A method for predicting the remaining life of a power device

Abstract

The application provides a power device residual life prediction method, comprising: obtaining full life cycle data and degradation characteristic-cycles data of the aging process of the power device under different stress levels; fitting stress-life data to establish an analytical life model thereof; representing the degradation characteristic-cycles data of the power device under different stress levels as degradation characteristic-cumulative damage data; obtaining historical stress data and degradation characteristic data of the power device to be predicted under dynamic working conditions; representing the degradation characteristic data as degradation characteristic-cumulative damage data; and using a degradation trajectory similarity life prediction method to predict the residual service life of the power device under dynamic working conditions. The application can more accurately predict the residual life of the power device.

Description

Technical Field

[0001] This invention belongs to the field of power semiconductor device reliability, and particularly relates to a method for predicting the remaining lifetime of power devices. Background Technology

[0002] Power semiconductor devices are the core components for power conversion and are widely used in smart grids, high-speed rail traction, industrial frequency conversion, and new energy vehicles. Their reliability is a key factor affecting the reliability of power electronic systems. Power semiconductor device failures account for 31% of all power electronic converter failures, making them the most vulnerable component in power electronic converters.

[0003] Failure modes of power semiconductor devices can be categorized into random failures and aging failures. Random failures are typically caused by overcurrent thermal stress resulting from improper use or design. The risk of random failures can be mitigated by adding protection circuits, optimizing system and device design, and improving manufacturing and packaging processes. However, thermal mismatch caused by differences in the coefficients of thermal expansion between multilayer packaging materials leads to alternating thermal stress when the device is subjected to high temperatures and temperature changes, which is a major cause of fatigue aging. Solder layer fatigue cracking and bond wire detachment are two main manifestations of aging failure. The aging failure process of devices inevitably occurs throughout their entire application lifecycle. Therefore, health status assessment and Remaining Useful Life (RUL) prediction for power devices are crucial for reducing the risk of unplanned downtime in power electronic systems and achieving long-term safe and reliable operation.

[0004] Existing methods for predicting the lifetime of power devices can be divided into analytical model methods and data-driven methods. Analytical lifetime models generally establish the number of thermal cycles N before power device failure by fitting data obtained from accelerated life tests. fThe analytical expression between thermal stress and degradation means that analytical lifetime modeling methods can only provide the expected lifetime of the same product model under a certain reliability / failure probability, and cannot provide updatable predictions for the degradation process of a specific individual. Data-driven lifetime prediction methods establish data-driven models based on the full lifecycle measurement data of degradation characteristics during the aging process of similar products, and then combine this with monitoring data from the actual operation of a specific product or system to achieve lifetime prediction for a specific individual. However, because bond wire detachment can occur suddenly during the aging process of power devices, existing data-driven methods based on degradation trend modeling, time series analysis, and regression prediction struggle to accurately model and predict the jumps and nonlinear characteristics of degradation characteristics. Furthermore, the degradation rate of a product is closely related to the stress level it experiences. Existing studies using data-driven models for power device lifetime prediction often rely on constant stress accelerated test degradation process data of the same product model, failing to model the stress changes, degradation rate, and lifetime within the data-driven model. This falls far short of the application requirement of real-time updatable lifetime prediction for specific power devices under actual dynamic operating conditions. Summary of the Invention

[0005] To overcome the shortcomings of existing power device lifetime prediction methods, such as the inability to predict dynamic operating conditions, individualized predictions, and predictions that can be updated in real time with monitoring data, this invention proposes a power device remaining lifetime prediction method. This method is a power device remaining lifetime prediction method that integrates an analytical lifetime model and a data-driven model based on degradation trajectory similarity, which can improve the accuracy of power device remaining lifetime prediction.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for predicting the remaining lifetime of power devices includes the following steps:

[0008] Step (1) Conduct constant stress accelerated life test through power device aging test platform to obtain full life cycle data of power device aging process under different stress levels, and extract degradation characterization quantity-cycle number data through full life cycle data;

[0009] Step (2) Obtain the stress-life data of the power device based on the degradation characterization quantity-cycle number data in step (1), and establish its analytical lifetime model by fitting the stress-life data;

[0010] Step (3) Based on the analytical lifetime model of the power device obtained in step (2), the Miner damage accumulation method is used to express the degradation characterization quantity-cycle number data of the power device under different stress levels as degradation characterization quantity-cumulative damage data.

[0011] Step (4) Perform online monitoring of the power devices under dynamic operating conditions to obtain their historical stress data and degradation characterization data;

[0012] Step (5) processes the historical stress data of the power device under dynamic operating conditions obtained in step (4), and uses the analytical lifetime model and Miner damage accumulation method in step 3 to represent the degradation characterization data as degradation characterization-cumulative damage data.

[0013] Step (6) uses the power device degradation characterization quantity-cumulative damage degree data obtained in step (3) as the reference sample degradation trajectory dataset, and the power device degradation characterization quantity-cumulative damage degree data obtained in step (5) as the test sample degradation trajectory dataset. The degradation trajectory similarity lifetime prediction method is used to predict the remaining service life of the power device under dynamic conditions.

[0014] Furthermore, in step (1), the degradation characterization quantity refers to the physical quantity of device parameters that changes monotonically with the aging process of the power device.

[0015] Further, in step (2), the lifetime of the power device under different stress levels refers to the number of cycles from the start of the test until the change in the degradation characterization quantity reaches the given failure threshold; the analytical lifetime model of the power device is selected from the Coffin-Manson model, the Less model, or the Bayer model, and their expressions are as follows:

[0016] N f =A(ΔT) j ) b

[0017]

[0018]

[0019] Where, N f It is the number of cycles before failure, i.e., the cycle life, ΔT j It is the junction temperature range, T m It is the average junction temperature, T jmax It is the maximum junction temperature, t on I is the heating time, U is the effective current through each bonding wire, D is the module's withstand voltage, and k is the diameter of the bonding wire. B and E a Let A, b, and β1 to β6 be the Boltzmann constant and the activation energy of the silicon chip, respectively.

[0020] Furthermore, in step (3), the expression for Miner damage accumulation is:

[0021]

[0022] Where D is the cumulative damage degree, k is the data point index at the current time, and n i N represents the number of cycles for the i-th stress amplitude. fi Let be the cycle life for the i-th stress amplitude.

[0023] Further, step (6) includes:

[0024] Step (6.1) uses the power device degradation characterization quantity-cumulative damage data under dynamic operating conditions obtained in step (5) as the test sample degradation trajectory dataset, and extracts the H+1 data points closest to the current time in the test sample degradation trajectory data as the test sample degradation trajectory segment. The data point sequence is X0(k,H)=[x0((k—H)·Δt),…,x0(k·Δt)], where H is a non-negative integer, Δt is the degradation trajectory sampling interval, k is the data point number at the current time, x is the data point of the degradation characterization quantity, and X is the data point sequence of the degradation trajectory segment;

[0025] Step (6.2) extracts any H+1 consecutive data points from the start of operation to the point of failure of each reference sample as a reference sample degradation trajectory segment, and its data point sequence is X. i (m,H)=[x i ((m—H)·Δt),…,x i [(m·Δt)], where i is the reference sample index, m is the index of any sampling point of the reference sample, and H+1≤m≤M i M i This represents the number of data points in reference sample i ranging from completely healthy to completely failed.

[0026] Step (6.3) involves iteratively matching the trajectory segments of the test sample similarity comparison with any similarity comparison trajectory segment of each reference sample from the start of operation to failure, and calculating the degradation trajectory segment X0(k,H) of the test sample and the degradation trajectory segment X of the reference sample. i Similarity of (m,H);

[0027] Step (6.4) obtains the reference sample degenerate trajectory segment that has the highest similarity to the test sample degenerate trajectory segment among the reference sample degenerate trajectory segments, and selects the end time N of the trajectory segment that makes the reference sample i most similar to the test sample. i The remaining lifetime at (k)·Δt is used as the estimated remaining lifetime, i.e., RUL. ei (k), where:

[0028]

[0029] RUL ei(k)=(M i —N i (k))·Δt

[0030] N i (k) represents the trajectory segment X that has the highest similarity to the test sample. i The value of m in (m,H);

[0031] Step (6.5) involves a weighted sum of the remaining lifetime estimates for all reference samples to obtain the remaining lifetime prediction result for the test sample at the current time, expressed as:

[0032]

[0033] Among them, w i (k) is the weighting factor, and L is the total number of reference samples.

[0034] Beneficial effects:

[0035] Compared with existing methods, this invention integrates analytical lifetime modeling and similarity-based data-driven methods, and combines lifetime data and degradation data. It can predict the remaining lifetime of a specific device under dynamic operating conditions. The prediction results can be updated in real time as the monitoring data of the specific device increases, and the remaining lifetime of power devices can be predicted more accurately. Attached Figure Description

[0036] Figure 1 This is a flowchart of a method for predicting the remaining lifetime of power devices according to the present invention.

[0037] Figure 2 This section presents stress-lifetime data and fitting curves for power devices.

[0038] Figure 3 This is a graph showing the change in on-resistance of power devices versus cumulative damage under constant stress accelerated testing.

[0039] Figure 4 This is a graph showing the change in on-resistance of power devices under dynamic operating conditions versus the cumulative damage level.

[0040] Figure 5 This is a diagram showing the dynamic operating condition test conditions.

[0041] Figure 6 This is a graph showing the predicted remaining lifetime of power device #1 under dynamic operating conditions.

[0042] Figure 7 This is a graph showing the predicted remaining lifetime of power device #2 under dynamic operating conditions. Detailed Implementation

[0043] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0044] The design concept of this invention is as follows: Existing power device lifetime prediction methods, such as analytical lifetime models, can only predict the average lifetime of a sample of the same type, and cannot predict the specific degradation process of a particular individual device. Data-driven methods based on degradation trend modeling, time series analysis, and regression prediction can predict the specific degradation process of a particular individual device, but they still have shortcomings such as the inability to perform cross-operating condition predictions and the difficulty in accurately modeling nonlinear trends. Therefore, this invention proposes a power device remaining lifetime prediction method, specifically a power device remaining lifetime prediction method that integrates an analytical lifetime model and a data-driven model based on degradation trajectory similarity, for predicting the remaining lifetime of individual power devices under dynamic operating conditions.

[0045] like Figure 1 As shown, a method for predicting the remaining lifetime of power devices according to the present invention includes the following steps:

[0046] Step 1: Conduct constant stress accelerated life test using a power device aging test platform to obtain full life cycle data of the power device aging process under different stress levels. Extract stress-life data and degradation characterization quantity-cycle number data of the power device from the full life cycle data.

[0047] Furthermore, in step 1, the degradation characterization quantity refers to the physical quantity of the device parameter that changes monotonically with the aging process of the power device, and the lifespan of the power device under different stress levels refers to the number of cycles from the start of the test until the change in the degradation characterization quantity reaches the given failure threshold.

[0048] In this embodiment, a domestically produced 450A / 1200V IGBT module was selected as the test object, and the junction temperature variation ΔT at different levels was measured. j The constant stress power cycle test under certain conditions was used to obtain the full life cycle data of this type of IGBT module, and the change in on-resistance was used as the degradation characterization quantity.

[0049] Step 2: Obtain the stress-lifetime data of the power device based on the degradation characterization quantity-cycle count data from Step 1, and establish its analytical lifetime model by fitting the stress-lifetime data.

[0050] Further, in step 2, the lifetime of the power device under different stress levels refers to the number of cycles from the start of the test until the change in the degradation characterization quantity reaches the given failure threshold. The analytical lifetime model for the power device can be the Coffin-Manson model, the Less model, or the Bayer model, and their expressions are as follows:

[0051] N f =A(ΔT) j )b

[0052]

[0053]

[0054] Where, N f It is the number of cycles before failure, i.e., the cycle life, ΔT j It is the junction temperature range, T m It is the average junction temperature, T jmax It is the maximum junction temperature, t on I is the heating time, U is the effective current through each bonding wire, D is the module's withstand voltage, and k is the diameter of the bonding wire. B and E a Let A, b, and β1 to β6 be the Boltzmann constant and the activation energy of the silicon chip, respectively.

[0055] In this embodiment, a failure threshold of 0.15mΩ on-resistance change is used. The number of cycles required for the power device to reach the failure threshold is extracted to obtain the stress-lifetime data of the power device. This data is then processed according to the Coffin-Manson model, i.e., N... f =A(ΔT) j ) b The stress-life data were fitted, and the fitting results are as follows: Figure 2 As shown.

[0056] Step 3: Based on the analytical lifetime model of the power device obtained in Step 2, the Miner damage accumulation method is used to express the degradation characterization quantity-cycle number data of the power device under different stress levels as degradation characterization quantity-cumulative damage data.

[0057] Furthermore, in step 3, the expression for the accumulation of Miner damage is:

[0058]

[0059] Where D is the cumulative damage degree, k is the data point index at the current time, and n i N represents the number of cycles for the i-th stress amplitude. fi Let be the cycle life for the i-th stress amplitude.

[0060] In this embodiment, by accumulating Miner damage, the degradation characterization quantity-cumulative damage degree data of the power device throughout its entire life cycle is obtained, as follows: Figure 3 As shown.

[0061] Step 4: Perform online monitoring of power devices under dynamic operating conditions to obtain their historical stress data and degradation characterization data.

[0062] In this embodiment, two domestically produced 450A / 1200V IGBT modules of the same model as those used in the constant stress power cycling test were used as test objects. Power cycling tests were conducted under dynamic operating conditions, and the historical junction temperature variation ΔT of the two samples was obtained. j Data and on-resistance change data, dynamic operating conditions such as Figure 4 .

[0063] Step 5: Process the historical stress data of the power device under dynamic operating conditions obtained in Step 4, and use the analytical lifetime model and Miner damage accumulation method in Step 3 to represent the degradation characterization data as degradation characterization-cumulative damage data.

[0064] In this embodiment, based on the historical junction temperature variation ΔT of the two samples... j The data calculator displays the cumulative damage degree as a function of cycle number. Then, the data of change in on-resistance minus cycle number is transformed into data of change in on-resistance minus cumulative damage degree. The results are as follows: Figure 5 .

[0065] Step 6: Use the power device degradation characterization quantity-cumulative damage degree data obtained in Step 3 from the constant stress accelerated life test as the reference sample degradation trajectory dataset, and use the power device degradation characterization quantity-cumulative damage degree data under dynamic conditions obtained in Step 5 as the test sample degradation trajectory dataset. Use the degradation trajectory similarity life prediction method to predict the remaining service life of the power device under dynamic conditions.

[0066] Furthermore, step 6 includes the following steps:

[0067] Step 601: Use the power device degradation characterization quantity-cumulative damage data obtained in Step 5 under dynamic operating conditions as the test sample degradation trajectory dataset. Extract the H+1 data points closest to the current time in the test sample degradation trajectory data as test sample degradation trajectory segments. The data point sequence is X0(k,H)=[x0((k—H)·Δt),...,x0(k·Δt)], where H is a non-negative integer, Δt is the degradation trajectory sampling interval, k is the data point number at the current time, x is the data point of the degradation characterization quantity, and X is the data point sequence of the degradation trajectory segment.

[0068] Step 602: Extract any H+1 consecutive data points from the start of operation to the point of failure of each reference sample as a reference sample degradation trajectory segment, and the data point sequence is X. i (m,H)=[x i ((m—H)·Δt),...,x i [(m·Δt)], where i is the reference sample index, m is the index of any data point in the reference sample, and H+1≤m≤M iM i This represents the number of data points in reference sample i ranging from fully healthy to fully failed.

[0069] Step 603: Perform cyclic matching between the trajectory segment of the test sample similarity comparison and any similarity comparison trajectory segment of each reference sample from the start of operation to the point of failure, and calculate the degradation trajectory segment X0(k,H) of the test sample and the degradation trajectory segment X of the reference sample. i Similarity of (m,H).

[0070] In this embodiment, the degradation trajectory segment X0(k,H) of the test sample and the degradation trajectory segment X of the reference sample are used. i The Euclidean distance (m, H) is used as a similarity measure, and the Euclidean distance d(X0, X) between the test sample and the reference sample is... i The smaller the value, the higher the degenerative similarity between the two. Its expression is:

[0071]

[0072] Where g is a non-negative integer. x0 and x i These are the data points representing the degradation characteristics of the test sample and the reference sample i, respectively, X0 and X... i These are the data point sequences of the degenerate trajectory segments of the test sample and the reference sample i, respectively.

[0073] Step 604: Obtain the reference sample degenerate trajectory segment that has the highest similarity to the test sample degenerate trajectory segment among the reference sample degenerate trajectory segments, and take the end time N of the trajectory segment that makes the reference sample i most similar to the test sample. i (k)·Δt(N i (k) is the trajectory segment X that has the highest similarity to the test sample. i The remaining lifetime at the value of m in (m,H) ​​is used as the remaining lifetime estimate, i.e., RUL. ei (k), where:

[0074]

[0075] RUL ei (k)=(M i —N i (k))·Δt

[0076] Step 605: Sum the weighted remaining lifetime estimates of all reference samples to obtain the remaining lifetime prediction result of the test sample at the current time. Its expression is:

[0077]

[0078] Among them, wi (k) is the weighting factor, and L is the total number of reference samples.

[0079] In this embodiment, using the kernel density weighting method, the remaining lifetime of reference sample i at time k·Δt is first obtained as follows:

[0080] RUL i (k)=(M i —k)·Δt

[0081] Then the probability density distribution of the remaining lifetime of the sample at time k·Δt for:

[0082]

[0083] Where L is the number of reference samples, r is an arbitrary point of remaining lifetime, and K h K is the scaling kernel function, and K is the kernel function. In this embodiment, the Gaussian kernel function is used; h>0 is a smoothing parameter called bandwidth.

[0084] Then the weighting function w based on kernel density estimation i The expression for (k) is:

[0085]

[0086] Finally, the weighted sum of the remaining lifetime estimates of all reference samples based on the weighting function is the predicted remaining lifetime of the test sample at the current moment. The lifetime prediction results for the two IGBT modules under dynamic operating conditions are as follows: Figure 6 and Figure 7 As shown in the figure. The prediction results show that the overall output of the prediction model can accurately track the cycle life consumption trend, and the absolute error of the prediction gradually decreases in the region nearing the end of life.

[0087] This invention proposes a power device remaining lifetime prediction method that integrates an analytical lifetime model and a data-driven model based on degradation trajectory similarity. This method is used to predict the remaining lifetime of individual power devices under dynamic operating conditions. Compared with existing methods, it can make full use of the historical degradation data and lifetime data of power devices, improve the accuracy of power device remaining lifetime prediction, and update the prediction results as monitoring data increases, thereby reducing the absolute error of the prediction.

Claims

1. A method for predicting the remaining lifetime of power devices, characterized in that, Includes the following steps: Step (1) Conduct constant stress accelerated life test through power device aging test platform to obtain full life cycle data of power device aging process under different stress levels, and extract degradation characterization quantity-cycle number data through full life cycle data; Step (2) Obtain the stress-lifetime data of the power device based on the degradation characterization quantity-cycle number data in step (1), and establish its analytical lifetime model by fitting the stress-lifetime data; Step (3) Based on the analytical lifetime model of the power device obtained in step (2), the Miner damage accumulation method is used to express the degradation characterization quantity-cycle number data of the power device under different stress levels as degradation characterization quantity-cumulative damage data. Step (4) Perform online monitoring of the power devices under dynamic operating conditions to obtain their historical stress data and degradation characterization data; Step (5) processes the historical stress data of the power device under dynamic operating conditions obtained in step (4), and uses the analytical lifetime model and Miner damage accumulation method in step (3) to represent the degradation characterization data as degradation characterization-cumulative damage data. Step (6) uses the power device degradation characterization quantity-cumulative damage degree data obtained in step (3) as the reference sample degradation trajectory dataset, and the power device degradation characterization quantity-cumulative damage degree data obtained in step (5) as the test sample degradation trajectory dataset. The degradation trajectory similarity lifetime prediction method is used to predict the remaining service life of the power device under dynamic conditions.

2. The method for predicting the remaining lifetime of power devices according to claim 1, characterized in that, In step (1), the degradation characterization quantity refers to the physical quantity of device parameters that changes monotonically with the aging process of the power device.

3. The method for predicting the remaining lifetime of power devices according to claim 1, characterized in that, In step (2), the lifetime of the power device under different stress levels refers to the number of cycles from the start of the test until the change in the degradation characterization reaches the given failure threshold; the analytical lifetime model of the power device is selected from the Coffin-Manson model, the Less model, or the Bayer model, and their expressions are as follows: ； ； ； Where, N f It is the number of cycles before failure, i.e., the cycle life, ΔT j It is the junction temperature range, T m It is the average junction temperature, T jmax It is the maximum junction temperature, t on I is the heating time, U is the effective current through each bonding wire, D is the module's withstand voltage, and k is the diameter of the bonding wire. B and E a Let A, B, and β1~β6 be the Boltzmann constant and the activation energy of the silicon chip, respectively.

4. The method for predicting the remaining lifetime of a power device according to claim 1, characterized in that, In step (3), the expression for Miner damage accumulation is: ； Where D is the cumulative damage degree, k is the data point index at the current time, and n i N represents the number of cycles for the i-th stress amplitude. fi Let be the cycle life for the i-th stress amplitude.

5. The method for predicting the remaining lifetime of a power device according to claim 1, characterized in that, Step (6) includes: Step (6.1) uses the power device degradation characterization quantity-cumulative damage data under dynamic operating conditions obtained in step (5) as the test sample degradation trajectory dataset. The H+1 nearest data points to the current time in the test sample degradation trajectory data are extracted as test sample degradation trajectory segments, and their data point sequences are as follows: Where H is a non-negative integer, ∆t is the sampling interval of the degenerate trajectory, k is the data point number at the current time, x is the data point of the degenerate characterization quantity, and X is the data point sequence of the degenerate trajectory segment; Step (6.2) extracts any H+1 consecutive data points from the start of operation to the point of failure for each reference sample as a reference sample degradation trajectory segment, and its data point sequence is as follows: Where i is the reference sample index, m is the index of any data point in the reference sample, and H+1≤m≤M i M i This represents the number of data points in reference sample i ranging from completely healthy to completely failed. Step (6.3) involves iteratively matching the trajectory segments of the test sample similarity comparison with any similarity comparison trajectory segment of each reference sample from the start of operation to the point of failure, and calculating the degradation trajectory segment X0(k,H) of the test sample and the degradation trajectory segment X of the reference sample. i Similarity of (m,H); Step (6.4) obtains the reference sample degenerate trajectory segment that has the highest similarity to the test sample degenerate trajectory segment among the reference sample degenerate trajectory segments, and selects the end time N of the trajectory segment that makes the reference sample i most similar to the test sample. i The remaining lifetime at (k)·Δt is used as the estimated remaining lifetime, i.e., RUL. ei (k), where: ； ； N i (k) represents the trajectory segment X that has the highest similarity to the test sample. i The value of m in (m,H); Step (6.5) involves a weighted sum of the remaining lifetime estimates for all reference samples to obtain the remaining lifetime prediction result for the test sample at the current time, expressed as: ； Among them, w i (k) is the weighting factor, and L is the total number of reference samples.

Images