GaN chip gate fault detection method and system based on multi-source feature fusion

Abstract

The application relates to the technical field of chip fault detection, in particular to a GaN chip gate fault detection method and system based on multi-source feature fusion, which comprises the following steps: applying a multi-pulse excitation to a GaN chip, collecting pulse response and environmental noise signals, removing abnormal pulses in the pulse response based on the environmental noise signals to obtain effective response segments; constructing a transient coefficient set and a steady-state index set; calculating the central tendency and dispersion degree of the transient coefficient set and the steady-state index set, and mapping the central tendency and dispersion degree into a transient aggregation index and a steady-state aggregation index; generating a defect quantification risk index according to the product of the calculated coupling factor, the transient aggregation index and the steady-state aggregation index; and determining whether the GaN chip has a gate defect according to the defect quantification risk index and a generated adaptive determination threshold. Through cross-pulse aggregation quantification and physical linkage fusion, and in combination with an adaptive determination threshold of the environment and temperature, the weak defect detection sensitivity is improved.

Description

Technical Field

[0001] This application relates to the technical field of chip fault detection, and in particular to a method and system for detecting gate faults in GaN chips based on multi-source feature fusion. Background Technology

[0002] In wafer-level mass production testing of gallium nitride (GaN) chips, automated test equipment (ATE) typically applies an excitation pulse to the gate of the device under test (DUT) and simultaneously acquires the gate leakage current response at the microampere level to determine the presence of early defects. Because the leakage current signal is extremely weak, strong electromagnetic interference from equipment such as motors and power supplies in the test environment can couple into the signal path as random glitches, resulting in a large amount of non-stationary noise superimposed on the acquired response sequence. Conventional denoising methods (such as fixed threshold filtering or moving smoothing) often obscure the transient spikes and steady-state fluctuations reflecting defects while suppressing background noise, thus destroying key diagnostic features. Furthermore, feature extraction methods relying solely on single-pulse response waveforms are limited by the low signal-to-noise ratio of a single pulse, making it difficult to reliably distinguish between interference and real defects, and are highly susceptible to misjudgments due to occasional interference pulses.

[0003] Most existing testing schemes fail to systematically utilize the repeatability of defect features across multi-pulse excitations, neglecting the physical characteristic that true gate degradation (such as trap-assisted tunneling) regularly reproduces transient distortion and increased steady-state fluctuations across multiple consecutive pulses. Furthermore, transient and steady-state features typically exhibit a positive correlation under this type of degradation mechanism, while existing criteria lack constraints on their physical consistency. When noise only causes an abnormal increase in one-sided features, false alarms are easily generated in the output results. In addition, the judgment thresholds are mostly fixed values, unable to dynamically adapt to fluctuations in the electromagnetic environment intensity of the testing station and drifts in the real-time wafer temperature. This leads to the incorrect rejection of qualified chips during periods of strong interference, while defective chips may be missed in quiet environments, severely impacting the stability of mass production yield.

[0004] Therefore, how to extract multi-pulse transient and steady-state features with defect repeatability indication capability from the extremely weak gate leakage current contaminated by strong electromagnetic interference, and how to fuse the two based on physical failure priors for noise resistance, while establishing a decision mechanism that adapts to environmental noise and temperature, are the technical problems that urgently need to be solved. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method and system for detecting gate faults in GaN chips based on multi-source feature fusion.

[0006] Firstly, this application provides a method for detecting gate faults in GaN chips based on multi-source feature fusion, employing the following technical solution: A gate fault detection method for GaN chips based on multi-source feature fusion includes the following steps: applying multi-pulse excitation to the GaN chip, acquiring pulse response and environmental noise signals, and removing abnormal pulses from the pulse response based on the environmental noise signal to obtain an effective response segment; removing DC from the effective response segment and dividing it into transient and steady-state segments, extracting the transient energy peak coefficient and steady-state fluctuation index, and constructing a transient coefficient set and a steady-state index set; calculating the central tendency and dispersion of the transient coefficient set and the steady-state index set, and mapping them to transient clustering index and steady-state clustering index; calculating the absolute value of the difference between the transient clustering index and the steady-state clustering index as the linkage coupling deviation, and generating a coupling factor that decreases as the linkage coupling deviation increases; using the coupling factor to weight the mean of the transient clustering index and the steady-state clustering index, and combining the product term of the transient clustering index and the steady-state clustering index to generate a defect quantification risk index; calculating a smoothed noise level estimate based on the environmental noise signal, and generating an adaptive judgment threshold based on the current temperature of the GaN chip; if the defect quantification risk index is greater than or equal to the adaptive judgment threshold, the GaN chip is determined to have a gate defect.

[0007] Optionally, the method for calculating the transient aggregation index includes: calculating the median of the transient coefficient set and the corrected absolute deviation of the median; dividing the corrected absolute deviation of the median by the sum of the median and a preset zero constant, and performing negative correlation normalization to obtain the transient aggregation index; similarly to calculating the transient aggregation index, the steady-state aggregation index is calculated.

[0008] Optionally, the step of obtaining a valid response segment by removing abnormal pulses from the pulse response based on the environmental noise signal includes the following steps: extracting a response segment corresponding to the duration of the pulse from the pulse response according to the trigger timestamp of each pulse; calculating the short-time energy value of the environmental noise signal within the same time window as the response segment; if the short-time energy value exceeds a preset upper limit for removing abnormal noise pulses, then marking the corresponding response segment as invalid and removing it, and using the response segments that are not removed as the valid response segments.

[0009] Optionally, the method for removing DC from the effective response segment and dividing it into transient and steady-state segments is as follows: extract the steady-state maintenance segment after the effective response segment ends and calculate the current time mean; subtract the current time mean from the entire effective response segment to eliminate baseline drift and obtain a zero-mean response sequence; determine the segmentation boundary based on the width of the excitation pulse and the preset transient duration; divide the preset time interval after the rising edge ends in the zero-mean response sequence into the transient segment and divide the remaining time interval into the steady-state segment.

[0010] Optionally, the extraction of the transient energy peak coefficient includes: calculating the maximum value of the absolute value of leakage current at all sampling points within the transient segment; calculating the root mean square of leakage current at all sampling points within the transient segment; and dividing the maximum value by the sum of the root mean square of leakage current and a preset zero constant to obtain the transient energy peak coefficient.

[0011] Optionally, the steady-state fluctuation index includes: calculating the absolute value of the standard deviation and the mean of all sampled values ​​within the steady-state range; dividing the standard deviation by the sum of the absolute value of the mean and a preset zero constant to obtain the steady-state fluctuation index; and replacing the denominator in calculating the steady-state fluctuation index with the current range normalized value in response to the absolute value of the mean being less than a preset minimum threshold.

[0012] Optionally, the defect quantification risk index is calculated as follows: calculate the arithmetic mean of the transient aggregation index and the steady-state aggregation index; multiply the arithmetic mean by the coupling factor to obtain the weighted mean term; calculate the product of the transient aggregation index and the steady-state aggregation index, and multiply the product by the preset risk fusion balance coefficient to obtain the product reinforcement term; add the weighted mean term and the product reinforcement term to generate the defect quantification risk index.

[0013] Secondly, this application provides a GaN chip gate fault detection system based on multi-source feature fusion, which adopts the following technical solution: the GaN chip gate fault detection system based on multi-source feature fusion includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the GaN chip gate fault detection method based on multi-source feature fusion described above is implemented.

[0014] This application has the following technical advantages: 1. By dividing the gate leakage current response of each pulse into transient and steady-state segments, and extracting the transient energy peak coefficient and steady-state fluctuation index as single-pulse features, a cross-pulse transient feature set and a steady-state feature set are constructed. The median absolute deviation is used to measure the set dispersion, and the transient clustering index and steady-state clustering index are generated through negative exponential mapping. This transforms the systematic reproduction of real defects under multiple excitations into a high-cluster numerical expression, while random electromagnetic interference is naturally suppressed due to its scattered distribution across pulses. Even if the signal-to-noise ratio of a single pulse is extremely low, it can still effectively widen the feature gap between defective chips and normal chips, and improve the detectability of weak early gate defects.

[0015] 2. Based on the physical prior that transient distortion and steady-state fluctuation are positively correlated in GaN HEMT gate degradation, the absolute value of the difference between the transient aggregation index and the steady-state aggregation index is defined as the linkage coupling deviation. The arithmetic mean of the two indices is dynamically modulated and then multiplied by the two indices and weighted to form a defect quantification risk index. The risk value will only increase significantly when both transient and steady-state features show high aggregation and their deviations are within the allowable range of physical linkage. This effectively suppresses noise-induced unilateral feature anomalies, thereby reducing the probability of misjudgment and improving detection specificity.

[0016] 3. By synchronously collecting environmental noise signals and implementing exponential moving averages, a smooth noise level estimate is obtained. At the same time, the real-time temperature of the wafer is acquired. An adaptive judgment threshold matching the interference intensity and temperature changes is generated based on a preset linear mapping relationship. The adaptive judgment threshold automatically rises when the electromagnetic environment deteriorates to suppress the false rejection of qualified chips, and decreases accordingly when the environment improves to prevent defects from being missed. This ensures that the probability of false judgment remains approximately constant under different test shifts and workstation conditions, effectively guaranteeing the stability of mass production yield and the robustness of the detection system. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for detecting gate faults in GaN chips based on multi-source feature fusion, according to an embodiment of this application. Detailed Implementation

[0018] This application discloses a GaN chip gate fault detection method based on multi-source feature fusion, referring to... Figure 1 This includes steps S1-S5: S1: Apply multi-pulse excitation to the GaN chip, collect the pulse response and environmental noise signal, and remove abnormal pulses in the pulse response based on the environmental noise signal to obtain the effective response segment.

[0019] In one embodiment, obtaining a valid response segment includes the steps of: extracting a response segment corresponding to the duration of the pulse from the pulse response based on the trigger timestamp of each pulse; calculating the short-time energy value of the environmental noise signal within the same time window as the response segment; if the short-time energy value exceeds a preset noise abnormal pulse rejection limit, the corresponding response segment is marked as invalid and rejected, and the response segments that are not rejected are taken as valid response segments.

[0020] Specifically, in GaN chip wafer-level automated testing, the automated test equipment (ATE) applies a series of excitation pulses to the gate of the device under test and simultaneously acquires two signals: one is the gate-drain current time-domain response sequence, denoted as the pulse response. Impulse response One path is a current signal, and the other is an ambient noise signal obtained through an open-circuit probe, denoted as the ambient noise signal. The ambient noise signal is a voltage signal. An open-circuit probe is placed near the test chuck, without contacting the chip's pins, to pick up electromagnetic interference from the surrounding environment. The physical quantity induced on the probe by this spatial electromagnetic interference is an induced voltage. Both signals are controlled by the same sampling clock to maintain time synchronization.

[0021] Strong electromagnetic interference generated by motors, power supplies, and other equipment in the test environment will couple to the leakage current signal in the form of random pulse spikes. The intensity changes of these interferences will also be reflected in the environmental noise signal. Therefore, the leakage current response pulse contaminated by strong interference can be identified based on the energy of the environmental noise in the corresponding time period.

[0022] Based on the trigger timestamp of each excitation pulse, extract the response segment corresponding to the pulse duration from the pulse response and align them to form the first... The response segment of each pulse is denoted as In relation to Same time window Internally, environmental noise signals The short-time energy value is calculated as follows: In the formula, Indicates environmental noise signal The short-time energy value, expressed in volts squared. For time windows Number of sampling points within, For time windows The sampling time within the range. This short-time energy value quantizes the first... The average power level of external electromagnetic interference during the duration of each pulse.

[0023] Obtain the short-time energy value of each pulse. Then, it is compared with the preset upper limit for noise abnormal pulse rejection. If a comparison is made, The corresponding response fragment Removed, Not Removed As a valid response fragment, its sequence set is denoted as Noise abnormal pulse rejection limit The setup is based on the statistical analysis of the baseline level of environmental noise before the test. An environmental noise signal with a duration of no less than 1 second is collected during a period without pulse excitation, and the variance is calculated. ,Pick The multiplier 3 here is a preset parameter, taken from experience, and can also be adjusted by the implementer according to the specific electromagnetic environment of the workstation.

[0024] S2: Remove DC from the effective response segment and divide it into transient and steady-state segments. Extract the transient energy peak coefficient and steady-state fluctuation index to construct the transient coefficient set and steady-state index set. Calculate the central tendency and dispersion of the transient coefficient set and steady-state index set and map them to the transient clustering index and steady-state clustering index.

[0025] In one embodiment, the method for removing DC from the effective response segment and dividing it into transient and steady-state segments is as follows: the steady-state maintenance segment after the effective response segment ends is extracted and the current time mean is calculated. The effective response segment is subtracted from the current time mean to eliminate baseline drift, resulting in a zero-mean response sequence. The segmentation boundary is determined based on the width of the excitation pulse and the preset transient duration. The preset time interval after the rising edge of the zero-mean response sequence ends is divided into the transient segment, and the remaining time interval is divided into the steady-state segment.

[0026] Regarding the aforementioned valid response fragments Because the test system exhibits a baseline drift that varies slowly with temperature and time under continuous pulse excitation, and each response segment is superimposed with a DC offset component, DC removal processing is required. In practice, the time from the end of the falling edge of the pulse is used as the reference, with a delay of... (Acquire experience points) After setting the preset parameters, a segment of time with a duration of 10 ... (Acquire experience points) The steady-state maintenance period (with preset parameters) is defined, and the mean current-time value of all sampling points within the steady-state maintenance period is calculated and denoted as . Then the entire valid response segment Subtract point by point The zero-mean response sequence is obtained. .

[0027] After obtaining the zero-mean response sequence, based on the rise time of the excitation pulse... Determine the transient duration coefficient The calculation formula is as follows: , These are empirical values ​​and can be adjusted based on the actual testing scenario. Next, interval division is performed to divide the zero-mean response sequence. From the end of the rising edge, the duration The interval is divided into transient segments, denoted as . And the remaining effective time interval of the pulse, i.e. from The period from the end of the pulse to the beginning of its falling edge is divided into a steady-state segment, denoted as... .

[0028] In one embodiment, the method for extracting the transient energy peak coefficient is as follows: Calculate the maximum absolute value of leakage current at all sampling points during the transient period. Within, take the maximum absolute value of the zero-mean leakage flow at all sampling points, and denot it as... .

[0029] Calculate the root mean square of the leakage current at all sampling points during the transient period, denoted as . The calculation formula is: ,in, Indicates the first The root mean square of leakage current for each pulse. Transient segment Number of sampling points within, Indicates the first A pulse in The zero-mean response at time t.

[0030] Dividing the maximum value by the sum of the root mean square of the leakage current and the preset zero constant yields the transient energy peak coefficient. The formula for calculating the transient energy peak coefficient is: , Indicates the first The peak transient energy coefficient of each pulse. To prevent taking empirical values ​​for the divide-by-zero constant A is a preset parameter whose function is to prevent division by zero errors when the denominator is extremely small or close to zero, thus ensuring the numerical stability of the algorithm in mathematical calculations. Dimensionless, it reflects the prominence of transient spikes relative to the overall energy; the larger the value, the more pronounced the transient spike.

[0031] In one embodiment, the method for extracting the steady-state fluctuation index is as follows: Calculate the absolute values ​​of the standard deviation and mean of all sampled values ​​within the steady-state range. The mean is denoted as... The standard deviation is denoted as ; ,in steady state segment Number of sampling points.

[0032] The steady-state fluctuation index is obtained by dividing the standard deviation by the sum of the absolute value of the mean and the preset zero constant; the mathematical expression is: , Indicates the first The steady-state fluctuation index of a pulse within the steady-state range Same as before A.

[0033] In response to the absolute value of the mean being less than a preset minimum threshold, the denominator in calculating the steady-state fluctuation index is replaced with the current range normalized value. Specifically, if... , To determine the minimum threshold, an empirical value is used. If A is a preset parameter, then the denominator will be replaced with the normalized value of the current range. In this embodiment, A, according to calculate, Dimensionless, representing the relative fluctuation intensity of current in the steady-state region.

[0034] The peak transient energy coefficient of all effective pulses from the same chip The transient coefficients are compiled according to the pulse timing sequence. All steady-state fluctuation indices The collection constitutes the steady-state index set. The number of elements in both sets is equal to the total number of effective pulses of the chip, and each element uniquely corresponds to the response of an excitation pulse, thus completing the construction of the single-chip transient and steady-state characteristic sets required in this step.

[0035] Obtaining the set of transient coefficients for the same chip With steady-state index set Subsequently, the phenomenon encountered in actual testing is that the leakage current transient distortion and steady-state fluctuations caused by real gate defects (such as degradation caused by trap-assisted tunneling) are increased and will systematically reproduce between multiple excitation pulses of the chip, resulting in a set of transient coefficients. Peak transient energy coefficient of each pulse They are close to each other; while random electromagnetic interference spikes are asynchronous and irregular in time, causing anomalies. The values ​​will be scattered in the form of outliers, resulting in a significant increase in the dispersion of values ​​within the set. Therefore, by quantifying the central tendency and dispersion measures of the two sets, transient and steady-state clustering indices are obtained to characterize the consistency of the recurrence of defect features between pulses.

[0036] In one embodiment, the method for calculating the transient aggregation index includes: calculating the median of the transient coefficient set and correcting the absolute deviation of the median; dividing the corrected absolute deviation of the median by the sum of the median and a preset zero constant, and performing negative correlation normalization to obtain the transient aggregation index.

[0037] For the transient coefficient set First, calculate the median of the set. As a robust measure of central tendency for outliers, it is defined as the peak energy coefficient of all transient pulses. After sorting by value, take the middle value. If the number of elements is even, take the average of the two middle values. Dimensionless. The median, rather than the arithmetic mean, is used as a measure of central tendency because, under the influence of occasional interference pulses that have not yet been completely eliminated, a few extremely large or small outliers have a strong pulling effect on the arithmetic mean, while the median can more accurately reflect the true clustering center of the majority of pulse characteristic values.

[0038] To calculate the measure of dispersion, the corrected median absolute deviation is used. First, the peak energy coefficient for each transient event is obtained. With median The absolute value of the difference, then take the median of these absolute differences, and multiply by the scaling factor. The mathematical expression for correcting the absolute deviation of the median is: In the formula, Set of transient coefficients The corrected absolute deviation of the median is dimensionless. This indicates the median operation; For the first The source of the transient energy peak coefficient of each pulse has been given in the preceding steps; Set of transient coefficients The median. The scaling factor of 1.5 is a preset empirical amplification coefficient used to moderately scale and amplify the dispersion of the set, thereby providing a more robust dispersion penalty weight for outliers when using negative exponential mapping.

[0039] The mathematical expression for the transient aggregation index is: In the formula, The transient aggregation exponent is dimensionless and its value ranges from 1 to 1. ; This refers to the aforementioned corrected absolute deviation of the median; Set of transient coefficients the median; To prevent division by zero constant, an empirical value is used. Dimensionless, used to prevent the median When the denominator approaches zero, the result is zero, leading to numerical anomalies. This can be adjusted by the implementer according to the specific implementation scenario. Represented by natural constant An exponential function with base 0.

[0040] The logic of this mapping lies in: when the set of transient coefficients Peak energy coefficients of each transient state Highly clustered at the median When near, correct the absolute deviation of the median. Minimal, factor Approaching The transient aggregation index approaches This indicates that the consistency of transient features across pulses is extremely strong, consistent with the behavioral patterns of real defects; conversely, when there are a large number of scattered outliers within the set, the absolute deviation of the corrected median increases significantly, the ratio increases, and the transient aggregation index decreases rapidly and approaches a certain value. This indicates that the dispersion of eigenvalues ​​is mainly caused by random disturbances.

[0041] In one embodiment, the steady-state clustering index is calculated similarly to the transient clustering index. For the set of steady-state indices... Perform the same central tendency and dispersion measures and negative exponential mapping process as described above. Specifically, first calculate the set of steady-state indices. the median of The median, as a measure of the central tendency of steady-state fluctuations, Dimensionless, its calculation method is the same as the median Same, that is, take all steady-state fluctuation indices. The median value after sorting. Then calculate the corrected absolute deviation of the median of the steady-state exponent set. The mathematical expression is: In the formula, Set of steady-state indices The corrected absolute deviation of the median is dimensionless. For the first The steady-state fluctuation index of each pulse has been given in the preceding steps; Set of steady-state indices The median. Scale factor. Its function is the same as described above: as an empirical amplification factor, it moderately scales the absolute deviation, thereby providing a robust estimate of the dispersion of outliers.

[0042] Steady-state aggregation index The calculation also uses a negative exponential mapping: In the formula, The steady-state aggregation index is dimensionless and its value ranges from 1 to 1. ; This refers to the aforementioned corrected absolute deviation of the median; Set of steady-state indices the median; To prevent division by zero constant, an empirical value is used. Dimensionless, this index can be adjusted by the implementer according to the specific implementation scenario. This index reflects whether the steady-state fluctuation values ​​of all effective pulses of the same chip have systematic consistency: the increase in steady-state leakage current fluctuation caused by real gate defects is stably presented across multiple pulses, indicating a steady-state aggregation index. It will be at a relatively high level; while the steady-state fluctuations caused by random noise do not have inter-pulse consistency, and the steady-state clustering index... Lower.

[0043] After the above processing, the transient coefficient set The numerical clustering degree is compressed into the transient clustering index. steady-state index set The degree of numerical clustering is compressed to Transient aggregation index With steady-state aggregation index All with Indicates high aggregation, approaching This indicates high dispersion, thereby completing the calculation of the central tendency measure and dispersion measure required in this step and mapping them to generate transient and steady-state clustering indices.

[0044] S3: Calculate the absolute value of the difference between the transient aggregation index and the steady-state aggregation index as the linkage coupling deviation, and generate a coupling factor that decreases as the linkage coupling deviation increases.

[0045] In practical GaN HEMT (Gallium Nitride High Electron Mobility Transistor) devices, during early gate degradation, physical mechanisms such as trap-assisted tunneling lead to a positive correlation between transient distortion and steady-state fluctuations in gate leakage current. That is, once a real defect forms, the transient aggregation exponential... With steady-state aggregation index Not only are they both at a high level, but their numerical values ​​should also be close to each other.

[0046] However, random electromagnetic interference in the test environment often lacks this inherent correlation, and transient clustering may occur due to occasional spikes during the transient phase. Abnormally high steady-state aggregation index Maintaining a low level, or the steady-state segment being affected by low-frequency noise causing a steady-state aggregation index. Inflated and transient aggregation index A stable situation.

[0047] To address this phenomenon, this step quantifies the degree to which the two aggregation indices of the current chip deviate from the ideal defect linkage mode, generating a coupling factor that decreases as the deviation increases, thus providing a basis for suppressing false triggering of noise unilateral features during subsequent fusion.

[0048] Specifically, the transient-steady-state linkage deviation is defined as the absolute value of the difference between the transient aggregation index and the steady-state aggregation index on the current chip, and its calculation formula is: In the formula, The linkage coupling deviation is dimensionless and its value range is [value range missing]. ; The transient aggregation index is dimensionless and has been generated in the preceding steps from the transient coefficient set through a measure of central tendency, a measure of dispersion, and a negative exponential mapping. The steady-state aggregation index is dimensionless and has been generated from the steady-state index set through the same process in the previous steps.

[0049] When the gate truly has early degradation defects and is not contaminated by strong interference. and If the values ​​are close and the linkage and coupling deviation is small, it indicates that the current feature combination conforms to the defect linkage law; if only one of the transient and steady-state features increases abnormally due to noise, the difference between the two indices widens, the linkage and coupling deviation increases accordingly, indicating that the current combination deviates from the physical prior, and its credibility decreases.

[0050] To convert the magnitude of the linkage coupling deviation into a modulation coefficient that can be applied to the fusion weights, a Gaussian attenuation function is used to generate the coupling factor. : In the formula, The coupling factor is dimensionless and its value range is... ; This is a linkage coupling deviation; The tolerance of the linkage constraint is set to an empirical value of 0.25, which is dimensionless. It can also be adjusted by the implementer according to the specific implementation scenario. This parameter determines the broadening of the Gaussian decay function, that is, the range within which the deviation is considered to basically meet the linkage constraint. Indicates An exponential function with base 0.

[0051] Index Term Maintaining a non-positive state, when the linkage coupling deviation... Approaching At that time, the term approaches 0. A value approaching 1 indicates that the transient and steady-state characteristics highly conform to the linkage law; when the linkage coupling deviation... Increase, this item increases negatively. The Gaussian function rapidly decays and approaches zero, effectively suppressing combinations of features that do not conform to physical priors. The gradient of change in smaller areas is relatively gentle, which can tolerate minor linkage deviations caused by measurement noise or normal process fluctuations between chips, avoiding excessive sensitivity of coupling factors that could lead to an incorrect underestimation of the true defect score; when linkage deviations... Once the tolerance control range is exceeded, the attenuation rate accelerates, ensuring strong suppression of the unilateral characteristics of noise.

[0052] Therefore, this step outputs the coupling factor. As a quantitative representation of the physical linkage consistency between transient and steady-state characteristics, it is used for dynamic modulation and fusion weights when generating the subsequent defect quantification risk index.

[0053] S4: The defect quantification risk index is generated by weighting the mean of the transient clustering index and the steady-state clustering index using the coupling factor and combining the product of the transient clustering index and the steady-state clustering index.

[0054] To obtain the transient aggregation index of the current chip Steady-state aggregation index and coupling factor Subsequently, in actual testing, random electromagnetic interference may cause one of the two indices to show a unilateral abnormal increase while the other remains low. If the arithmetic mean is directly used for fusion, the unilateral abnormality will still increase the fusion result to some extent, creating a false alarm risk. Furthermore, real gate degradation defects physically inevitably drive both transient distortion and steady-state fluctuations simultaneously, making... and Both are at a high level and are interconnected. Based on this phenomenon, this step modulates the arithmetic mean term using a coupling factor, and introduces a product term to apply additional reinforcement when both indices are at the same level. The weighted mean term and the product reinforcement term are generated sequentially and added together to obtain the defect quantification risk index.

[0055] The defect quantification risk index is calculated as follows: The arithmetic mean of the transient and steady-state clustering indices is calculated; the arithmetic mean is multiplied by the coupling factor to obtain the weighted mean term; when the transient and steady-state characteristics highly conform to the defect linkage law, the coupling factor... Approaching The weighted mean term almost completely retains the average value; when the linkage deviation is large, the coupling factor... Approaching This factor was significantly suppressed, effectively reducing the contribution of one-sided noise features to the fusion result.

[0056] Calculate the product of the transient aggregation index and the steady-state aggregation index; the product is only valid for the transient aggregation index. Steady-state aggregation index A larger value can only be obtained when both are at relatively high levels. Even if one index is high while the other is low, the product will still remain at a low level.

[0057] Multiplying the product by a preset risk fusion balance coefficient yields the product enhancement term. For example, the risk fusion balance coefficient can be 0.3, or it can be adjusted by the implementer according to the specific detection scenario to control the tolerance for false alarms and missed detections. This coefficient is used to control the weight of the product term in the final risk index, so that when both indices are at a high level, the effect of the product term on raising the risk index is within a reasonable range.

[0058] Adding the weighted mean term to the product reinforcement term generates the defect quantification risk index, denoted as... Dimensionless, a larger value indicates that the leakage current response characteristics of the current chip more closely match the behavior pattern of a real gate defect. In the presence of a real defect, the transient aggregation index... With steady-state aggregation index High average and small linkage deviation, coupling factor near Transient aggregation index With steady-state aggregation index The arithmetic mean is relatively large, and the product term provides additional enhancement, making the defect quantification risk index more robust. Significantly higher than the level achievable by the mean term alone; when the coupling factor is artificially inflated due to noise alone, The index exhibits severe attenuation, with the weighted mean term being suppressed and the product term contributing only a limited amount due to the low value of another index, indicating a defect-quantified risk index. It is effectively limited to a low value, thereby achieving a balance between noise resistance and high detection specificity at the fusion level.

[0059] This completes the generation of the current chip defect quantification risk index.

[0060] S5: Calculate a smoothed noise level estimate based on the environmental noise signal, and generate an adaptive judgment threshold by combining it with the current temperature of the GaN chip; if the defect quantification risk index is greater than or equal to the adaptive judgment threshold, the GaN chip is determined to have a gate defect.

[0061] To obtain the defect quantification risk index of the current chip Subsequently, the electromagnetic interference intensity at the actual testing station fluctuated significantly between different testing cycles, while the wafer temperature deviated from the calibrated value due to the accuracy of environmental temperature control and the self-heating effect of the device. A fixed judgment threshold is difficult to keep the false judgment probability stable under different operating conditions. This application utilizes synchronously acquired environmental noise signals and real-time wafer temperature to generate an adaptive judgment threshold that dynamically adjusts with environmental conditions.

[0062] Specifically, the environmental noise signal collected during the current chip testing cycle. Calculate the root mean square of all sampled values ​​as the instantaneous noise level. External electromagnetic interference often contains random spikes, which can be directly addressed by... Adjusting the threshold will cause a short-term jump in the threshold, therefore a first-order exponential moving average is used to... Smoothing is performed, and the smoothed noise level estimate is obtained recursively. The mathematical expression is: In the formula, This is an estimate of the smoothed noise level for the current period. This represents the root mean square value of the instantaneous noise in the current period. The noise smoothing forgetting factor is taken as an empirical value of 0.1. It is dimensionless and is a preset parameter, but it can also be adjusted by the implementer according to the specific implementation scenario. This is the estimated smooth noise level calculated in the previous test period. When the value is small, Primarily determined by historical information, the response to instantaneous spikes is smooth, allowing the estimated value to maintain good stability while tracking changes in noise trends.

[0063] For the first chip's test cycle, it can be Initialize to the chip Value. Current temperature of the wafer. Temperature data is collected in real time by a temperature sensor mounted on the probe station or wafer chuck, and the unit is Celsius.

[0064] The base current and fluctuation amplitude of the gate leakage current of GaN chips drift with changes in junction temperature. When the temperature deviates from the calibrated reference temperature, the transient energy peak coefficient and steady-state fluctuation index corresponding to the same defect severity will exhibit systematic shifts, requiring corresponding adjustments to the judgment threshold. Therefore, the obtained smoothed noise level estimate will be... and current temperature Substituting the preset linear mapping function, an adaptive judgment threshold for the current chip is generated. : In the formula, The threshold is dimensionless and is designed for adaptive judgment. The baseline threshold is set at an empirical value of 0.6. It is dimensionless and is a preset parameter that represents the baseline for judgment when there is no noise interference and the temperature is equal to the reference temperature. It can also be adjusted by the implementer according to the specific implementation scenario. The noise gain coefficient, taken as an empirical value of 2, is dimensionless and is a preset parameter used to smooth the estimated noise level. A linear mapping is used to represent the threshold increment; The preset reference noise root mean square value, for example, can be 0.1 volts, is used for normalization, to eliminate dimensions and limit the threshold rise. The temperature compensation coefficient is taken as an empirical value of 0.01. It is dimensionless and is a preset parameter that reflects the threshold correction amount required per unit temperature deviation. The reference temperature is fixed at 25 degrees Celsius for calibration. The preset reference temperature difference, for example, can be 10 degrees Celsius, is used to normalize the actual temperature deviation to the standard temperature fluctuation range.

[0065] Smoothing noise level estimate when external interference increases Increase, adaptive judgment threshold The temperature rises accordingly; current temperature Deviation from the calibration reference temperature When, adaptively determine the threshold The temperature compensation term is adjusted accordingly. Thus, each chip obtains an adaptive decision threshold that dynamically adapts to ambient noise intensity and temperature. .

[0066] Because early defects in the real gate cause a high degree of aggregation and correlation between multi-pulse transient and steady-state characteristics, thus increasing the defect quantification risk index. Increased environmental interference or temperature deviation from the calibration value will push up the adaptive judgment threshold through linear mapping. Quantify the risk index of defects With adaptive decision threshold Direct comparison can distinguish between defective and normal chips under a wide range of testing conditions.

[0067] If satisfied This indicates that the comprehensive score of the trans-pulse transient consistency, steady-state fluctuation clustering, and the correlation between the two in the chip's leakage current response has exceeded the upper limit of normal fluctuations allowed by the current environmental noise intensity and temperature. Therefore, the GaN chip is determined to have a gate defect and is marked as a failed device. This indicates that the chip's multi-pulse feature aggregation level has not reached the intensity required for the defect mode, or that the current noise and temperature conditions have kept the threshold at a level adapted to the environment and have not triggered the judgment. The chip is then judged to be qualified, and the corresponding normal test result is output.

[0068] This application also discloses a GaN chip gate fault detection system based on multi-source feature fusion, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the GaN chip gate fault detection method based on multi-source feature fusion according to this application.

[0069] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0070] In this application, the aforementioned memory can be any tangible medium that contains or stores a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as resistive random access memory, dynamic random access memory, static random access memory, etc., or any other medium that can be used to store required information and can be accessed by an application program, module, or both.

[0071] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for detecting gate faults in GaN chips based on multi-source feature fusion, characterized in that, Including the following steps: A multi-pulse excitation was applied to the GaN chip, and the pulse response and environmental noise signal were collected. Based on the environmental noise signal, abnormal pulses in the pulse response were removed to obtain the effective response segment. DC is removed from the effective response segment and the transient segment and steady-state segment are divided. The transient energy peak coefficient and steady-state fluctuation index are extracted, and the transient coefficient set and steady-state index set are constructed. Calculate the central tendency and dispersion of the transient coefficient set and the steady-state index set, and map them to transient clustering index and steady-state clustering index; The absolute value of the difference between the transient aggregation index and the steady-state aggregation index is calculated as the linkage coupling deviation, and a coupling factor that decreases as the linkage coupling deviation increases is generated. The defect quantification risk index is generated by weighting the transient and steady-state clustering indices using coupling factors and combining the product of the transient and steady-state clustering indices. The smoothed noise level estimate is calculated based on the environmental noise signal, and an adaptive judgment threshold is generated by combining the current temperature of the GaN chip. If the defect quantification risk index is greater than or equal to the adaptive judgment threshold, the GaN chip is determined to have a gate defect.

2. The GaN chip gate fault detection method based on multi-source feature fusion according to claim 1, characterized in that, The method for calculating the transient aggregation index includes: calculating the median of the transient coefficient set and the corrected absolute deviation of the median; dividing the corrected absolute deviation of the median by the sum of the median and a preset zero constant, and performing negative correlation normalization to obtain the transient aggregation index; Similarly, the steady-state aggregation index is calculated.

3. The GaN chip gate fault detection method based on multi-source feature fusion according to claim 1, characterized in that, The method for obtaining an effective response segment by removing abnormal pulses from the impulse response based on environmental noise signals includes the following steps: Based on the trigger timestamp of each pulse, extract the response segment corresponding to the pulse duration from the pulse response; Within the same time window as the response segment, calculate the short-time energy value of the environmental noise signal; If the short-time energy value exceeds the preset noise abnormal pulse rejection limit, the corresponding response segment is marked as invalid and rejected, and the response segments that are not rejected are taken as the valid response segments.

4. The GaN chip gate fault detection method based on multi-source feature fusion according to claim 1, characterized in that, The method for removing DC from the effective response segment and dividing it into transient and steady-state segments is as follows: After the effective response segment ends, the steady-state maintenance segment is extracted and the current-time mean is calculated. The effective response segment is then subtracted from the current-time mean to eliminate baseline drift, resulting in a zero-mean response sequence. The segmentation boundary is determined based on the width of the excitation pulse and the preset transient duration. The preset time interval after the rising edge ends in the zero-mean response sequence is divided into the transient segment, and the remaining time interval is divided into the steady-state segment.

5. The GaN chip gate fault detection method based on multi-source feature fusion according to claim 1, characterized in that, The extraction of the transient energy peak coefficient includes: Calculate the maximum absolute value of leakage current at all sampling points during the transient period; Calculate the root mean square of the leakage current at all sampling points during the transient period; Divide the maximum value by the sum of the root mean square of the leakage current and the preset zero constant to obtain the transient energy peak coefficient.

6. The GaN chip gate fault detection method based on multi-source feature fusion according to claim 1, characterized in that, The steady-state fluctuation index includes: Calculate the absolute values ​​of the standard deviation and mean of all sampled values ​​within the steady-state range; The steady-state fluctuation index is obtained by dividing the standard deviation by the sum of the absolute value of the mean and the preset zero constant. In response to the absolute value of the mean being less than a preset minimum threshold, the denominator in the calculation of the steady-state fluctuation index is replaced with the current range normalized value.

7. The method for detecting gate faults in GaN chips based on multi-source feature fusion according to claim 1, characterized in that, The method for calculating the defect quantification risk index is as follows: Calculate the arithmetic mean of the transient aggregation index and the steady-state aggregation index; Multiplying the arithmetic mean by the coupling factor yields the weighted mean term; Calculate the product of the transient aggregation index and the steady-state aggregation index, and multiply the product by the preset risk fusion balance coefficient to obtain the product reinforcement term; The weighted mean term is added to the product reinforcement term to generate the defect quantification risk index.

8. A GaN chip gate fault detection system based on multi-source feature fusion, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the GaN chip gate fault detection method based on multi-source feature fusion according to any one of claims 1-7.

Images