A method for screening IGBT parallel devices based on clustering and artificial bee colony algorithms

Abstract

This invention provides a method for screening parallel IGBT devices based on clustering and artificial bee colony algorithms. The method includes: clustering electrical parameters based on the number of parallel IGBTs; constructing a fitness function with the intra-group current imbalance rate as the objective; and finally, using the artificial bee colony algorithm to collaboratively optimize the electrical parameters based on the clusters and the fitness function, selecting the parallel chip combination with the minimum intra-group current imbalance rate. The clustering algorithm provides a high-quality initial solution for the artificial bee colony algorithm, and the combination of Gaussian mutation and truncation mutation strategies improves the algorithm's convergence efficiency, enabling it to quickly approach the global optimum in large-scale chip screening. This optimizes the threshold voltage and on-state voltage drop, suppressing transient and steady-state current imbalances in parallel IGBTs, significantly reducing both transient current deviation and steady-state current imbalance rate. This method significantly improves the current sharing performance of the devices.

Description

Technical Field

[0001] This invention generally relates to the field of IGBT chip technology, and specifically to a method for screening IGBT parallel devices based on clustering algorithms and artificial bee colony algorithms. Background Technology

[0002] With the development of renewable energy and rail transportation, the current capacity of a single IGBT chip has reached its limit. Parallel connection of multiple devices to increase current capacity has become the mainstream solution for high-power converters. However, due to the inherent variations in wafer fabrication and packaging processes, the threshold voltage and on-state voltage drop of IGBTs of the same model will inevitably differ. The threshold voltage determines the switching transient, and its difference can lead to inconsistent switching timing, causing some chips to be subjected to excessive stress. The on-state voltage drop determines the steady-state current shunt, and its difference can cause uneven current distribution, leading to localized heat concentration. These two imbalances are coupled together, and in severe cases, can lead to device overheating or even damage, becoming a fundamental obstacle to system reliability.

[0003] To suppress current imbalance, the industry commonly employs device screening and matching methods, but existing technologies have significant limitations. Screening based on a single parameter can only address one aspect at the expense of others, failing to achieve synergistic optimization of transient and steady-state performance. While multi-level classification based on simple rules considers two parameters, it faces the challenge of an excessive number of combinations in large-scale chip assembly. This is because existing algorithms only focus on electrical parameters, ignoring process information such as wafer batch and packaging station, leading to a disconnect between theoretical grouping and practical application. This results in algorithms that can only find feasible solutions, but cannot approximate the globally optimal solution. Summary of the Invention

[0004] In view of the above-mentioned defects or deficiencies in the existing technology, it is desirable to provide a method for screening IGBT parallel devices based on clustering algorithm and artificial bee colony algorithm.

[0005] This invention provides a method for screening IGBT parallel devices based on clustering algorithms and artificial bee colony algorithms, comprising: S1: Obtain the electrical parameters and process information of all IGBT chips to be screened; the electrical parameters include: threshold voltage and on-state voltage drop; the process information includes: wafer batch and packaging station information; S2: Obtain the number of IGBT chips required for parallel connection; S3: Based on the number of IGBT chips, the electrical parameters of the IGBT chips to be screened with the same process information are clustered to obtain multiple clusters; S4: Set a fitness function based on the number of IGBT chips; the fitness function is used to calculate the intra-group current imbalance rate of multiple parallel IGBT chips; S5: Based on the cluster and the fitness function, the electrical parameters are optimized using the artificial bee colony algorithm to obtain the optimal electrical parameters; the group of IGBT chips with the optimal electrical parameters has the minimum current imbalance rate. S6: The IGBT chips with the same process information and the best electrical parameters are selected as the final screening result.

[0006] According to the technical solution provided by this invention, the electrical parameters and process information of all IGBT chips to be screened are obtained, including: Collect the initial threshold voltage and initial on-state voltage drop of all IGBT chips to be screened, as well as wafer batch and packaging station information; The initial threshold voltage and initial on-state voltage drop were standardized to obtain electrical parameters; wafer batch and packaging station information were used as process information.

[0007] According to the technical solution provided by the present invention, based on the number of IGBT chips, the electrical parameters of IGBT chips to be screened with the same process information are clustered to obtain multiple clusters, including: Extract the distribution characteristics of the threshold voltage based on the electrical parameters that have the same process information; Set a clustering objective function; the clustering objective function is used to calculate the sum of differences in all electrical parameters within the current cluster; The first number of clusters is determined based on the number of IGBT chips and the number of IGBT chips to be screened. Based on the clustering objective function and the distribution characteristics, the electrical parameters are clustered to obtain multiple clusters.

[0008] According to the technical solution provided by the present invention, a fitness function is set according to the number of IGBT chips, including: If the number of IGBT chips to be screened is divisible by the total number of IGBT chips, then the fitness function is set as follows:

[0009] If the number of IGBT chips to be screened is not divisible by the total number of IGBT chips, then the fitness function is set as follows:

[0010] in, fitness G is the fitness value. i For the i-th group, and The weighting coefficients for the influence of current are summed to equal 1. Let be the normalized threshold voltage mean of the i-th group. Let λ be the normalized mean on-state voltage drop of the i-th group, and λ be the batch constraint penalty coefficient. This is a batch consistency judgment function. When the number of differences in the batch tags of IGBT chips within the i-th group exceeds the difference threshold, It equals 1, otherwise it equals 0.

[0011] According to the technical solution provided by the present invention, based on the cluster and the fitness function, the artificial bee colony algorithm is used to optimize the electrical parameters to obtain the optimal electrical parameters, including: S5-1: Based on the electrical parameters of all IGBT chips to be screened, initially set the artificial bee colony algorithm, including: setting the initial honey source, maximum number of iterations, and termination condition according to the threshold voltage and on-state voltage drop corresponding to the clusters; S5-2: Gaussian mutation is performed on the threshold voltage of the IGBT chip to be screened during the iteration process of the artificial bee colony algorithm, and truncation mutation is performed on the on-state voltage drop to obtain the iterative threshold voltage and iterative on-state voltage drop respectively. S5-3: Based on the number of IGBT chips, substitute the iteration threshold voltage and the iteration on-state voltage drop into the fitness function to calculate the iteration fitness value; S5-4: If the iteration fitness value meets the termination condition, then the current iteration threshold voltage and iteration on-state voltage drop are taken as the optimal electrical parameters.

[0012] According to the technical solution provided by the present invention, step S5-4 further includes: if the termination condition is not met, then the following steps are performed: S5-5: Based on the iterative fitness value, perform a local search within the cluster; S5-6: Update the current honey source; the current honey source is initially the initial honey source, and in subsequent iterations it is the honey source obtained in the previous iteration; S5-7: Repeat steps S5-2 to S5-7 until the termination condition is met.

[0013] According to the technical solution provided by the present invention, a local search is performed within the cluster based on the iterative fitness value, including: Select the current nectar source for the observed bees according to the fitness probability formula; The fitness probability formula is:

[0014] in, For the first The iterative fitness value of each hired bee corresponding to a nectar source. For the number of hired bees, For the first The probability of a bee selecting a nectar source is positively correlated with the iterative fitness value of the nectar source.

[0015] According to the technical solution provided by the present invention, updating the current honey source includes: If the fitness value of the nectar source corresponding to the hired bee does not decrease after a preset number of iterations, then the scout bee will randomly generate a new nectar source location within the cluster corresponding to the nectar source to replace the nectar source.

[0016] According to the technical solution provided by the present invention, the termination conditions include: The rate of decrease in the iterative fitness value is less than the decrease rate threshold, and satisfies: The current imbalance rate corresponding to the iteration threshold voltage and the iteration on-state voltage drop is less than the current imbalance rate threshold or reaches the maximum number of iterations.

[0017] The beneficial effects of this invention are as follows: To address the technical problems in existing IGBT parallel device screening methods, such as the inability of single-parameter screening to consider both transient and steady-state performance, low search efficiency and susceptibility to local optima in large-scale combinations due to simple rule-based multi-level classification, and the disconnect between theoretical grouping and practical application caused by ignoring process information such as wafer batch and packaging station, this invention adopts the following technical solution: First, the threshold voltage, on-state voltage drop, wafer batch, and packaging station information of the IGBT chips to be screened are collected; based on the number of parallel connections, the electrical parameters of the IGBT chips to be screened with the same process information are clustered to obtain multiple clusters; a fitness function is constructed with the current imbalance rate within the group as the objective; finally, based on the clusters and the fitness function, the artificial bee colony algorithm is used to collaboratively optimize the electrical parameters to screen the parallel chip combination with the minimum current imbalance rate within the group.

[0018] By simultaneously optimizing the threshold voltage and on-state voltage drop, the transient and steady-state current imbalance of parallel IGBTs was synergistically suppressed, reducing the transient current deviation from 11.3% to 4.82% and the steady-state current imbalance rate from 8.68% to 1.96%, significantly improving the current sharing performance of the devices. A clustering algorithm was used to provide high-quality initial solutions for the artificial bee colony algorithm. Combined with customized Gaussian mutation and truncation mutation strategies, the algorithm's convergence efficiency was improved by more than 42%, and it can still quickly approach the global optimum in large-scale chip screening. A batch process constraint penalty term was introduced into the fitness function to ensure that the selected grouping schemes not only have high electrical parameter matching, but also consistent wafer batch and packaging station information, avoiding the problem of "good laboratory results but poor engineering applications." This can directly guide the industrial batch screening of parallel IGBT devices in high-power converters. Attached Figure Description

[0019] 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 flowchart illustrating a method for screening IGBT parallel devices based on clustering and artificial bee colony algorithms. Figure 2 This is a distribution diagram of threshold voltage and on-state voltage drop under conventional screening methods; Figure 3 This is a distribution diagram of threshold voltage and on-state voltage drop under the screening method of this embodiment; Figure 4 The current waveforms of parallel devices under conventional screening methods are shown. Figure 5 The current waveform of the parallel devices is shown in the screening method of this embodiment. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] refer to Figure 1 This invention provides a method for screening IGBT parallel devices based on clustering algorithms and artificial bee colony algorithms, comprising: S1: Obtain the electrical parameters and process information of all IGBT chips to be screened; the electrical parameters include: threshold voltage V th and conduction voltage drop V CE(on) The process information includes: wafer batch and packaging station information; Step S1 specifically includes: Collect the initial threshold voltage and initial on-state voltage drop of all IGBT chips to be screened, as well as wafer batch and packaging station information; after obtaining all the data, establish a database for storage and retrieval; During the specific data acquisition process, a symmetrical dual-pulse test platform was used for data acquisition. and The specific parameters and test conditions are: collector current. DC bus voltage Grid emitter voltage This allows for a direct reflection of the parameter discreteness of the original device, and enables clear observation. normal distribution and The skewed distribution pattern.

[0023] The initial threshold voltage and initial on-state voltage drop were standardized to obtain electrical parameters; wafer batch and packaging station information were used as process information.

[0024] The purpose of standardization is to eliminate dimensional differences between different parameters. The specific formula is:

[0025] in, The global minimum value of the electrical parameters among all IGBT chips to be screened. This represents the global maximum value of the electrical parameters among all IGBT chips to be screened. For the first The original values ​​of the electrical parameters of each IGBT chip. For the first Standardized values ​​of electrical parameters for each IGBT chip.

[0026] S2: Obtain the number of IGBT chips required for parallel connection; the specific number of IGBT chips is set according to actual needs. The specific value of m, the number of IGBT chips required for parallel connection, is generally between 2 and 200.

[0027] S3: Based on the number of IGBT chips, the electrical parameters of the IGBT chips to be screened with the same process information are clustered to obtain multiple clusters, including: The distribution characteristics of the threshold voltage are extracted based on the electrical parameters with identical process information; the distribution characteristics specifically refer to the distribution patterns of multiple threshold voltages. Distribution patterns include normal distribution, skewed distribution, etc.

[0028] Set a clustering objective function; the clustering objective function is used to calculate the sum of differences in all electrical parameters within the current cluster; The first number of clusters is determined based on the number of IGBT chips and the number of IGBT chips to be screened. Specifically, the clustering objective function is:

[0029] in, C is the function value of the clustering objective function, K is the number of clusters (i.e., the first number), and is equal to the ratio (rounded up) of the number of IGBT chips to be screened n to the required number of IGBT chips m. k For the k-th cluster, For the first A grouped standardized threshold voltage, μ k The standard value of the threshold voltage in the k-th cluster is given. The distance is Euclidean.

[0030] Based on the clustering objective function and the distribution characteristics, the electrical parameters are clustered (specifically using the K-means clustering method) to obtain multiple clusters (a first number), including: Q1: Randomly extract multiple initial clusters (equal to the ratio of the number of IGBT chips to be screened n to the required number of IGBT chips m). Q2: Assign multiple data points to the current iterative cluster in a random manner that satisfies the distribution characteristics; each data point corresponds to a combination of threshold voltage and on-state voltage drop; the current iterative cluster is initially the initial cluster, and the subsequent clusters are the iterative clusters obtained from the previous iteration; Q3: Calculate the data for each data point and substitute it into the clustering objective function to obtain the clustering objective function value; Q4: If the clustering objective function value is less than the clustering convergence threshold, or the number of iterations of K-means clustering is reached, then the current iterative cluster is taken as the final cluster; otherwise, repeat steps Q2 to Q4.

[0031] Specifically, the number of iterations for K-means clustering is Next, the clustering convergence threshold is... .

[0032] Specifically, steps Q1 to Q4 achieve high-quality honey source targeted initialization by deeply integrating the normal distribution characteristics of the IGBT threshold voltage with the K-means clustering algorithm. First, the number of clusters is dynamically set based on the number of parallel connections, decomposing the large-scale chip assembly problem into local optimizations within each cluster, thus ensuring a natural fit between the initial honey sources. The statistical distribution pattern effectively avoids the blindness of random initialization, ensuring that each cluster corresponds to one hired bee, making the initial nectar source naturally compatible with the IGBT chip. By analyzing the parameter distribution characteristics, the initial search range of the algorithm is narrowed down from the root cause; secondly, by iteratively optimizing the clustering objective function to the convergence threshold, the differences in the electrical parameters of the chips within the cluster are minimized, so that subsequent iterations can quickly approach the global optimum.

[0033] S4: Set a fitness function based on the number of IGBT chips; the fitness function is used to calculate the intra-group current imbalance rate of multiple parallel IGBT chips; Step S4 specifically includes: If the number of IGBT chips to be screened is divisible by the total number of IGBT chips, then the fitness function is set as follows:

[0034] If the number of IGBT chips to be screened is not divisible by the total number of IGBT chips, then the fitness function is set as follows:

[0035] in, fitness G is the fitness value. i For the i-th group (a group is a combination of multiple IGBT chips that need to be connected in parallel), and The weighting coefficients for the influence of current are summed to equal 1. Let be the normalized threshold voltage mean of the i-th group. Let λ be the normalized mean on-state voltage drop of the i-th group, and λ be the batch constraint penalty coefficient. This is a batch consistency judgment function. When the number of differences in the batch tags of IGBT chips in the i-th group exceeds the difference threshold T... B hour, It equals 1, otherwise it equals 0.

[0036] Specifically, , Batch constraint penalty coefficient Difference threshold .

[0037] In this embodiment, a smaller fitness value indicates a lower current imbalance rate and a better grouping effect of multiple IGBTs.

[0038] The case-specific fitness function in step S4 accurately maps the physical mechanism of IGBT parallel connection: by quantifying the coupling effect of threshold voltage and on-state voltage drop on current imbalance through weighted summation, eliminating dimensional interference by normalizing the ratio of range to mean, and introducing a batch penalty term to constrain process consistency, so that the optimization objective directly corresponds to the engineering current sharing index; at the same time, different calculation methods are set for whether the total number of chips is divisible, ensuring that the algorithm can output feasible grouping schemes that meet actual production needs in all scenarios, significantly improving the engineering adaptability of the selection and the suppression effect of current imbalance rate.

[0039] S5: Based on the cluster and the fitness function, the electrical parameters are optimized using the artificial bee colony algorithm to obtain the optimal electrical parameters; the group of IGBT chips with the optimal electrical parameters has the minimum current imbalance rate. The Artificial Bee Colony Algorithm (ASA) is a swarm intelligence optimization algorithm that simulates the foraging behavior of bees. It involves three types of bees—hired bees, observer bees, and scout bees—collaborating to complete the search task. In the selection of IGBT parallel devices, each nectar source represents a chip combination scheme, and the amount of nectar in the source corresponds to the fitness function value (i.e., the current imbalance rate). Hired bees are responsible for local searches within clusters; observer bees select high-quality nectar sources for focused development based on fitness probabilities; and scout bees randomly generate new solutions when nectar sources have not been updated for a long time, avoiding getting trapped in local optima. By introducing Gaussian mutation and truncation mutation strategies, the algorithm can take into account the physical constraints of threshold voltage and on-state voltage drop, achieving coordinated optimization of electrical parameters while ensuring search efficiency, and quickly approaching the global optimum.

[0040] Step S5 specifically includes: S5-1: Based on the electrical parameters of all IGBT chips to be screened, initially set the artificial bee colony algorithm, including: The initial honey source, maximum number of iterations, and termination condition are set according to the threshold voltage and on-state voltage drop corresponding to the clusters. In some implementations, the algorithm parameter settings are configured in ways such as: the number of IGBT chips required for parallel connection. Number of clusters Number of hired bees Observe the number of bees (1:1 ratio with mercenary bees), number of scout bees (This represents 12.5% ​​of the employed bees, within the design range of 10% to 20%); K-means clustering has a maximum of 50 iterations and a convergence threshold. ; Weight , Weight Batch penalty coefficient Batch difference threshold ;Recon bee iteration threshold Second; Two-stage (including S5 and local search) convergence parameters: , descent rate threshold Current imbalance rate threshold Maximum number of iterations The above data are for illustrative purposes only and can be adjusted according to actual circumstances or experience.

[0041] Furthermore, the search space is the search range corresponding to the electrical parameters of all IGBT chips to be screened; The positions of the threshold voltage and on-state voltage drop corresponding to the clusters within the search range are used as the initial honey sources; The termination conditions include: The rate of decrease in the iterative fitness value is less than the decrease rate threshold (set to 1% in this embodiment, and based on the fitness value corresponding to the global optimum), and satisfies: The current imbalance rate corresponding to the iteration threshold voltage and the iteration on-state voltage drop is less than the current imbalance rate threshold (set to 2% in this embodiment) or reaches the maximum number of iterations.

[0042] In some implementations, to avoid misjudgment by the algorithm, it is necessary to combine the results of multiple iterations to determine whether the termination condition has been met.

[0043] The specific method is as follows: Set the descent rate calculation window. Number of consecutive convergence checks If the above termination condition is met at least 3 out of 5 consecutive iterations during the iteration process, the algorithm is finally considered to have converged.

[0044] Step S5-1 provides a suitable V-shape for the artificial bee colony algorithm by directly mapping the clusters to the initial nectar source. th The high-quality initial solution with normal distribution characteristics fundamentally solves the problem of blindness in random initialization; at the same time, setting a 1:1 ratio of hired bees to observation bees and the proportion of scout bees ensures the balance between local search and global exploration; and combined with two-stage convergence parameters, the algorithm effectively avoids invalid iterations while ensuring optimization accuracy, significantly improving search efficiency and engineering adaptability.

[0045] S5-2: (S5-2 is performed in each iteration) Gaussian mutation is performed on the threshold voltage of the IGBT chip to be screened in the artificial bee colony algorithm iteration process, and truncation mutation is performed on the on-state voltage drop to obtain the iterative threshold voltage and iterative on-state voltage drop respectively. The formula for Gaussian mutation is:

[0046] in, The threshold voltage after Gaussian variation. The threshold voltage before Gaussian transformation. Standardize all chips to be screened The standard deviation of N is given by N(0,1), which is a standard normally distributed random number.

[0047] The formula for truncation variation is:

[0048] in, To obtain the largest value, To minimize the impact, To cut off the mutated conduction voltage drop. To cut off the conduction voltage drop before the mutation, Standardize all chips to be screened standard deviation For standardization The global minimum value, For standardization The global maximum value.

[0049] In step S5-2, the threshold voltage adopts Gaussian variation and the on-state voltage drop adopts truncation variation, which is designed based on the essential differences between the two in terms of physical meaning and distribution characteristics.

[0050] The threshold voltage mainly determines the switching transient behavior of IGBT chips. During wafer fabrication and packaging, it is affected by random process fluctuations and usually exhibits a normal distribution (i.e., Gaussian distribution). Therefore, Gaussian mutation is adopted. By superimposing random perturbations that follow a standard normal distribution, while maintaining population diversity, a fine search is performed in the parameter concentration region, thereby effectively optimizing the transient current sharing performance.

[0051] The on-state voltage drop determines the steady-state shunt capability of the IGBT chip. Its distribution is often skewed and has clear physical boundaries. Directly using Gaussian mutation can easily generate invalid parameters that exceed the actual process capabilities. Truncation mutation, based on standard Gaussian perturbation, forces the mutation result to be between the global minimum and maximum values ​​through a maximization / minimization operation. This ensures that candidate solutions always correspond to realistically manufacturable chip parameters, guaranteeing the engineering feasibility and convergence stability of the search process. These two differentiated mutation strategies work together to achieve synergistic optimization of transient and steady-state performance.

[0052] The above method introduces random perturbations that follow a standard normal distribution for local search, which can achieve fine mining of transient characteristic parameters while maintaining population diversity, and effectively improve the search efficiency of the algorithm in the parameter set region.

[0053] The truncation operation ensures that the mutated parameters are always within the range allowed by the actual process, avoids generating invalid parameter combinations, and guarantees the engineering feasibility and convergence stability of the search process.

[0054] S5-3: Based on the number of IGBT chips, substitute the iteration threshold voltage and the iteration on-state voltage drop into the fitness function to calculate the iteration fitness value; S5-4: If the iteration fitness value meets the termination condition, then the current iteration threshold voltage and iteration on-state voltage drop are taken as the optimal electrical parameters.

[0055] If the termination condition is not met, proceed with the following steps: S5-5: Based on the iterative fitness value, perform a local search within the cluster; Specifically, the current nectar source for the observed bees is selected according to the fitness probability formula; The fitness probability formula is:

[0056] in, For the first The iterative fitness value of each hired bee corresponding to a nectar source. For the number of hired bees, For the first The probability of a bee selecting a nectar source is positively correlated with the iterative fitness value of the nectar source.

[0057] In step S5-5, the bees are observed prioritizing the selection of high-quality nectar sources for localized development based on a fitness probability formula. This formula calculates the selection probability based on the fitness value of the nectar source, thus tilting the search resources towards IGBT chip combinations with lower current imbalance rates, achieving adaptive resource allocation.

[0058] The selected honey source will be locally searched within its cluster neighborhood. This mechanism effectively balances global exploration and local mining capabilities, ensuring that the global optimal solution is quickly approached in large-scale chip screening and significantly improving current sharing performance.

[0059] Step S5-5 guides the observation bees to prioritize high-quality nectar sources for in-depth searching using the fitness probability formula, thereby achieving adaptive allocation of search resources and significantly improving local mining capabilities while maintaining population diversity.

[0060] S5-6: Update the current honey source; the current honey source is initially the initial honey source, and in subsequent iterations it is the honey source obtained in the previous iteration; Furthermore, update the current honey source, including: If the fitness value of the nectar source corresponding to the hired bee does not decrease after a preset number of iterations, the nectar source is determined to be "exhausted". The scout bee then randomly generates a new nectar source location within the cluster corresponding to the nectar source to replace the nectar source, in order to maintain population diversity while avoiding ineffective searches.

[0061] In this embodiment, the preset number of times Second-rate.

[0062] This mechanism can effectively break the deadlock of local optima and maintain population diversity, while also avoiding ineffective searches and preventing the algorithm from jumping out of the locked high-quality parameter range, thus balancing global exploration capability and search efficiency.

[0063] S5-7: Repeat steps S5-2 to S5-7 until the termination condition is met.

[0064] The mechanism described above for determining whether the termination condition is met abandons the traditional method of fixing the number of iterations. This not only ensures the optimization accuracy of the algorithm but also avoids redundant iterations, further improving the efficiency of the algorithm in engineering applications.

[0065] S6: The IGBT chips with the same process information and the best electrical parameters are selected as the final screening result.

[0066] This invention achieves accurate selection of IGBT parallel devices by deeply integrating K-means clustering and a customized artificial bee colony algorithm. The clustering algorithm initializes honey sources based on the normal distribution characteristics of threshold voltage, providing a high-quality solution set for subsequent optimization. A case-specific fitness function collaboratively optimizes the threshold voltage and on-state voltage drop, and batch process constraints are introduced to ensure that the grouping scheme has both theoretical matching and engineering adaptability. Differential Gaussian mutation and truncated mutation are adapted to the distribution patterns of the parameters, combined with the probability selection of observation bees and the update mechanism within the scout bee cluster, effectively balancing local search and global exploration capabilities. A two-stage convergence strategy avoids ineffective iterations.

[0067] The experiment was designed and four IGBT chips selected by this method were assembled into a parallel module. Under the test conditions of collector current of 185A, DC bus voltage of 600V and gate emitter voltage of 15V, transient current deviation was measured by a dual-pulse test platform, and steady-state current distribution was recorded by steady-state thermal imaging and current probe.

[0068] refer to Figure 2 and Figure 4 Following conventional screening methods, the distribution of threshold voltage and on-state voltage drop is relatively dispersed, and the current (Curent) changes over time in the waveform diagram of multiple IGBT chips (including...) I d1 , I d2 , I d3 , I d4 There are significant differences.

[0069] refer to Figure 3 and Figure 5 According to the screening method in this embodiment, the distribution of threshold voltage and on-state voltage drop is relatively concentrated, and the current (Curent) changes with time in the waveform diagram of multiple IGBT chips (including...) I d1 , I d2 , I d3 , I d4 There was no significant difference.

[0070] Experimental results show that the transient current deviation decreased from 11.3% before screening to 4.82%, the steady-state current imbalance rate decreased from 8.68% to 1.96%, and the number of algorithm iterations was reduced by more than 42% compared with the standard artificial bee colony algorithm. This method can still stably output a grouping scheme that meets the current sharing requirements in chip pools of 100 or more, verifying its industrial applicability and robustness.

[0071] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the specific combination of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A method for screening IGBT parallel devices based on a clustering algorithm and an artificial bee colony algorithm, characterized in that, include: S1: Obtain the electrical parameters and process information of all IGBT chips to be screened; The electrical parameters include: threshold voltage and on-state voltage drop; the process information includes: wafer batch and packaging station information; S2: Obtain the number of IGBT chips required for parallel connection; S3: Based on the number of IGBT chips, the electrical parameters of the IGBT chips to be screened with the same process information are clustered to obtain multiple clusters; S4: Set a fitness function based on the number of IGBT chips; the fitness function is used to calculate the intra-group current imbalance rate of multiple parallel IGBT chips; S5: Based on the cluster and the fitness function, the electrical parameters are optimized using the artificial bee colony algorithm to obtain the optimal electrical parameters; the group of IGBT chips with the optimal electrical parameters has the minimum current imbalance rate. S6: The IGBT chips with the same process information and the best electrical parameters are selected as the final screening result.

2. The IGBT parallel device screening method based on clustering algorithm and artificial bee colony algorithm according to claim 1, characterized in that, Obtain the electrical parameters and process information of all IGBT chips to be screened, including: Collect the initial threshold voltage and initial on-state voltage drop of all IGBT chips to be screened, as well as wafer batch and packaging station information; The initial threshold voltage and initial on-state voltage drop were standardized to obtain electrical parameters; wafer batch and packaging station information were used as process information. 3.The IGBT parallel device screening method based on the clustering algorithm and the artificial bee colony algorithm of claim 1, wherein, Based on the number of IGBT chips, the electrical parameters of the IGBT chips to be screened with the same process information are clustered to obtain multiple clusters, including: Extract the distribution characteristics of the threshold voltage based on the electrical parameters that have the same process information; Set a clustering objective function; the clustering objective function is used to calculate the sum of differences in all electrical parameters within the current cluster; The first number of clusters is determined based on the number of IGBT chips and the number of IGBT chips to be screened. Based on the clustering objective function and the distribution characteristics, the electrical parameters are clustered to obtain a first number of clusters.

4. The IGBT parallel device screening method based on the clustering algorithm and the artificial bee colony algorithm according to claim 1, characterized in that, The fitness function is set according to the number of IGBT chips, including: If the number of IGBT chips to be screened is divisible by the total number of IGBT chips, then the fitness function is set as follows: If the number of IGBT chips to be screened is not divisible by the total number of IGBT chips, then the fitness function is set as follows: in, fitness G is the fitness value. i For the i-th group, and The weighting coefficients for the influence of current are summed to equal 1. Let be the normalized threshold voltage mean of the i-th group. Let λ be the normalized mean on-state voltage drop of the i-th group, and λ be the batch constraint penalty coefficient. This is a batch consistency judgment function. When the number of differences in the batch tags of IGBT chips within the i-th group exceeds the difference threshold, It equals 1, otherwise it equals 0.

5. The IGBT parallel device screening method based on the clustering algorithm and the artificial bee colony algorithm according to claim 4, characterized in that, Based on the clusters and the fitness function, the electrical parameters are optimized using the artificial bee colony algorithm to obtain the optimal electrical parameters, including: S5-1: Based on the electrical parameters of all IGBT chips to be screened, initially set the artificial bee colony algorithm, including: setting the initial honey source, maximum number of iterations, and termination condition according to the threshold voltage and on-state voltage drop corresponding to the clusters; S5-2: Gaussian mutation is performed on the threshold voltage of the IGBT chip to be screened during the iteration process of the artificial bee colony algorithm, and truncation mutation is performed on the on-state voltage drop to obtain the iterative threshold voltage and iterative on-state voltage drop respectively. S5-3: Based on the number of IGBT chips, substitute the iteration threshold voltage and the iteration on-state voltage drop into the fitness function to calculate the iteration fitness value; S5-4: If the iteration fitness value meets the termination condition, then the current iteration threshold voltage and iteration on-state voltage drop are taken as the optimal electrical parameters.

6. The IGBT parallel device screening method based on the clustering algorithm and the artificial bee colony algorithm according to claim 5, characterized in that, Step S5-4 further includes: if the termination condition is not met, then the following steps are performed: S5-5: Based on the iterative fitness value, perform a local search within the cluster; S5-6: Update the current honey source; the current honey source is initially the initial honey source, and in subsequent iterations it is the honey source obtained in the previous iteration; S5-7: Repeat steps S5-2 to S5-7 until the termination condition is met.

7. The IGBT parallel device screening method based on the clustering algorithm and the artificial bee colony algorithm according to claim 6, characterized in that, Based on the iterative fitness value, a local search is performed within the cluster, including: Select the current nectar source for the observed bees according to the fitness probability formula; The fitness probability formula is: in, For the first The iterative fitness value of each hired bee corresponding to a nectar source. For the number of hired bees, For the first The probability of a bee selecting a nectar source is positively correlated with the iterative fitness value of the nectar source.

8. The IGBT parallel device screening method based on the clustering algorithm and the artificial bee colony algorithm according to claim 6, characterized in that, Update the current honey source, including: If the fitness value of the nectar source corresponding to the hired bee does not decrease after a preset number of iterations, then the scout bee will randomly generate a new nectar source location within the cluster corresponding to the nectar source to replace the nectar source.

9. The IGBT parallel device screening method based on clustering algorithm and artificial bee colony algorithm according to claim 5, characterized in that, The termination conditions include: The rate of decrease in the iterative fitness value is less than the decrease rate threshold, and satisfies: The current imbalance rate corresponding to the iteration threshold voltage and the iteration on-state voltage drop is less than the current imbalance rate threshold or reaches the maximum number of iterations.

Images