimensional operating condition feature dataset, a time-series decomposition operation is performed on each operating condition feature parameter. A local weighted regression smoothing algorithm is applied to the complete time-series sequence of the junction temperature feature parameter, setting the smoothing window width to the number of sampling points corresponding to one hour, resulting in the long-term evolution trend component sequence and the short-term dynamic fluctuation component sequence of the junction temperature. The same decomposition operation is repeated for the other nine feature parameters to generate their respective long-term trend components and short-term fluctuation components. Based on the long-term trend component, the Pearson correlation coefficient between any two feature parameters is calculated, constructing a 10-row by 10-column static coupling correlation matrix. Based on the short-term fluctuation component, a Granger causality test is performed, constructing a 10-row by 10-column dynamic linkage matrix. The static coupling correlation matrix and the dynamic linkage matrix are fused by taking the maximum value of each element to generate a bidirectional correlation matrix for the operating condition features. Based on this matrix, the entire operating period is segmented, dividing the day into four operating periods. Within each period, a spectral clustering algorithm is executed to identify the core correlation feature combinations. For each core associated feature combination, state representation attributes and temporal evolution attributes are extracted. The state representation attributes include the mean vector and covariance matrix, while the temporal evolution attributes include the rate of change vector and the direction of change vector. The attributes of all core associated feature combinations are merged to generate a multi-dimensional operating condition associated feature set of power devices covering the entire operating period.

[0013] Step S3: Construct an adaptive density clustering model based on the multi-dimensional operating condition association feature set, complete the clustering of the power device's full operating condition status, and generate various operating condition clusters, including steady-state operating condition clusters and transient operating condition clusters. In this embodiment, based on a multi-dimensional operating condition associated feature set, a three-dimensional Gaussian distribution function is constructed for each differentiated operating condition associated feature unit. The probability density value in the feature space is calculated, and the temporal clustering characteristics of each unit are extracted to construct an operating condition feature space distribution model. Based on this model, thresholds for high-density clustered regions and low-density discrete regions are set. Grid points with probability density values ​​higher than the high-density threshold are marked as cluster core regions, and grid points with probability density values ​​lower than the low-density threshold are marked as operating condition discrete regions. A merging operation is performed on data points within the cluster core regions, grouping core regions with similar mean vectors and covariance matrices into the same steady-state operating condition cluster. A differentiated splitting operation is performed on data points within the operating condition discrete regions, treating each discrete region as a separate transient operating condition unit. Local outlier factor calculation is performed on data points in both the steady-state operating condition cluster and the transient operating condition unit. Data points with local outlier factors exceeding the threshold are marked as isolated abnormal operating condition points and removed. After completing the clustering of all operating conditions, several steady-state operating condition clusters and several transient operating condition clusters are generated. Each cluster has a clear feature space boundary and a representative operating condition description. The steady-state operating condition clusters correspond to the operating conditions of power devices under stable load conditions, while the transient operating condition clusters correspond to the operating conditions under abnormal conditions such as sudden load increases or sudden temperature changes.

[0014] Step S4: Based on the characteristic distribution patterns of various operating condition clusters, match the internal loss evolution mechanism of power devices under different operating conditions, and generate the device loss dynamic characteristic sequence corresponding to each operating condition cluster. In this embodiment, based on various operating condition clusters, the mean values ​​of six core operating characteristic parameters—effective current, on-state resistance, maximum junction temperature, switching frequency, voltage rise rate, and current fall rate—are extracted for each operating condition cluster to construct a six-dimensional operating condition cluster feature representation vector. Based on the semiconductor carrier transport characteristics of power devices, on-state resistance and effective current are determined as the core influencing factors of conduction loss, switching frequency, voltage rise rate, and current fall rate are determined as the core influencing factors of switching loss, and maximum junction temperature is determined as a dual influencing factor. A mapping relationship is established between the operating condition cluster feature representation vector and the core influencing factors of loss, mapping each element in the six-dimensional vector to the corresponding physical variable. Based on this mapping relationship, a nonlinear correlation curve between the operating condition characteristics and conduction and switching losses is fitted. The fitting process uses a radial basis function network model, and the model is trained and temperature drift corrected using measured loss data. Based on the corrected nonlinear correlation curve, the real-time power loss and loss accumulation rate of the device under various operating condition clusters are calculated. Based on the time-series operating characteristics of each cluster, local weighted regression smoothing is performed on the time-series power loss sequence of each cluster to generate the time-series dynamic evolution law of the loss parameters. Three dynamic feature parameters—mean power loss, power loss fluctuation amplitude, and power loss change rate—are extracted from the evolution law and concatenated into a three-dimensional vector as the device loss dynamic feature sequence corresponding to each cluster.

[0015] Step S5: Based on the dynamic characteristic sequence of device loss of each operating condition cluster coupled with the device fatigue damage accumulation mechanism, construct a device fatigue damage evolution model coupled with multiple operating conditions, and complete the real-time fatigue damage quantification calculation of power devices based on the device fatigue damage evolution model to generate device dynamic damage accumulation data. In this embodiment, a multi-condition coupled device fatigue damage evolution model is constructed based on the device loss dynamic characteristic sequence of each operating condition cluster and coupled with the fatigue damage accumulation mechanism of power devices. The fatigue damage accumulation mechanism is based on the linear cumulative damage theory, which sets that the fatigue life of a power device under a single operating condition satisfies an inverse power law relationship with the power loss, i.e., the greater the power loss, the shorter the fatigue life. For each operating condition cluster, the mean power loss in its loss dynamic characteristic sequence is substituted into the inverse power law relationship to calculate the fatigue damage amount of a single operation corresponding to that operating condition cluster. The fatigue damage amount of a single operation of each operating condition cluster is multiplied by the proportion of the operating time of that operating condition cluster to obtain the contribution of each operating condition cluster to the total fatigue damage. The contributions of all operating condition clusters are summed to obtain the total fatigue damage increment in the current operating period. The total fatigue damage increment is added to the historical fatigue damage accumulation to update the device dynamic damage accumulation data. This device fatigue damage evolution model takes the loss dynamic characteristic sequence and the proportion of the operating time of the operating condition cluster as input and the device dynamic damage accumulation data as output, and quantifies the degree of fatigue damage accumulation of power devices in real time throughout their entire life cycle. Each time new operating data arrives, the dynamic characteristics of loss and the percentage of operating time for each cluster under different operating conditions are recalculated, and the cumulative dynamic damage data of the devices is iteratively updated.

[0016] Step S6: Combining the device dynamic damage accumulation data with the runtime proportion of various operating condition clusters, iteratively correct the operating condition coupling coefficient of the device fatigue damage evolution model, and extrapolate the remaining service life of the power device based on the corrected device fatigue damage evolution model, thereby completing the full life cycle assessment of the power device.

[0017] In this embodiment, the operating condition coupling coefficient of the device fatigue damage evolution model is iteratively corrected based on the device's dynamic damage accumulation data and the runtime proportion of various operating condition clusters. The operating condition coupling coefficient is used to adjust the influence of the interaction between different operating condition clusters on fatigue damage, and its initial value is set to 1. The device's dynamic damage accumulation data at the current moment is compared with the failure threshold specified in the power device's technical manual, and the difference between the current damage level and the failure threshold is calculated. The runtime proportion of each operating condition cluster and the corresponding loss dynamic characteristic sequence are input into the device fatigue damage evolution model to calculate the theoretical damage accumulation. The theoretical damage accumulation is compared with the actual device dynamic damage accumulation data to calculate the deviation between the two. Based on this deviation, a gradient descent correction is performed on the operating condition coupling coefficient, with a correction step size set to 0.01, so that the corrected theoretical damage accumulation approaches the actual damage accumulation. Based on the corrected device fatigue damage evolution model, the current device dynamic damage accumulation data is substituted into the model to deduce the remaining running time required to reach the failure threshold. The remaining operating time is calculated as follows: subtract the current accumulated damage from the failure threshold to obtain the remaining tolerable damage; divide the remaining tolerable damage by the average damage accumulation rate within the current period to obtain the estimated remaining service life. As operating data accumulates, the operating condition coupling coefficient and the estimated remaining service life are iteratively corrected to complete the full life cycle assessment of the power device.

[0018] Furthermore, as an embodiment of the present invention, reference is made to... Figure 2 As shown, Figure 1 A detailed flowchart of step S1 is shown below. In this embodiment, step S1 includes the following steps: Step S11: Synchronously collect electrical operation data, thermal operation data, environmental operation data and mechanical vibration operation data during the operation of power devices, and integrate them to generate a set of multi-source raw operation data corresponding to the power devices; In this embodiment, four parallel data acquisition links are used to acquire the operating status data of the power device at different time sections. The first link is responsible for acquiring electrical operating data, obtaining instantaneous voltage and current values ​​at the input and output terminals of the power device through voltage and current transformers, with an acquisition frequency of 10,000 sampling points per second, while also recording derived electrical parameters such as power factor and reactive power. The second link is responsible for acquiring thermal operating data, arranging thermocouple temperature measurement points on the casing and heat sink surface of the power device to acquire casing temperature, estimated junction temperature, and heat sink surface temperature distribution data, with an acquisition frequency of 100 sampling points per second. The third link is responsible for acquiring environmental operating data, obtaining ambient temperature, ambient humidity, and atmospheric pressure data through a thermometer and hygrometer installed inside the power device cabinet, with an acquisition frequency of 1 sampling point per second. The fourth link is responsible for acquiring mechanical vibration operating data, installing accelerometers at the mounting base and terminal connections of the power device to acquire time-domain waveform data of vibration acceleration in three axes, with an acquisition frequency of 5,000 sampling points per second. The clock sources of the four data acquisition links are synchronized via a GPS timing module to ensure that all data are aligned in the time dimension. Electrical data, thermal state data, environmental data, and vibration data collected at the same timestamp are horizontally stitched together to form a record containing multi-dimensional fields. After 24 hours of continuous collection, all records are vertically stacked in chronological order to generate a multi-source raw data set containing 86,400 records, each with hundreds of fields.

[0019] Step S12: Based on the semiconductor conduction loss, switching loss and thermal fatigue failure mechanism of power devices, the feature validity of the original multi-source operation data set is identified, and redundant data dimensions that are not coupled with device lifetime degradation are eliminated. In this embodiment, feature selection rules are established based on the semiconductor conduction loss mechanism of power devices. Conduction loss is proportional to the square of the on-state current and linearly related to the on-state resistance. Therefore, the effective current value and on-state resistance value are retained as effective feature dimensions, while harmonic content and phase angle difference data, which are not directly related to conduction loss, are eliminated. A second set of feature selection rules is established based on the switching loss mechanism of power devices. Switching loss is closely related to switching frequency, voltage change rate, and current change rate. Therefore, switching frequency, voltage rise rate, and current fall rate are retained as effective feature dimensions, while DC bias and ripple coefficient data, which are unrelated to the switching process, are eliminated. A third set of feature selection rules is established based on the thermal fatigue failure mechanism of power devices. Thermal fatigue failure is directly related to the junction temperature fluctuation amplitude and the number of cycles. Therefore, the maximum junction temperature, minimum junction temperature, and junction temperature change rate are retained as effective feature dimensions, while ambient light intensity and electromagnetic field intensity data, which are unrelated to thermal fatigue, are eliminated. The effective feature dimensions generated by the three rules are merged and deduplicated to obtain a feature list containing twenty-six effective dimensions. Each record in the original dataset from multiple sources is traversed, retaining only the fields listed in the feature list and discarding the rest of the redundant fields. After feature validity verification, the number of fields in the original dataset is reduced from hundreds to twenty-six, significantly reducing the data dimensionality while retaining core information that is physically coupled with device lifespan degradation.

[0020] Step S13: Extract the real-time operating features, dynamic fluctuation features, and state change features from the remaining valid data dimensions to form the initial operating condition feature parameter set corresponding to the power device; In this embodiment, three types of operating condition features are extracted from the obtained simplified data set. The first type is real-time operation features, which directly read the instantaneous values ​​of five fields—effective current, on-state resistance, switching frequency, maximum junction temperature, and minimum junction temperature—at each time segment to form a real-time operation feature vector. The second type is dynamic fluctuation features, which perform sliding window statistical analysis on the historical sequences of the effective current and maximum junction temperature fields, with a window length set to 60 seconds, calculate the standard deviation and peak-to-peak value of the data within the window, and add the calculation results as dynamic fluctuation features to the feature parameter group. The third type is state change features, which perform difference operations on the continuous sequences of the junction temperature change rate and voltage rise rate fields, calculate the absolute value of the difference between two adjacent sampling points, and when the absolute value of the difference exceeds a preset threshold, record the time and magnitude of the change, and add the frequency and average magnitude of the change event as state change features to the feature parameter group. The extracted real-time operation features, dynamic fluctuation features, and state change features are arranged in chronological order, forming an initial operating condition feature parameter group containing thirty-one feature dimensions at each time segment. This parameter set comprehensively describes the operating status, dynamic behavior characteristics, and abnormal event patterns of power devices under current operating conditions.

[0021] Step S14: Based on the rated operating condition boundary conditions of the power device, perform boundary constraint calibration on the initial operating condition characteristic parameter set to eliminate the characteristic offset problem caused by the rated operating condition deviation. In this embodiment, the rated operating condition boundary conditions specified in the power device's manufacturer's technical manual are read, including the rated voltage, rated current, rated junction temperature upper limit, and rated switching frequency. The voltage characteristic value in the initial operating condition characteristic parameter group generated in step S13 is divided by the rated voltage value to obtain the voltage normalization coefficient. The current characteristic value is divided by the rated current value to obtain the current normalization coefficient. The junction temperature characteristic value is divided by the rated junction temperature upper limit to obtain the junction temperature normalization coefficient. The switching frequency characteristic value is divided by the rated switching frequency value to obtain the frequency normalization coefficient. These four normalization coefficients all fluctuate within the range of 0 to 1.5, eliminating the dimensional inconsistency problem caused by differences in power device models and rated operating conditions. Further, operating condition boundary constraint calibration is performed on the normalized characteristic parameters. For characteristic values ​​exceeding the rated operating condition upper limit, they are truncated to the upper limit; for characteristic values ​​below the rated operating condition lower limit, they are raised to the lower limit. Simultaneously, the deviation between each characteristic value and the rated operating condition reference value is calculated, and this deviation is added to the characteristic parameter group as an independent calibration characteristic dimension. After calibration by operating condition boundary constraints, each characteristic value in the characteristic parameter group is within a unified dimensional system and standard operating condition range, eliminating the characteristic offset problem caused by rated operating condition deviation.

[0022] Step S15: The calibrated multi-dimensional working condition feature parameters are structurally integrated to construct a time-aligned, dimensionally unified multi-dimensional working condition feature dataset.

[0023] In this embodiment, the calibrated multi-dimensional operating condition feature parameters are structurally integrated to create a two-dimensional table structure. The number of rows in the table equals the total number of data acquisition time points, i.e., 86,400 rows, and the number of columns equals the total number of calibrated feature dimensions, i.e., thirty-three feature dimensions plus four deviation dimensions, totaling thirty-seven columns. The table is filled row by row in chronological order, with each row corresponding to all calibrated feature parameters at a given time segment. Missing value checks are performed on each column of the table. For missing data due to acquisition failures, the mean of the previous and next time points is used to fill in the missing data. Outlier detection is performed on each column of the table, using the interquartile range (IIR) method to identify outliers, and replacing outliers with the median of that column. The filled and replaced two-dimensional table is saved as a time-aligned, dimensionally unified multi-dimensional operating condition feature dataset. Each row in this dataset represents a complete description of the power device's operating condition at a certain time, and each column represents an operating condition feature dimension with clear physical meaning. The rows are strictly arranged in chronological order, providing a structurally sound and reliable data foundation for subsequent operating condition clustering analysis and life assessment modeling.

[0024] Furthermore, step S2 includes the following steps: Step S21: Based on the temporal distribution characteristics of the multidimensional working condition feature dataset, decompose the long-term evolution trend component and short-term dynamic fluctuation component of each working condition feature parameter. In this embodiment, based on the generated multidimensional operating condition feature dataset, a time-series decomposition operation is performed on each operating condition feature parameter. Taking the junction temperature feature parameter as an example, the complete time-series sequence of this parameter across 86,400 time segments is obtained. A local weighted regression smoothing algorithm is applied to this time-series sequence, with a smoothing window width of 3,600 time points, i.e., a one-hour time span. At each time point, the algorithm selects a subset of data from 1,800 time points before and after that point, and assigns distance weights to the data points in the subset, with greater weights closer to the center point and smaller weights farther away. A second-order polynomial is fitted based on the weighted data subset, and the value of this polynomial at the center point is used as the estimate of the long-term evolution trend component. This operation is repeated for all time points to obtain the long-term evolution trend component sequence of the junction temperature feature parameter. Subtracting the long-term evolution trend component sequence from the original junction temperature time-series sequence yields the short-term dynamic fluctuation component sequence. The same local weighted regression smoothing operation was repeated for the other thirty-six operating condition characteristic parameters to generate their respective long-term evolution trend component sequences and short-term dynamic fluctuation component sequences. All long-term evolution trend component sequences constitute a long-term trend matrix with 37 rows and 86,400 columns, and all short-term dynamic fluctuation component sequences constitute a short-term fluctuation matrix with 37 rows and 86,400 columns. These two matrices respectively characterize the slow drift pattern of power device operating condition characteristics over a long time scale and the rapid oscillation characteristics over a short time scale.

[0025] Step S22: Explore the static coupling correlation between different long-term evolution trend components and the dynamic linkage pattern between short-term dynamic fluctuation components to generate a two-way correlation matrix of working condition characteristics. In this embodiment, based on the long-term trend matrix, the Pearson correlation coefficient between the long-term evolution trend components of any two operating condition characteristic parameters is calculated. Taking the long-term trend components of junction temperature and on-state resistance as examples, the synchronous observation sequence of the two at 86,400 time points is obtained, and the ratio of the covariance to their respective standard deviations is calculated to obtain the correlation coefficient value. This calculation is performed pairwise on all thirty-seven characteristic parameters to obtain a static coupling correlation matrix of thirty-seven rows by thirty-seven columns. Each element in the matrix represents the degree of linear correlation between the long-term trends of the corresponding two characteristic parameters, with a value ranging from negative one to positive one. Based on the short-term fluctuation matrix, the Granger causality between the short-term dynamic fluctuation components of any two operating condition characteristic parameters is calculated. Taking the short-term fluctuation components of current and junction temperature as examples, a vector autoregressive model containing current lag terms and junction temperature lag terms is constructed, with the model order set to 10 time lag steps. The F-test is used to determine whether introducing a current hysteresis term significantly improves the prediction accuracy of the junction temperature fluctuation component. If the F-statistic exceeds a critical value, a one-way Granger causal relationship between current fluctuation and junction temperature fluctuation is considered to exist. This test is performed pairwise on all thirty-seven feature parameters, resulting in a dynamic linkage matrix of thirty-seven rows by thirty-seven columns. The elements in the matrix are binary values, where 1 indicates the existence of a Granger causal relationship and 0 indicates its absence. The static coupling correlation matrix and the dynamic linkage matrix are then fused by taking the maximum value of each element to generate a bidirectional correlation matrix of operating condition features. This matrix reflects both the long-term static dependencies between features and captures the short-term dynamic transmission paths between features.

[0026] Step S23: Based on the bidirectional correlation matrix of operating condition features, identify the core correlation feature combinations of the operating status of the dominant device in different operating periods; In this embodiment, based on the generated bidirectional correlation matrix of operating condition features, the entire 86,400 time segments of the operating time are segmented. The 24 hours of the day are divided into four operating segments, each containing 21,600 time segments. For the first operating segment, a subset of operating condition feature data from all time segments within that segment is extracted, and the corresponding local correlation matrix is ​​calculated. A spectral clustering algorithm is then performed on this local correlation matrix, with a set number of clusters of 5. The spectral clustering algorithm first constructs a Laplace matrix based on the local correlation matrix, calculates the eigenvectors corresponding to the top 5 smallest eigenvalues ​​of the Laplace matrix, and concatenates these eigenvectors column-wise to form a 21,600-row by 5-column matrix. K-means clustering is then performed on the row vectors of this matrix, dividing the 21,600 time segments into 5 clusters. For each cluster, the average correlation strength of each operating condition feature parameter within the cluster is calculated, and the top three feature parameter combinations with the highest average correlation strength are determined as the core correlation feature combinations for that cluster. The same spectral clustering and statistical operations were repeated for the second, third, and fourth operating periods to identify the core correlation feature combinations within each period. A total of twenty core correlation feature combinations were generated across the four operating periods. Each combination contains three operating condition feature parameters, and these combinations represent the key feature set of the dominant power device operating states in different operating periods.

[0027] Step S24: Extract the state representation attributes and temporal evolution attributes of each core associated feature combination to form differentiated operating condition associated feature units; aggregate all differentiated operating condition associated feature units and integrate them to generate a multi-dimensional operating condition associated feature set covering power devices throughout the entire operating period.

[0028] In this embodiment, based on the twenty identified core correlation feature combinations, state characterization attributes and temporal evolution attributes are extracted for each combination. Taking the core correlation feature combination consisting of junction temperature, current, and on-state resistance in the first operating segment as an example, the state characterization attributes include the mean vector and covariance matrix of the three feature parameters within the combination during the time period. The mean vector reflects the typical operating point of the combination, and the covariance matrix reflects the dispersion and correlation structure among the features within the combination. The temporal evolution attributes include the rate of change vector and direction of change vector of the three feature parameters within the combination during the time period. The rate of change is obtained by taking the first difference of the time series and averaging the absolute values, and the direction of change is obtained by calculating the slope sign of the time series. The state characterization attributes and temporal evolution attributes are merged to form a differential condition correlation feature unit containing multi-dimensional values. The same attribute extraction operation is repeated for the remaining nineteen core correlation feature combinations to obtain twenty differential condition correlation feature units. These twenty differential condition correlation feature units are arranged in chronological order, with the five units generated in the first operating segment at the beginning, followed by the five units generated in the second operating segment, and so on. All units together constitute a multi-dimensional operating condition-related feature set for power devices covering the entire operating period. This feature set not only includes statistical information of the original operating condition feature parameters, but also incorporates the correlation between features and the temporal evolution law, providing semantically rich feature representations for subsequent operating condition clustering and lifetime assessment.

[0029] Furthermore, step S22 includes the following steps: Time-series normalization mapping is performed on each feature parameter in the multidimensional working condition feature dataset to eliminate the dimensional differences between different feature dimensions and generate a set of time-series feature sequences of equal dimensions. In this embodiment, based on a multi-dimensional operating condition feature dataset, a time-series normalization mapping operation is performed on each operating condition feature parameter. Taking the current RMS value feature parameter as an example, the complete time series sequence of this parameter across 86,400 time segments is obtained, and the minimum and maximum values ​​of the sequence are calculated to obtain the minimum and maximum observed values ​​of the current RMS value. A min-max normalization transformation is performed on each value in the sequence, subtracting the minimum observed value from the original value and then dividing by the difference between the maximum and minimum observed values ​​to obtain a normalized value mapped to a closed interval of 0 to 1. The same normalization operation is repeated for the on-state resistance feature parameter, calculating the minimum and maximum values ​​of the parameter sequence and mapping each original value to the 0 to 1 interval. The same min-max normalization transformation is performed on each of the other thirty-five operating condition feature parameters, such as the maximum and minimum junction temperatures and the switching frequency, to ensure that the numerical range of each feature parameter is compressed to the standard interval of 0 to 1. After normalization, all thirty-seven operating condition characteristic parameters generate a time-series characteristic sequence group with thirty-seven rows and 86,400 columns. Each row in this sequence group represents the normalized trajectory of a characteristic parameter across all time sections, and each column represents the normalized value of all characteristic parameters at the same time section. Since the dimensional differences of all characteristic parameters have been eliminated, the difference in units between current ampere-level values ​​and junction temperature Celsius-level values ​​no longer affects the fairness of subsequent correlation analysis.

[0030] Based on equal-dimensional time-series feature sequence groups, a nonlinear correlation fitting algorithm is used to calculate the static correlation degree between any two working condition feature parameters and construct a feature static correlation matrix. In this embodiment, based on an equal-dimensional time-series feature sequence group, a nonlinear correlation fitting algorithm is used to calculate the static correlation degree between any two operating condition feature parameters. Taking the junction temperature normalized sequence and the on-state resistance normalized sequence as examples, synchronous observation pairs of the two at 86,400 time points are obtained, forming a two-dimensional scatter plot containing 86,400 data points. Local polynomial regression fitting is performed on this scatter plot, setting the fitting order to second order and the bandwidth parameter to 0.1, that is, each fitting point only considers 10% of the data points around it. At each fitting point, the local polynomial regression selects a subset of data points within the bandwidth centered on that point, assigns ternary quadratic weights to the data points in the subset, with the weight increasing the closer to the center point, and fits a second-order polynomial surface, using the value of this surface at the center point as the fitted value. The coefficient of determination between the original observation value and the fitted value at all data points is calculated. This coefficient measures the proportion of the original data variation explained by the fitted surface, and its value ranges from 0 to 1. This coefficient of determination is used as the static correlation degree between junction temperature and on-state resistance. By performing the same local polynomial regression fitting and coefficient of determination calculation on each pair of all thirty-seven feature parameters, a symmetric matrix of thirty-seven rows by thirty-seven columns is obtained, which is the feature static correlation matrix. All diagonal elements in this matrix are 1, and the off-diagonal elements represent the strength of the nonlinear static correlation between corresponding feature parameters.

[0031] The short-term fluctuation components of characteristic parameters under various operating conditions are processed by sliding time-series slicing to obtain the characteristic fluctuation amplitude and fluctuation rate in different time-series slices; the dynamic linkage response relationship of each characteristic fluctuation parameter in different time-series slices is calculated to generate a time-series dynamic correlation sequence set. In this embodiment, a sliding time-slicing process is performed based on the short-term dynamic fluctuation component sequences of various operating condition characteristic parameters. Taking the junction temperature short-term fluctuation component sequence as an example, the sliding window length is set to 600 time points, i.e., a ten-minute time span, and the sliding step size is 60 time points, i.e., a one-minute time interval. Starting from the first time point of the sequence, 60 time points are slid forward each time, and 600 data points within the window are extracted as a time-series slice. This sliding slicing operation is performed on the entire sequence of 86,400 time points, generating a total of 1,430 time-series slices. For each time-series slice, the standard deviation of the fluctuation data within the slice is calculated as the fluctuation amplitude, and the average of the absolute values ​​of the first-order differences of the fluctuation data within the slice is calculated as the fluctuation rate. The same sliding slicing operation is repeated for the short-term fluctuation component sequences of the current short-term fluctuation component sequence, the on-state resistance short-term fluctuation component sequence, and other thirty-six characteristic parameters to obtain the fluctuation amplitude sequence and fluctuation rate sequence corresponding to each characteristic parameter. For the same time-series slice, the Pearson correlation coefficient between the junction temperature fluctuation amplitude and the current fluctuation amplitude is calculated to obtain the dynamic linkage response relationship within that slice. The same correlation coefficient calculation is repeated for junction temperature fluctuation rate and current fluctuation rate. The same operation is repeated for combinations of junction temperature with on-state resistance, junction temperature with switching frequency, and other characteristic parameters. The dynamic linkage response relationships calculated in all time slices are arranged in chronological order to generate a time-series dynamic correlation sequence set. Each element in this sequence set corresponds to the dynamic linkage strength of a specific combination of characteristic parameters in a specific time slice.

[0032] By integrating the static correlation matrix of features and the time-series dynamic correlation sequence set, a bidirectional correlation matrix of working condition features is generated.

[0033] In this embodiment, a fusion operation is performed based on the feature static correlation matrix and the time-series dynamic correlation sequence set to generate a bidirectional correlation matrix of operating condition features. Taking the combination of junction temperature and on-state resistance as an example, the static correlation degree value between junction temperature and on-state resistance is read from the feature static correlation matrix and recorded as the static correlation value. The dynamic linkage response relationship values ​​between junction temperature and on-state resistance in all 1430 time slices are read from the time-series dynamic correlation sequence set to form a dynamic correlation sequence of length 1430. The mean of this dynamic correlation sequence is calculated to obtain the dynamic correlation mean. The static correlation value and the dynamic correlation mean are weighted and summed, with the weight of the static correlation value set to 0.4 and the weight of the dynamic correlation mean set to 0.6, to obtain the bidirectional correlation degree between junction temperature and on-state resistance. The same weighted summation operation is repeated for the other 666 feature parameter combinations, such as junction temperature and current, junction temperature and switching frequency, to obtain the bidirectional correlation degree of all feature parameter combinations. All bidirectional correlation values ​​are filled into a matrix of 37 rows by 37 columns. The element in the i-th row and j-th column of the matrix represents the bidirectional correlation between the i-th feature parameter and the j-th feature parameter, and the diagonal elements are set to 1. This matrix is ​​the bidirectional correlation matrix of working condition features, in which each element comprehensively reflects the overall correlation strength between the corresponding two feature parameters in both long-term static dependence and short-term dynamic linkage.

[0034] Furthermore, the calculation of the dynamic linkage response relationship of each characteristic fluctuation parameter within different time slices includes the following steps: The characteristic fluctuation amplitude and fluctuation rate within each time slice are classified into stable fluctuation state and sudden fluctuation state. In this embodiment, a state classification operation is performed on each time-series slice based on the characteristic fluctuation amplitude and fluctuation rate data within each time-series slice. Taking the junction temperature characteristic parameter data in the first time-series slice as an example, the values ​​of the junction temperature fluctuation amplitude and fluctuation rate within this slice are obtained. The fluctuation amplitude classification threshold is set to 1.5 times the average fluctuation amplitude of all characteristic parameters within this slice, and the fluctuation rate classification threshold is set to 1.5 times the average fluctuation rate of all characteristic parameters within this slice. The junction temperature fluctuation amplitude is compared with the fluctuation amplitude classification threshold. If the junction temperature fluctuation amplitude is lower than the threshold, the junction temperature is determined to be in a stable state in the fluctuation amplitude dimension; if the junction temperature fluctuation amplitude is higher than or equal to the threshold, the junction temperature is determined to be in a sudden change state in the fluctuation amplitude dimension. The same threshold comparison operation is performed on the junction temperature fluctuation rate. Combining the determination results of the two dimensions of fluctuation amplitude and fluctuation rate, if both dimensions are determined to be in a stable state, the overall state of the junction temperature within this time-series slice is classified as a stable fluctuation state; if at least one dimension is determined to be in a sudden change state, the overall state of the junction temperature within this time-series slice is classified as a sudden fluctuation state. The same threshold comparison and comprehensive judgment operation is repeated for the other thirty-six characteristic parameters, such as current, on-state resistance, and switching frequency, within the same time slice to obtain the state classification labels for all characteristic parameters within that time slice. The same state classification operation is performed on the remaining one thousand four hundred and twenty-nine time slices one by one, and each characteristic parameter within each time slice is assigned a state classification label, with the label value being either a stable fluctuation state or a sudden fluctuation state.

[0035] For a stable fluctuation state, calculate the synchronous change coefficient between the fluctuations of different characteristic parameters to characterize the steady-state linkage relationship; In this embodiment, for time slices determined to be in a stable fluctuation state, the synchronization coefficient between fluctuations of different characteristic parameters is calculated. Taking the first time slice as an example, it is assumed that the junction temperature and current within this slice are both determined to be in a stable fluctuation state. The junction temperature fluctuation amplitude sequence and the current fluctuation amplitude sequence within this slice are obtained, both of which contain 600 data points. The covariance of the two sequences is calculated and then divided by the product of the standard deviations of the two sequences to obtain the Pearson correlation coefficient, which is the synchronization coefficient between junction temperature and current. The same correlation coefficient calculation is repeated for other characteristic parameter combinations such as junction temperature and on-state resistance, junction temperature and switching frequency, and current and on-state resistance, provided that both characteristic parameters involved in the calculation are in a stable fluctuation state within this slice. If at least one characteristic parameter in a certain characteristic parameter combination is in a sudden fluctuation state within this slice, the synchronization coefficient of that combination is marked as invalid. All valid synchronization coefficients are arranged in the index order of the characteristic parameter combinations to form a stable state linkage vector, the length of which is equal to the number of characteristic parameter combinations in a stable fluctuation state. For the remaining 1,429 time-series slices that were also determined to be in a stable fluctuation state, the same synchronous change coefficient calculation was performed one by one, generating a corresponding stable-state linkage vector for each stable fluctuation state slice. All stable-state linkage vectors together constitute a steady-state linkage relationship dataset, which records the coordinated change patterns among the characteristic parameters of various operating conditions of the power devices during the stable operation period.

[0036] For abrupt fluctuations, the difference between the fluctuation response delay and the fluctuation amplitude of each characteristic parameter is extracted to characterize the transient linkage relationship. In this embodiment, for time-series slices identified as experiencing abrupt fluctuations, the difference between the fluctuation-induced change response delay and the change amplitude of each feature parameter is extracted. Taking the first time-series slice as an example, it is assumed that both the junction temperature and current within this slice are identified as experiencing abrupt fluctuations. The junction temperature fluctuation amplitude sequence and the current fluctuation amplitude sequence within this slice are obtained, each containing 600 data points. A peak detection algorithm is performed on the junction temperature fluctuation amplitude sequence to find the maximum point in the sequence that exceeds the mean plus three standard deviations. The time index position of this maximum point within the slice is recorded as the time of the junction temperature abrupt change. The same peak detection operation is performed on the current fluctuation amplitude sequence, and the time of the current abrupt change is recorded. The difference between the time of the current abrupt change and the time of the junction temperature abrupt change is calculated. If the difference is positive, it indicates that the junction temperature abrupt change precedes the current abrupt change, and this difference is the current response delay to the junction temperature fluctuation abrupt change; if the difference is negative, it indicates that the current abrupt change precedes the junction temperature abrupt change, and the absolute value of this difference is the junction temperature response delay to the current fluctuation abrupt change. Simultaneously, the absolute value of the difference between the fluctuation amplitude at the moment of junction temperature abrupt change and the fluctuation amplitude at the moment of current abrupt change is calculated as the abrupt change amplitude difference. The same peak detection and delay calculation operations are repeated for other characteristic parameter combinations such as junction temperature and on-state resistance, junction temperature and switching frequency, and current and on-state resistance, provided that both characteristic parameters involved in the calculation are in abrupt fluctuation state within the slice. All effective response delays and abrupt change amplitude differences are arranged in the index order of the characteristic parameter combinations to form an abrupt change state linkage vector. The same response delay and amplitude difference extraction operations are performed on each of the remaining 1429 time-series slices that are also determined to be in abrupt fluctuation state, generating a corresponding abrupt change state linkage vector for each slice. All abrupt change state linkage vectors together constitute a transient linkage relationship dataset, which records the causal conduction speed and impact intensity between the characteristic parameters of various operating conditions of the power device during the abnormal fluctuation period.

[0037] By combining steady-state and transient linkage relationships, a feature dynamic linkage model within a single time slice is constructed; the feature dynamic linkage models of all time slices are traversed and integrated to generate a set of time-series dynamic association sequences.

[0038] In this embodiment, a feature dynamic linkage model is constructed for each time series slice based on the steady-state linkage relationship dataset and the transient linkage relationship dataset. Taking the first time series slice as an example, if the slice is determined to be in a stable fluctuation state, the steady-state linkage vector corresponding to the slice is read from the steady-state linkage relationship dataset, and this vector is directly used as the feature dynamic linkage model for the slice. The model only includes the synchronization change coefficient as the linkage relationship metric. If the slice is determined to be in a sudden fluctuation state, the sudden fluctuation state linkage vector corresponding to the slice is read from the transient linkage relationship dataset, and this vector is directly used as the feature dynamic linkage model for the slice. The model includes two linkage relationship metrics: fluctuation sudden change response delay and sudden change amplitude difference. If some feature parameter combinations in the slice are in a stable fluctuation state and some feature parameter combinations are in a sudden fluctuation state, the steady-state linkage vector and the sudden fluctuation state linkage vector are concatenated to form a hybrid linkage vector, which is used as the feature dynamic linkage model for the slice. The same model construction operation is performed on the remaining 1,429 time series slices one by one, with each time series slice corresponding to a unique feature dynamic linkage model. All 1,430 dynamic linkage models are arranged in chronological order, with the models from the first slice at the beginning and the models from the last slice at the end, forming a time-series dynamic correlation sequence set. Each element in this sequence set is a vector, and the values ​​in the vectors reflect the dynamic linkage relationships between the feature parameters of each operating condition within the corresponding time-series slice. The vector length varies depending on the number of feature parameter combinations in stable and abrupt states within the slice.

[0039] Furthermore, step S3 includes the following steps: Step S31: Based on the multi-dimensional working condition association feature set, extract the spatial distribution density and temporal aggregation characteristics of each working condition feature, and construct a spatial distribution model of working condition features; In this embodiment, a multi-dimensional operating condition associated feature set is used, which contains twenty differentiated operating condition associated feature units. Each unit consists of state representation attributes and temporal evolution attributes. For each differentiated operating condition associated feature unit, its spatial distribution density attribute is extracted. Taking the first differentiated operating condition associated feature unit as an example, this unit contains the mean vector and covariance matrix of three feature parameters—junction temperature, current, and on-state resistance—during the first operating segment. Based on the mean vector and covariance matrix, a three-dimensional Gaussian distribution function is constructed. The center of this function is located at the mean vector, and the covariance matrix determines the shape and direction of the distribution. The probability density value of this three-dimensional Gaussian distribution function in the feature space is calculated to obtain the spatial distribution density value at each feature point. The same Gaussian distribution construction and probability density calculation operations are repeated for the second to twentieth differentiated operating condition associated feature units to obtain twenty spatial distribution density fields. Simultaneously, the temporal aggregation characteristics of each unit are extracted. Taking the first differentiated operating condition associated feature unit as an example, the continuous distribution of 21,600 time segments within the first operating period of this unit on the time axis is statistically analyzed. The average time interval between adjacent time segments is calculated. If the average time interval is less than a preset threshold, the unit is determined to have high temporal aggregation characteristics. The same temporal aggregation characteristic calculation is performed on all twenty differentiated operating condition associated feature units to obtain the temporal aggregation degree index for each unit. The spatial distribution density field and the temporal aggregation degree index are associated and stored to construct a spatial distribution model of operating condition characteristics. This model contains twenty density fields and corresponding twenty aggregation degree indices, which fully describe the distribution law of power device operating condition characteristics in both spatial and temporal dimensions.

[0040] Step S32: Based on the working condition feature space distribution model, adaptively identify high-density clustered regions and low-density discrete regions in the feature space, and divide the working condition clustering core domain and the working condition discrete domain. In this embodiment, based on the constructed working condition feature space distribution model, adaptive identification is performed on regions in the feature space. Taking the three-dimensional Gaussian distribution density field corresponding to the first differentiated working condition associated feature unit as an example, the probability density values ​​of all grid points in the density field are traversed, and the average and standard deviation of the probability density of all grid points are calculated. The threshold for determining high-density clustered regions is set to the average value plus twice the standard deviation, and grid points with probability density values ​​higher than this threshold are marked as high-density clustered regions. The threshold for determining low-density discrete regions is set to the average value minus one standard deviation, and grid points with probability density values ​​lower than this threshold are marked as low-density discrete regions. Grid points with probability density values ​​between the two thresholds are marked as transition regions. Connectivity analysis is performed on the grid points in the high-density clustered regions, and adjacent grid points are merged into a connected domain, with each connected domain corresponding to a clustering core domain. The same connectivity analysis is performed on the grid points in the low-density discrete regions, and adjacent grid points are merged into a connected domain, with each connected domain corresponding to a working condition discrete domain. Repeat the same threshold determination and connectivity analysis operations for the density fields corresponding to the second to twentieth differential working condition associated feature units to obtain the distribution of cluster core domains and working condition discrete domains in all twenty density fields. Cluster core domains and working condition discrete domains belonging to the same density field are associated and recorded. Each cluster core domain is accompanied by its covered grid point coordinate range and corresponding probability density value range, and each working condition discrete domain is accompanied by its covered grid point coordinate range and corresponding probability density value range.

[0041] Step S33: Based on the feature aggregation attributes of the clustering core domain, complete the merging of similar steady-state operating conditions, and based on the feature difference attributes of the discrete operating condition domain, complete the differentiated splitting of transient operating conditions. In this embodiment, based on the identified clustering core domains and discrete operating condition domains, steady-state operating conditions are merged into similar types, and transient operating conditions are differentiated and split. Taking the first clustering core domain in the first density field as an example, this core domain covers a continuous grid point region in the feature space, where the probability density values ​​of all grid points are higher than a high-density threshold. The original operating condition feature data corresponding to all grid points within this core domain are extracted. This data comes from a multi-dimensional operating condition feature dataset and includes the original values ​​of thirty-seven feature parameters such as junction temperature, current, and on-state resistance. The mean vector and covariance matrix of these original data are calculated as representative feature templates for this core domain. The same mean vector and covariance matrix calculation is performed on the clustering core domains in the second density field. Clustering core domains with similar mean vectors and covariance matrices in different density fields are merged. The similarity is determined by the Euclidean distance between the two mean vectors being less than a preset distance threshold and the Frobenius norm difference between the two covariance matrices being less than a preset norm threshold. All clustering core domains that meet the merging criteria are grouped into the same steady-state operating condition cluster. For the operating condition discrete domain, taking the first operating condition discrete domain in the first density field as an example, this discrete domain covers a low-density region in the feature space. The original operating condition feature data corresponding to all grid points within this discrete domain are extracted, and the mean vector and covariance matrix of these data are calculated. Since the data points within the discrete domain are sparsely and dispersed, different discrete domains are not merged; instead, each discrete domain is treated as a separate transient operating condition unit. In cases where multiple discrete domains overlap or are nested, they are divided into non-overlapping sub-regions according to their spatial location, with each sub-region corresponding to an independent transient operating condition unit.

[0042] Step S34: Remove isolated abnormal operating conditions in the feature space to avoid interference from accidental abnormal operating conditions on the clustering results; In this embodiment, isolated anomalous operating condition points in the feature space are removed based on steady-state operating condition clusters and transient operating condition units. Taking the first steady-state operating condition cluster as an example, this cluster contains the merged results of multiple clustering core domains, covering a large continuous region in the feature space. The local reachability density of all data points in this cluster is calculated. For each data point, neighboring points within a radius centered on it are selected, with the number of neighboring points set to fifty. The sum of the reachability distances from the center point to all neighboring points is calculated, and the reciprocal is taken to obtain the local reachability density. The local outlier factor of each data point is calculated, and the local reachability density of the center point is compared with that of its neighboring points. If the local reachability density of the center point is significantly lower than that of its neighboring points, the local outlier factor is greater than 1. The local outlier factor threshold is set to 2, and data points with a local outlier factor greater than 2 are marked as isolated anomalous operating condition points. These isolated anomalous operating condition points are removed from the steady-state operating condition cluster, and the removed data points are reclassified as discrete points. The same local reachability density calculation and local outlier determination operations were repeated for the second to all steady-state operating condition clusters, removing isolated outlier points from each cluster. The same local outlier calculation was performed on each data point in the transient operating condition unit. Due to the sparse data in the transient operating condition unit, the local outlier factor was generally high; only extreme outliers with a local outlier factor exceeding 5 were removed. All removed isolated outlier points were collected into a separate outlier set and did not participate in subsequent operating condition clustering and lifetime assessment.

[0043] Step S35: Complete the clustering and division of the full operating conditions, and generate various operating condition clusters, including steady-state operating condition clusters and transient operating condition clusters.

[0044] In this embodiment, based on the steady-state operating condition clusters and transient operating condition units, and after removing isolated abnormal operating condition points, the clustering and partitioning of the entire operating condition is completed. All data points corresponding to the core domains of the clusters after merging and removing data points are assigned to a set of steady-state operating condition clusters. Each steady-state operating condition cluster contains a group of data points with similar characteristic distributions. These data points correspond to the operating states of power devices under stable load, constant temperature, and normal vibration conditions. All data points corresponding to the discrete operating condition domains after splitting and removing data points are assigned to a set of transient operating condition clusters. Each transient operating condition cluster contains a group of data points with dispersed characteristic distributions but spatially close proximity. These data points correspond to the operating states of power devices under abnormal conditions such as sudden load increases, sudden temperature changes, or vibration shocks. For each steady-state operating condition cluster, the number of data points and their proportion of all data points are counted, and the mean vector and standard deviation vector of the data within the cluster are calculated as a representative operating condition description of the cluster. For each transient operating condition cluster, the number and proportion of data points are counted, and the mean vector and standard deviation vector are calculated. All steady-state and transient operating condition clusters are sorted in descending order of the number of data points, and each cluster is assigned a unique cluster number. The final set of operating condition clusters contains several steady-state clusters and several transient clusters. Each cluster has a clear feature space boundary and a representative operating condition description, providing a refined basis for subsequent lifetime assessment of power devices based on different operating condition types.

[0045] Furthermore, step S4 includes the following steps: Step S41: Analyze the characteristic distribution attributes of various operating condition clusters and match the conduction and switching states of power devices under the corresponding operating conditions; In this embodiment, feature distribution attribute parsing is performed on each operating condition cluster based on various operating condition clusters. Taking the first steady-state operating condition cluster as an example, this cluster contains a set of data points clustered in the feature space, with each data point corresponding to thirty-seven operating condition feature parameters at a time cross-section. The average junction temperature, average current, and average on-state resistance of all data points in this cluster are extracted. These three averages reflect the typical conduction operating state corresponding to this cluster. The higher the average junction temperature and the larger the average current, the greater the conduction current density and the higher the conduction voltage drop of the power device under this operating condition. The average switching frequency, average voltage rise rate, and average current fall rate of all data points in this cluster are extracted. These three averages reflect the typical switching operating state corresponding to this cluster. The higher the average switching frequency, the faster the voltage rise rate and the faster the current fall rate, the more frequent the switching action of the power device under this operating condition, and the greater the voltage and current stress during the switching process. The same extraction operations for the average junction temperature, average current, average on-state resistance, average switching frequency, average voltage rise rate, and average current fall rate were repeated for the second to last steady-state operating condition clusters. Each steady-state cluster yielded a set of on-state parameters and a set of switching state parameters. The same parameter extraction operations were performed on the transient operating condition clusters. Due to the dispersed distribution of data points within the transient clusters, the median was used instead of the mean as the representative parameter of the cluster to reduce the impact of extreme values ​​on parameter estimation. The on-state and switching state parameters for all operating condition clusters were compiled into a lookup table, with each row corresponding to one operating condition cluster and each column corresponding to one operating state parameter.

[0046] Step S42: Based on the semiconductor physical loss mechanism of the device, establish a corresponding correlation model between the operating condition cluster characteristics and the device conduction loss and switching loss; In this embodiment, based on the analyzed conduction and switching state parameters of each operating condition cluster, a corresponding correlation model is established in conjunction with the semiconductor physical loss mechanism of the device. The conduction loss is calculated as follows: conduction loss power equals the on-state resistance multiplied by the square of the effective current value. Taking the first steady-state operating condition cluster as an example, the average on-state resistance and average current of the cluster are obtained, and the average on-state resistance is multiplied by the square of the average current value to obtain the estimated conduction loss power value for that cluster. The switching loss is calculated as follows: switching loss power equals the switching frequency multiplied by the single-switch energy loss, which is proportional to the product of the voltage rise rate and the current fall rate. Taking the first steady-state operating condition cluster as an example, the average switching frequency, average voltage rise rate, and average current fall rate of the cluster are obtained, and the average voltage rise rate is multiplied by the average current fall rate, then multiplied by a proportionality coefficient to obtain the estimated single-switch energy loss value. This estimated value is then multiplied by the average switching frequency to obtain the estimated switching loss power value for that cluster. The same calculations for conduction and switching losses were repeated for the second through last steady-state clusters, yielding a set of loss estimates for each cluster. The same calculations were then performed on the transient clusters, using the median parameter and the same formula. The estimated conduction and switching losses for all clusters were then arranged by cluster number to form a correlation model between cluster characteristics and device losses. This model takes the cluster's characteristic parameters as input and outputs conduction and switching losses.

[0047] Step S43: Calculate the real-time power loss and power loss accumulation rate of devices under various operating conditions using the correlation model; In this embodiment, based on the established correspondence model, the real-time power loss and power accumulation rate of devices under various operating conditions are calculated. Taking the first steady-state operating condition cluster as an example, the estimated conduction power loss and the estimated switching power loss of the cluster are obtained, and the two are added together to obtain the estimated total power loss of the cluster. The number of data points contained in the cluster is obtained, that is, the number of operating time segments corresponding to the cluster, with each time segment representing one second of operating time. The estimated total power loss is multiplied by the number of data points of the cluster to obtain the power accumulation of the cluster over all operating periods. The power accumulation is divided by the number of data points of the cluster to obtain the average power loss of the cluster, which is the estimated real-time power loss of the cluster. The power accumulation is divided by the total operating time of the cluster to obtain the power accumulation rate of the cluster, in watts per second. The same calculation operations of total power loss, power accumulation, real-time power loss, and power accumulation rate are repeated for the second to the last steady-state operating condition clusters. The same calculation operations are performed for the transient operating condition cluster. The real-time power loss and cumulative power loss rate of all operating condition clusters are compiled into a table. Each row in the table corresponds to an operating condition cluster and includes five fields: cluster number, number of data points, total power loss, real-time power loss, and cumulative power loss rate. This table quantifies the energy consumption level and aging rate of power devices under different operating conditions.

[0048] Step S44: Combine the time-series operating characteristics of each operating condition cluster to generate the time-series dynamic evolution law of loss parameters; In this embodiment, based on the calculated real-time power loss and power loss accumulation rate of various operating condition clusters, and combined with the time-series operating characteristics of each operating condition cluster, the time-series dynamic evolution law of the power loss parameters is generated. Taking the first steady-state operating condition cluster as an example, the time section index list corresponding to the cluster is obtained. This list records the specific position of each data point in the cluster in 86,400 time sections. With the time section index as the horizontal axis and the real-time power loss of the cluster as the vertical axis, the power loss values ​​are plotted at all time section indices corresponding to the cluster, forming a time-series scatter plot of the power loss of the cluster. Local weighted regression smoothing is performed on the data points in the scatter plot, and the smoothing window width is set to 600 time points to obtain a smooth curve of the power loss of the cluster changing with time. This curve is the time-series evolution law of the power loss of the cluster. The same time-series scatter plot plotting and smoothing curve generation operation is repeated for the second to the last steady-state operating condition clusters, and a power loss time-series evolution curve is obtained for each steady-state operating condition cluster. For transient operating condition clusters, since their data points are scattered and discontinuous, no smoothing operation is performed. Instead, the power loss values ​​at each data point are directly arranged in chronological order to form a discrete power loss pulse sequence. This sequence represents the temporal evolution of power loss for the transient operating condition cluster. The temporal evolution of power loss for all operating condition clusters is superimposed onto the same time axis in chronological order, forming a dynamic evolution map of power loss parameters covering the entire operating period. This map visually demonstrates the fluctuation pattern of power loss when different operating conditions alternate.

[0049] Step S45: Integrate the loss time-series evolution patterns of various operating condition clusters to generate the device loss dynamic characteristic sequence corresponding to each operating condition cluster.

[0050] In this embodiment, based on the time-series evolution patterns of power loss in various operating condition clusters, a dynamic feature sequence of device power loss corresponding to each operating condition cluster is generated. Taking the first steady-state operating condition cluster as an example, the smoothed curve of the time-series evolution of power loss for this cluster is obtained. This curve contains the power loss values ​​at all time points corresponding to this cluster. Three dynamic feature parameters are extracted from this curve: mean power loss, power loss fluctuation amplitude, and power loss change rate. The mean power loss is obtained by taking the arithmetic mean of all values ​​on the curve. The power loss fluctuation amplitude is obtained by averaging the absolute values ​​of the differences between all values ​​on the curve and the mean. The power loss change rate is obtained by performing a first-order difference operation on the curve and then averaging the absolute values ​​of the difference results. These three dynamic feature parameters are concatenated into a three-dimensional vector in the order of mean, fluctuation amplitude, and change rate, which serves as the device power loss dynamic feature sequence for the first steady-state operating condition cluster. The same extraction and vector concatenation operations for mean, fluctuation amplitude, and change rate are repeated for the second to the last steady-state operating condition clusters, generating a three-dimensional dynamic feature sequence of power loss for each steady-state operating condition cluster. For transient operating condition clusters, the discrete loss power pulse sequence of the cluster is obtained. Three dynamic feature parameters—the average pulse peak value, pulse duration, and pulse occurrence frequency—are extracted from this sequence and concatenated into a three-dimensional vector as the device loss dynamic feature sequence of the transient operating condition cluster. The three-dimensional loss dynamic feature sequences of all operating condition clusters are arranged according to cluster number, forming a multi-row, three-column matrix. Each row of the matrix corresponds to the loss dynamic feature description of one operating condition cluster. This matrix is ​​the set of device loss dynamic feature sequences corresponding to each operating condition cluster.

[0051] Furthermore, step S42 includes the following steps: Extract the core operational characteristic parameters of clusters under various operating conditions and construct a cluster feature representation vector; In this embodiment, based on various operating condition clusters, a core operating feature parameter extraction operation is performed on each operating condition cluster. Taking the first steady-state operating condition cluster as an example, this cluster contains a set of data points clustered in the feature space, with each data point corresponding to thirty-seven operating condition feature parameters at a time cross-section. From these thirty-seven feature parameters, six core parameters directly related to the operating state of the power device are selected: RMS current value, on-state resistance value, maximum junction temperature, switching frequency, voltage rise rate, and current fall rate. The mean values ​​of all data points in this cluster on these six core parameters are calculated, resulting in six mean values. These six mean values ​​are arranged in a fixed order: RMS current value, on-state resistance value, maximum junction temperature, switching frequency, voltage rise rate, and current fall rate, forming a six-dimensional vector, which serves as the operating condition cluster feature representation vector for the first steady-state operating condition cluster. The same six core parameter mean calculation and six-dimensional vector construction operation are repeated for the second to the last steady-state operating condition clusters, generating a unique six-dimensional feature representation vector for each steady-state operating condition cluster. For transient operating condition clusters, due to the dispersed distribution of data points, the median is used instead of the mean to calculate the representative values ​​of the six core parameters, which are also arranged in a fixed order to form a six-dimensional vector. The six-dimensional feature representation vectors of all operating condition clusters are organized into a matrix, with the number of rows equal to the total number of operating condition clusters and the number of columns equal to six. Each row in the matrix corresponds to the digital feature description of one operating condition cluster.

[0052] Based on the semiconductor carrier transport characteristics of power devices, the core influencing factors of device loss under different operating conditions are determined. In this embodiment, based on the semiconductor carrier transport characteristics of power devices, the core influencing factors of device losses under different operating conditions are determined. The core influencing factors for conduction losses are on-state resistance and RMS current, because conduction loss power equals on-state resistance multiplied by the square of the RMS current. On-state resistance determines the scattering resistance encountered by carriers moving in the drift region, and the RMS current determines the number of carriers passing through the device per unit time. The core influencing factors for switching losses are switching frequency, voltage rise rate, and current fall rate, because switching loss power equals switching frequency multiplied by the energy loss per switching cycle. The energy loss per switching cycle is proportional to the product of the voltage rise rate and the current fall rate. The voltage rise rate reflects the rate of change of electric field strength during turn-off, and the current fall rate reflects the rate of change of carrier concentration gradient during turn-on. The maximum junction temperature serves as an auxiliary influencing factor, affecting both conduction and switching losses, because the mobility and threshold voltage of semiconductor materials change with temperature. At high temperatures, carrier mobility decreases, leading to increased on-state resistance, and the minority carrier lifetime shortens, resulting in changes in switching speed. The on-state resistance and RMS current are labeled as the core influencing factors of conduction loss; the switching frequency, voltage rise rate, and current fall rate are labeled as the core influencing factors of switching loss; and the maximum junction temperature is labeled as a dual influencing factor. These six core influencing factors correspond one-to-one with the six extracted core operating characteristic parameters, establishing a physical mechanism-level correlation.

[0053] Establish a mapping relationship between the characteristic vector of the working condition cluster and the core influencing factors of loss; In this embodiment, a mapping relationship is established between the characteristic representation vector of the operating condition cluster and the core influencing factor of losses. Taking the six-dimensional characteristic representation vector of the first steady-state operating condition cluster as an example, the first element of this vector is the mean effective value of current, which is mapped to the effective value variable of current in the core influencing factor of conduction losses; the second element is the mean on-state resistance, which is mapped to the on-state resistance variable in the core influencing factor of conduction losses; the third element is the mean maximum junction temperature, which is mapped to the junction temperature variable in the dual influencing factor; the fourth element is the mean switching frequency, which is mapped to the switching frequency variable in the core influencing factor of switching losses; the fifth element is the mean voltage rise rate, which is mapped to the voltage rise rate variable in the core influencing factor of switching losses; and the sixth element is the mean current fall rate, which is mapped to the current fall rate variable in the core influencing factor of switching losses. The same element-wise mapping operation is performed on the six-dimensional characteristic representation vectors of the second to the last operating condition clusters. The characteristic representation vector of each operating condition cluster is decomposed into six independent physical variables, corresponding to the core influencing factors of conduction losses and switching losses, respectively. The mapping results of all operating condition clusters were compiled into a mapping table. Each row in the table corresponds to an operating condition cluster, and each column corresponds to a core influencing factor variable. The values ​​in the table are the specific values ​​of that operating condition cluster for that influencing factor. This mapping table establishes a bridge from the digital characteristics of operating condition clusters to the physical loss mechanism.

[0054] The nonlinear correlation curves between the operating condition characteristics and the conduction loss and switching loss are fitted based on the mapping relationship; In this embodiment, based on the established mapping relationship, a nonlinear correlation curve between operating condition characteristics and conduction loss and switching loss is fitted. Taking conduction loss as an example, three independent variable data are obtained: the average effective current value, the average on-state resistance, and the average maximum junction temperature for all operating condition clusters. The calculated estimated conduction loss power for each operating condition cluster is used as the dependent variable data. The three independent variables and one dependent variable form a training set containing multiple data points, with each data point corresponding to one operating condition cluster. A three-layer neural network model is constructed: the input layer contains three neurons corresponding to the three independent variables, the hidden layer contains ten neurons using the hyperbolic tangent activation function, and the output layer contains one neuron corresponding to the conduction loss power. The neural network model is trained using the training set data, with the loss function set to mean squared error, the optimizer set to the Adam optimizer, the learning rate set to 0.001, and the training epochs set to 500. After training, the neural network model fits the nonlinear correlation curve between operating condition characteristics and conduction loss. Taking switching losses as an example, we obtain four independent variable data points for all operating condition clusters: the average switching frequency, the average voltage rise rate, the average current fall rate, and the average maximum junction temperature. We also obtain the calculated estimated switching loss power for each operating condition cluster as the dependent variable. We construct another three-layer neural network model: the input layer contains four neurons, the hidden layer contains twelve neurons, and the output layer contains one neuron. We train this model using the same loss function, optimizer, and training parameters to obtain nonlinear correlation curves between operating condition characteristics and switching losses. These two nonlinear correlation curves respectively characterize the nonlinear response modes of conduction losses and switching losses as operating condition characteristics change.

[0055] A correlation model between operating condition cluster characteristics and device losses is constructed based on nonlinear correlation curves.

[0056] In this embodiment, a correlation model between the operating condition cluster characteristics and device losses is constructed based on the fitted two nonlinear correlation curves. The neural network models corresponding to the conduction loss nonlinear correlation curve and the switching loss nonlinear correlation curve are encapsulated in parallel to form a dual-output model. The input of this dual-output model is a six-dimensional vector containing six parameters: RMS current, on-state resistance, maximum junction temperature, switching frequency, voltage rise rate, and current fall rate. The first three parameters in the input vector are routed to the input layer of the conduction loss neural network model, and the last four parameters are routed to the input layer of the switching loss neural network model. The maximum junction temperature parameter is routed to the input layers of both models. The conduction loss neural network model outputs an estimated conduction loss power, and the switching loss neural network model outputs an estimated switching loss power. The two output values ​​are added to obtain the estimated total loss power. A forward propagation verification is performed on this dual-output model. Using the six-dimensional feature representation vector of the first steady-state operating condition cluster as input, the model outputs estimated conduction loss power and estimated switching loss power. The two estimates are compared with reference values ​​calculated using physical formulas, and the relative error is calculated. If the relative error exceeds 5%, the weight parameters of the two neural network models are fine-tuned. The fine-tuning data uses the mapping table and loss reference values ​​of all operating condition clusters, and the fine-tuning rounds are set to one hundred. After fine-tuning, the dual-output model is the final correlation model between operating condition cluster characteristics and device losses, which can directly output the corresponding conduction loss and switching loss estimates based on the six-dimensional feature representation vector of any operating condition cluster.

[0057] Furthermore, the fitting of the nonlinear correlation curves between the operating condition characteristics and the conduction loss and switching loss based on the mapping relationship includes the following steps: Based on the characteristic representation vector of the working condition cluster, the gradient working condition operating range is divided; In this embodiment, a gradient interval partitioning operation is performed on the feature space based on the six-dimensional feature representation vector of each operating condition cluster. Taking the effective current value as an example, the values ​​of all operating condition clusters in this dimension are obtained, and the minimum and maximum values ​​are found. The numerical range between the minimum and maximum values ​​is evenly divided into ten equally wide gradient intervals, with the width of each interval equal to the difference between the maximum and minimum values ​​divided by ten. For the mean effective current value of the first steady-state operating condition cluster, it is determined which gradient interval it falls into, and the interval number is recorded. The same minimum and maximum value search and ten equally wide gradient interval partitioning operation are repeated for the five dimensions of on-state resistance, maximum junction temperature, switching frequency, voltage rise rate, and current fall rate, generating ten gradient intervals for each dimension. For each element in the six-dimensional feature representation vector of the first steady-state operating condition cluster, its corresponding gradient interval number is determined, resulting in a gradient encoding vector composed of six interval numbers. The same gradient encoding vector generation operation is repeated for the second to last operating condition clusters, and each operating condition cluster obtains a unique six-bit gradient code. All operating condition clusters with the same gradient encoding vector are grouped into the same gradient-based operating condition interval. Within each interval, the operating condition clusters are all within similar numerical ranges across the six core feature dimensions, ensuring the homogeneity of operating condition features within the interval. All operating condition clusters are divided into several gradient-based operating condition intervals, with each interval containing one or more operating condition clusters.

[0058] The measured loss parameters of power devices within the operating range of each gradient chemical condition are obtained to form a sample set corresponding to the operating condition and loss. In this embodiment, based on the defined gradient operating condition intervals, the measured loss parameters of power devices within each interval are obtained. Taking the first gradient operating condition interval as an example, this interval contains several operating condition clusters with the same gradient code. From the multi-dimensional operating condition feature dataset, the time section indices corresponding to all operating condition clusters within this interval are extracted. These indices point to the specific moments when the power devices are actually operating within this gradient interval. The measured total power loss values ​​corresponding to these moments are read from the power consumption measurement device at the power device operating site. These values ​​are directly measured by a high-precision power analyzer and include the sum of conduction loss and switching loss. The measured total power loss value corresponding to each time section index is associated with the operating condition cluster number to which the time section belongs, forming a sample record. The sample record contains four fields: gradient interval number, operating condition cluster number, time section index, and measured total power loss. The same measured power loss reading and sample record generation operation is performed on all time section indices within the first gradient operating condition interval to obtain the sample record set corresponding to this interval. The same measured power loss reading and sample record generation operations were repeated for the second to last gradient operating condition intervals, generating a corresponding sample record set for each gradient interval. The sample record sets from all gradient intervals were merged to form an operating condition-loss correspondence sample set. This sample set contains the measured power device loss data under different gradient operating condition intervals, providing a real physical measurement basis for subsequent correlation curve fitting.

[0059] Based on the corresponding sample set of operating conditions and losses, a nonlinear fitting algorithm is used to fit the basic correlation curve; In this embodiment, a nonlinear fitting algorithm is used to fit the basic correlation curve based on the operating condition-loss correspondence sample set. Taking conduction loss as an example, sample data related to conduction loss are selected from the operating condition-loss correspondence sample set. Each sample contains three independent variables: effective current value, on-state resistance, and maximum junction temperature, as well as the conduction loss component after deducting the switching loss component from the measured total power loss as the dependent variable. The switching loss component is deducted by estimating the switching loss value using the corresponding switching frequency, voltage rise rate, and current fall rate, and then subtracting the estimated value from the measured total power loss. All three independent variables and one dependent variable of all samples are combined into a training dataset. A radial basis function network model is constructed. The center points of the radial basis function are automatically selected from the training data using the K-means clustering algorithm, with the number of clusters set to twenty, and the width parameter of the radial basis function set to twice the average distance between all center points. The radial basis function network model is trained using the training dataset, and the weight coefficients of the output layer are solved using the least squares method to minimize the sum of squared residuals between the model prediction value and the actual conduction loss component. After training, the radial basis function network model becomes the basic correlation curve for conduction loss. Taking switching loss as an example, sample data related to switching loss are selected from the operating condition-loss correspondence sample set. Each sample includes four independent variables: switching frequency, voltage rise rate, current fall rate, and maximum junction temperature, as well as the switching loss component after deducting the conduction loss component from the measured total power loss as the dependent variable. Another radial basis function network model is constructed, with the number of clusters set to twenty-five and the width parameter set to twice the average distance between all center points. The same least squares method is used to solve for the output layer weight coefficients, obtaining the basic correlation curve for switching loss. The two basic correlation curves respectively describe the basic mapping relationship between operating condition characteristics and conduction loss and switching loss.

[0060] The deviation component of the basic correlation curve is corrected by incorporating the temperature drift characteristics of the device carriers. In this embodiment, the deviation component of the basic correlation curve is corrected based on the fitted basic correlation curve and the carrier temperature drift characteristics of the device. The carrier temperature drift characteristic is that the mobility of the semiconductor material decreases with increasing temperature, leading to an increase in on-state resistance with increasing junction temperature, which in turn causes the conduction loss to deviate from the predicted value of the basic correlation curve. Taking the conduction loss basic correlation curve as an example, the on-state resistance temperature coefficient curve provided in the power device's technical manual is obtained. This curve gives the correction factor of the on-state resistance relative to the room temperature reference value at different junction temperatures. For each sample in the operating condition-loss corresponding sample set, the maximum junction temperature of the sample is read, and the on-state resistance correction factor at that junction temperature is obtained from the table in the on-state resistance temperature coefficient curve. The actual on-state resistance value of the sample is divided by this correction factor to obtain the equivalent on-state resistance value corrected to the room temperature reference. The equivalent on-state resistance value is substituted into the conduction loss basic correlation curve to recalculate the predicted loss value. The difference between this predicted value and the actual conduction loss component of the sample is calculated; this difference is the deviation component caused by temperature drift. The same deviation component calculation was performed on all samples to obtain a set of deviation component values. A polynomial fitting was then performed on this set of deviation component values, with the maximum junction temperature as the independent variable and the deviation component as the dependent variable. The fitting order was set to second order to obtain a correction function for the deviation component as a function of junction temperature. A similar temperature drift correction operation was performed on the basic correlation curve of switching losses to obtain the switching loss temperature coefficient curve. The correction factor for switching losses at different junction temperatures was calculated, and a deviation component correction function for switching losses was fitted to obtain the result. The two deviation component correction functions respectively describe the influence of temperature drift on the predicted values ​​of conduction loss and switching loss. The generation process of the deviation component correction function is as follows: Taking the conduction loss deviation component correction function as an example, the maximum junction temperature data of all samples in the operating condition-loss corresponding sample set is obtained, denoted as T. j The sequence, and the corresponding conduction loss deviation component data, denoted as ΔP cond Sequence. ΔP cond The calculation method is as follows: For each sample, divide the actual on-state resistance value by the correction factor obtained from the on-state resistance temperature coefficient curve to obtain the room temperature reference equivalent on-state resistance value. Substitute this equivalent value into the basic correlation curve of conduction loss to obtain the basic predicted value. Subtract this basic predicted value from the actual conduction loss component of the sample, and the difference is ΔP. cond The maximum junction temperature T j As the independent variable, let the conduction loss deviation component ΔP be the independent variable. cond Using [variable name] as the dependent variable, perform a second-order polynomial fitting. The fitting model is defined as follows: Let a0, a1, and a2 be the coefficients of the polynomial to be fitted. These three coefficients are solved using the least squares method to minimize the sum of squared residuals between the fitted values ​​and the actual deviation components at all sample points. The solution process is as follows: Construct a Vandermonde matrix A, where the first column is a vector of all 1s, and the second column is T. j Vector, third column is Vector; construct the dependent variable vector b as ΔP cond Vectors; solved by solving normal equations The coefficient vector x = [a0, a1, a2] is obtained. T Taking the switching loss deviation component correction function as an example, the maximum junction temperature data T of all samples is obtained from the operating condition-loss corresponding sample set. j The sequence, and the corresponding switching loss deviation component data, denoted as ΔP sw Sequence. ΔP sw The calculation method is as follows: For each sample, divide the actual switching loss value by the correction factor obtained from the table of the switching loss temperature coefficient curve to obtain the room temperature baseline equivalent switching loss value. Substitute this equivalent value into the basic correlation curve of switching loss to obtain the basic predicted value. Subtract this basic predicted value from the actual switching loss component of the sample, and the difference is ΔP. sw The maximum junction temperature T j The independent variable is the switching loss deviation component ΔP. sw As the dependent variable, perform a second-order polynomial fitting, and the fitted model is in the form of: Where b0, b1, and b2 are the polynomial coefficients to be fitted. Using the same Vandermonde matrix construction and normal equation solution method as for conduction loss, the coefficient vector y = [b0, b1, b2] is obtained. T The specific coefficients of the two deviation component correction functions are determined by fitting the actual sample data using the least squares method. Taking the sample data of the first gradient operating condition range as an example, if the coefficients of the fitted conduction loss deviation component correction function are a0=0.12, a1=0.0035, and a2=0.00002, then the specific form of the correction function is as follows: If the coefficients of the correction function for the switching loss deviation component obtained from the fitting are b0=-0.08, b1=0.0021, and b2=0.000015, then the specific form of the correction function is: These two second-order polynomial functions are the final deviation component correction functions, which are used to superimpose the deviation correction caused by temperature drift on the predicted value of the basic correlation curve, and the specific parameters are only examples.

[0061] The nonlinear correlation curves between the operating condition characteristics and conduction loss and switching loss are mapped and fitted based on the corrected deviation components and the basic correlation curves.

[0062] In this embodiment, based on the fitted basic correlation curve and the calculated deviation component correction function, the nonlinear correlation curves between the fitted operating condition characteristics and the conduction loss and switching loss are mapped. Taking conduction loss as an example, the basic correlation curve of conduction loss and the deviation component correction function of conduction loss are combined in series. For any set of input operating condition characteristic parameters, the effective value of current, on-state resistance, and maximum junction temperature are first input into the basic correlation curve of conduction loss to obtain the basic predicted value. At the same time, the maximum junction temperature is input into the deviation component correction function of conduction loss to obtain the deviation correction value. The basic predicted value and the deviation correction value are added to obtain the final nonlinear correlation curve predicted value. Taking switching loss as an example, the switching frequency, voltage rise rate, current fall rate, and maximum junction temperature are input into the basic correlation curve of switching loss to obtain the basic predicted value. At the same time, the maximum junction temperature is input into the deviation component correction function of switching loss to obtain the deviation correction value. The two are added to obtain the final predicted value. The combined model is verified using all samples in the operating condition-loss corresponding sample set generated in step S4232. The relative error between the predicted value and the measured value of each sample is calculated, and the mean and standard deviation of the relative error of all samples are statistically analyzed. If the mean relative error exceeds 3%, the second-order polynomial coefficients of the deviation component correction function are fine-tuned using gradient descent with a learning rate of 0.0001 and 50 iterations, until the mean relative error drops below 3%. After fine-tuning, the two combined nonlinear correlation curves represent the final mapping relationship between the operating characteristics and conduction and switching losses, accurately reflecting the nonlinear deviations caused by physical effects such as carrier temperature drift.

[0063] Furthermore, the correction of the deviation component of the basic correlation curve based on the combined device carrier temperature drift characteristics includes the following steps: Extract the device thermal state characteristics corresponding to each gradient operating condition range to obtain the real-time temperature distribution characteristics of charge carriers; In this embodiment, based on the defined gradient operating condition intervals, device thermal state feature extraction is performed for each interval. Taking the first gradient operating condition interval as an example, this interval contains several operating condition clusters with the same gradient encoding. From the multi-dimensional operating condition feature dataset, the maximum junction temperature data corresponding to all time section indices within this interval are extracted to form a time series sequence of maximum junction temperature. Statistical feature extraction is performed on this time series sequence, and the arithmetic mean of the sequence is calculated to obtain the average junction temperature value within this interval. The standard deviation of the sequence is calculated to obtain the junction temperature fluctuation amplitude. The difference between the maximum and minimum values ​​of the sequence is calculated to obtain the junction temperature variation range. Simultaneously, from the collected thermal state operating data, the case temperature data and heat sink temperature data corresponding to all time section indices within this interval are extracted, and the temperature difference between the case temperature and the heat sink temperature is calculated to obtain the thermal resistance feature parameters within this interval. The four values—average junction temperature, junction temperature fluctuation amplitude, junction temperature variation range, and thermal resistance feature parameters—are combined into a four-dimensional thermal state feature vector, which serves as the device thermal state feature description for the first gradient operating condition interval. The same operations of extracting the time series sequence of the maximum junction temperature, calculating statistical features, and constructing the four-dimensional thermal state feature vector were repeated for the second to the last gradient condition operating intervals, generating a unique thermal state feature vector for each gradient condition operating interval. The thermal state feature vectors of all gradient condition operating intervals were arranged into a matrix with the number of rows equal to the total number of gradient condition operating intervals and the number of columns equal to four.

[0064] Based on the real-time temperature distribution characteristics of charge carriers, the influence of temperature drift on the offset of device loss parameters is analyzed, and the loss deviation caused by temperature drift is quantified to construct the deviation component that needs to be compensated for device loss. In this embodiment, based on the extracted thermal state characteristics of the device in each gradient operating condition range, the influence of temperature drift on the offset of device loss parameters is analyzed. Taking the first gradient operating condition range as an example, the average junction temperature and thermal resistance characteristic parameters of this range are obtained. The temperature coefficient curve of on-state resistance is read from the power device's technical data sheet. This curve shows the relative rate of change of on-state resistance with junction temperature, expressed as a percentage per degree Celsius. The average junction temperature is substituted into the on-state resistance temperature coefficient curve to calculate the percentage change in on-state resistance relative to a room temperature reference value of 25 degrees Celsius. This percentage is multiplied by the measured average on-state resistance across all time segments within this range to obtain the on-state resistance offset caused by temperature drift. The temperature coefficient curve of switching loss is read from the power device's technical data sheet. This curve shows the relative rate of change of switching loss with junction temperature. The average junction temperature is substituted into the switching loss temperature coefficient curve to calculate the percentage change in switching loss relative to a room temperature reference value. This percentage is multiplied by the measured average switching loss across all time segments within this range to obtain the switching loss offset caused by temperature drift. Substituting the on-state resistance offset into the conduction loss calculation formula yields the conduction loss deviation. The switching loss offset is then used as the switching loss deviation. The same calculations for on-state resistance offset, switching loss offset, conduction loss deviation, and switching loss deviation are repeated for the second to last gradient operating condition intervals, resulting in a set of conduction loss and switching loss deviations for each interval. The deviations from all gradient operating condition intervals are arranged by interval number to form a dataset of deviation components required for device loss compensation. This dataset quantifies the quantitative impact of temperature drift on loss parameters under different temperature conditions.

[0065] The deviation component that needs to be compensated for device loss is mapped and coupled to the basic correlation curve to complete the curve deviation correction, thereby eliminating the operating condition-loss correlation deviation caused by temperature drift and obtaining the corrected nonlinear correlation curve.

[0066] In this embodiment, based on the dataset of deviation components requiring compensation for device losses, the deviation components are mapped and coupled to the fitted basic correlation curve. Taking conduction loss as an example, a basic correlation curve for conduction loss is obtained. This curve takes the effective value of current, on-state resistance, and maximum junction temperature as inputs and outputs the basic predicted value of conduction loss. The calculated conduction loss deviation for each gradient operating condition range is obtained. Using the maximum junction temperature as the independent variable and the conduction loss deviation as the dependent variable, a deviation interpolation function is constructed. This interpolation function uses cubic spline interpolation, taking the average junction temperature value for all gradient operating condition ranges as interpolation nodes and the corresponding conduction loss deviation as interpolation node values. A cubic polynomial is fitted between two adjacent nodes to ensure that the function values ​​at the nodes are continuous and the first derivative is continuous. For any given maximum junction temperature input, the corresponding conduction loss deviation is calculated using this cubic spline interpolation function. The output of the basic correlation curve for conduction loss is added to this deviation to obtain the corrected predicted value of conduction loss. Taking switching loss as an example, the basic correlation curve of switching loss and the switching loss deviation in each gradient operating condition range are obtained. Another cubic spline interpolation function is constructed, with the maximum junction temperature as the independent variable and the switching loss deviation as the dependent variable. For any given maximum junction temperature input, the switching loss deviation is calculated using the interpolation function. The output of the basic correlation curve is added to this deviation to obtain the corrected predicted switching loss value. The corrected predicted conduction loss value and the predicted switching loss value are used as new outputs to replace the original output in the basic correlation curve, completing the curve deviation correction. The two corrected nonlinear correlation curves eliminate the operating condition-loss correlation deviation caused by temperature drift, accurately reflecting the true mapping relationship between operating condition characteristics and loss parameters under different temperature conditions.

[0067] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0068] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily 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 invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A power device lifetime assessment method based on multi-condition operating state clustering, characterized in that, Includes the following steps: Step S1: Collect multi-dimensional raw operating condition data of power devices during continuous operation throughout their entire life cycle, filter core operating condition characteristic parameters based on device operating mechanism, and construct a multi-dimensional operating condition characteristic dataset; Step S2: Perform differential mining of operating conditions on the multi-dimensional operating condition feature dataset, extract the dynamic evolution correlation features of each feature parameter, and generate a multi-dimensional operating condition correlation feature set corresponding to the power device. Step S3: Construct an adaptive density clustering model based on the multi-dimensional operating condition association feature set, complete the clustering of the power device's full operating condition status, and generate various operating condition clusters, including steady-state operating condition clusters and transient operating condition clusters. Step S4: Based on the characteristic distribution patterns of various operating condition clusters, match the internal loss evolution mechanism of power devices under different operating conditions, and generate the device loss dynamic characteristic sequence corresponding to each operating condition cluster. Step S5: Based on the dynamic characteristic sequence of device loss of each operating condition cluster coupled with the device fatigue damage accumulation mechanism, construct a device fatigue damage evolution model coupled with multiple operating conditions, and complete the real-time fatigue damage quantification calculation of power devices based on the device fatigue damage evolution model to generate device dynamic damage accumulation data. Step S6: Combining the device dynamic damage accumulation data with the runtime proportion of various operating condition clusters, iteratively correct the operating condition coupling coefficient of the device fatigue damage evolution model, and extrapolate the remaining service life of the power device based on the corrected device fatigue damage evolution model, thereby completing the full life cycle assessment of the power device.

2. The power device lifetime assessment method based on multi-condition operating state clustering according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Synchronously collect electrical operation data, thermal operation data, environmental operation data and mechanical vibration operation data during the operation of power devices, and integrate them to generate a set of multi-source raw operation data corresponding to the power devices; Step S12: Based on the semiconductor conduction loss, switching loss and thermal fatigue failure mechanism of power devices, the feature validity of the original multi-source operation data set is identified, and redundant data dimensions that are not coupled with device lifetime degradation are eliminated. Step S13: Extract the real-time operation features, dynamic fluctuation features, and state change features from the remaining valid data dimensions to form the initial operating condition feature parameter set corresponding to the power device; Step S14: Based on the rated operating condition boundary conditions of the power device, perform boundary constraint calibration on the initial operating condition characteristic parameter set to eliminate the characteristic offset problem caused by the rated operating condition deviation. Step S15: The calibrated multi-dimensional working condition feature parameters are structurally integrated to construct a time-aligned, dimensionally unified multi-dimensional working condition feature dataset.

3. The power device lifetime assessment method based on multi-condition operating state clustering according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Based on the temporal distribution characteristics of the multidimensional working condition feature dataset, decompose the long-term evolution trend component and short-term dynamic fluctuation component of each working condition feature parameter. Step S22: Explore the static coupling correlation between different long-term evolution trend components and the dynamic linkage pattern between short-term dynamic fluctuation components to generate a two-way correlation matrix of working condition characteristics. Step S23: Based on the bidirectional correlation matrix of operating condition features, identify the core correlation feature combinations of the operating status of the dominant device in different operating periods; Step S24: Extract the state representation attributes and temporal evolution attributes of each core associated feature combination to form differentiated operating condition associated feature units; aggregate all differentiated operating condition associated feature units and integrate them to generate a multi-dimensional operating condition associated feature set covering power devices throughout the entire operating period.

4. The power device lifetime assessment method based on multi-condition operating state clustering according to claim 3, characterized in that, Step S22 includes the following steps: Time-series normalization mapping is performed on each feature parameter in the multidimensional working condition feature dataset to eliminate the dimensional differences between different feature dimensions and generate a set of time-series feature sequences of equal dimensions. Based on equal-dimensional time-series feature sequence groups, a nonlinear correlation fitting algorithm is used to calculate the static correlation degree between any two working condition feature parameters and construct a feature static correlation matrix. The short-term fluctuation components of characteristic parameters under various operating conditions are processed by sliding time-series slicing to obtain the characteristic fluctuation amplitude and fluctuation rate in different time-series slices; the dynamic linkage response relationship of each characteristic fluctuation parameter in different time-series slices is calculated to generate a time-series dynamic correlation sequence set. By integrating the static correlation matrix of features and the time-series dynamic correlation sequence set, a bidirectional correlation matrix of working condition features is generated.

5. The power device lifetime assessment method based on multi-condition operating state clustering according to claim 4, characterized in that, The calculation of the dynamic linkage response relationship of each characteristic fluctuation parameter within different time slices includes the following steps: The characteristic fluctuation amplitude and fluctuation rate within each time slice are classified into stable fluctuation state and sudden fluctuation state. For a stable fluctuation state, calculate the synchronous change coefficient between the fluctuations of different characteristic parameters to characterize the steady-state linkage relationship; For abrupt fluctuations, the difference between the fluctuation response delay and the fluctuation amplitude of each characteristic parameter is extracted to characterize the transient linkage relationship. By combining steady-state and transient linkage relationships, a feature dynamic linkage model within a single time slice is constructed; the feature dynamic linkage models of all time slices are traversed and integrated to generate a set of time-series dynamic association sequences.

6. The power device lifetime assessment method based on multi-condition operating state clustering according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Based on the multi-dimensional working condition association feature set, extract the spatial distribution density and temporal aggregation characteristics of each working condition feature, and construct a spatial distribution model of working condition features; Step S32: Based on the working condition feature space distribution model, adaptively identify high-density clustered regions and low-density discrete regions in the feature space, and divide the working condition clustering core domain and the working condition discrete domain. Step S33: Based on the feature aggregation attributes of the clustering core domain, complete the merging of similar steady-state working conditions, and based on the feature difference attributes of the discrete domain of the working conditions, complete the differentiated splitting of transient working conditions. Step S34: Remove isolated abnormal operating conditions in the feature space to avoid interference from accidental abnormal operating conditions on the clustering results; Step S35: Complete the clustering and division of the full operating conditions, and generate various operating condition clusters, including steady-state operating condition clusters and transient operating condition clusters.

7. The power device lifetime assessment method based on multi-condition operating state clustering according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Analyze the characteristic distribution attributes of various operating condition clusters and match the conduction and switching states of power devices under the corresponding operating conditions; Step S42: Based on the semiconductor physical loss mechanism of the device, establish a corresponding correlation model between the operating condition cluster characteristics and the device conduction loss and switching loss; Step S43: Calculate the real-time power loss and power loss accumulation rate of devices under various operating conditions using the correlation model; Step S44: Combine the time-series operating characteristics of each operating condition cluster to generate the time-series dynamic evolution law of loss parameters; Step S45: Integrate the loss time-series evolution patterns of various operating condition clusters to generate the device loss dynamic characteristic sequence corresponding to each operating condition cluster.

8. The power device lifetime assessment method based on multi-condition operating state clustering according to claim 7, characterized in that, Step S42 includes the following steps: Extract the core operational characteristic parameters of clusters under various operating conditions and construct a cluster feature representation vector; Based on the semiconductor carrier transport characteristics of power devices, the core influencing factors of device loss under different operating conditions are determined. Establish a mapping relationship between the characteristic vector of the working condition cluster and the core influencing factors of loss; The nonlinear correlation curves between the operating condition characteristics and the conduction loss and switching loss are fitted based on the mapping relationship; A correlation model between operating condition cluster characteristics and device losses is constructed based on nonlinear correlation curves.

9. The power device lifetime assessment method based on multi-condition operating state clustering according to claim 8, characterized in that, The method of fitting the nonlinear correlation curves between operating condition characteristics and conduction loss and switching loss based on the mapping relationship includes the following steps: Based on the characteristic representation vector of the working condition cluster, the gradient working condition operating range is divided; The measured loss parameters of power devices within the operating range of each gradient chemical condition are obtained to form a sample set corresponding to the operating condition and loss. Based on the corresponding sample set of operating conditions and losses, a nonlinear fitting algorithm is used to fit the basic correlation curve; The deviation component of the basic correlation curve is corrected by incorporating the temperature drift characteristics of the device carriers. The nonlinear correlation curves between the operating condition characteristics and conduction loss and switching loss are mapped and fitted based on the corrected deviation components and the basic correlation curves.

10. The power device lifetime assessment method based on multi-condition operating state clustering according to claim 9, characterized in that, The deviation component of the basic correlation curve for correcting the carrier temperature drift characteristics of the combined device includes the following steps: Extract the device thermal state characteristics corresponding to each gradient operating condition range to obtain the real-time temperature distribution characteristics of charge carriers; Based on the real-time temperature distribution characteristics of charge carriers, the influence of temperature drift on the offset of device loss parameters is analyzed, and the loss deviation caused by temperature drift is quantified to construct the deviation component that needs to be compensated for device loss. The deviation component that needs to be compensated for device loss is mapped and coupled to the basic correlation curve to complete the curve deviation correction, thereby eliminating the operating condition-loss correlation deviation caused by temperature drift and obtaining the corrected nonlinear correlation curve.

Images