An IGBT aging feature extraction method based on a multi-scale geometric edge sharpening operator

Abstract

The application discloses an IGBT aging feature extraction method based on a multi-scale geometric edge sharpening operator and belongs to the technical field of power electronic device state monitoring. V ce Meanwhile, physical parameters required by the initialization algorithm are initialized; a physically self-adapting flat linear structural element is constructed based on the device physical parameters; a multi-scale regularization morphological gradient spectrum is calculated by introducing a scale regularization factor; a geometric edge sharpening trajectory is generated by deeply suppressing noise and sharpening edges by using a gating effect of a nonlinear cascade product; and a Miller platform duration and a voltage change rate are quantitatively extracted by searching local modulus maximum points of the trajectory to accurately position start and end moments of a voltage rising edge. The extraction method effectively overcomes defects of traditional differential methods, such as amplifying noise and blurring edges by linear filtering, and significantly improves a signal-to-noise ratio and precision of feature extraction.

Description

Technical Field

[0001] This invention belongs to the field of power electronic device condition monitoring technology, specifically relating to an IGBT aging feature extraction method based on a multi-scale geometric edge sharpening operator. Background Technology

[0002] Insulated-gate bipolar transistors (IGBTs) are critical power devices in core equipment such as offshore wind power converters, electric vehicles, and rail transportation. In harsh application scenarios such as offshore wind power, IGBT modules are subjected to the impact of wind loads from frequency and amplitude changes over long periods, making them highly susceptible to aging failure mechanisms such as bond wire detachment and solder layer fatigue. If their health status is not monitored in a timely manner, device failure may lead to converter shutdown or even trigger cascading failures, resulting in high operation and maintenance costs.

[0003] Existing research indicates that the collector-emitter voltage (Vc) during the IGBT turn-off transient process... ce The waveform contains a wealth of aging information. In particular, the Miller plateau duration (t) gp The gate oxide capacitance and stray inductance parameters (dv / dt) are directly related to the gate oxide capacitance and stray inductance parameters inside the device, and are sensitive indicators of the health status of the IGBT.

[0004] However, in actual offshore wind power operations, achieving accurate extraction of the aforementioned microsecond-level transient characteristics faces severe challenges, and existing signal processing technologies are insufficient to meet the following engineering requirements: 1) Noise overwhelming problem caused by strong electromagnetic interference: Offshore wind power converters operate in a high-voltage, high-current, high-frequency switching state (usually at the kHz level), and the switching action will generate an extremely high voltage change rate ( dv / dt ) and rate of change of current ( di / dt This induces significant broadband electromagnetic noise (EMI) in the signal acquisition circuit; this strong background noise is often superimposed on... V ce The waveform masks subtle aging characteristics.

[0005] 2) Limitations of existing differentiation methods: Traditional feature extraction methods often use numerical differentiation to find the inflection points of voltage waveforms (to determine the start and end times of the Miller plateau). However, numerical differentiation is essentially a high-pass filter, which is extremely sensitive to high-frequency noise. In noisy environments, differentiation operations will significantly amplify the noise signal, causing the real feature inflection points to be submerged by spurious noise peaks, making accurate time positioning impossible.

[0006] 3) The contradictions of traditional filtering methods: To suppress noise, existing technologies often employ linear filtering techniques such as low-pass filtering or wavelet transform. While these methods can smooth noise to some extent, they are often accompanied by an "edge blurring" effect, causing the steep geometric edges in the IGBT turn-off waveform to become gentler, resulting in critical time characteristics (such as...). t gp Distortion or decreased positioning accuracy may occur. Furthermore, methods such as wavelet transform rely on the selection of basis functions and lack physical adaptability to the specific waveform topology of the IGBT.

[0007] Therefore, there is an urgent need to design an extraction method that can both deeply suppress strong background noise and maintain or even sharpen the geometric edge features of the signal, so as to solve the technical problem of "not being able to see clearly or accurately extract" the weak aging features of IGBTs in a noisy environment. Summary of the Invention

[0008] The purpose of this invention is to address the problem of inaccurate extraction results in existing methods for extracting features characterizing the health status of IGBTs, and to provide an IGBT aging feature extraction method based on a multi-scale geometric edge sharpening operator. This method utilizes physically adaptive structuring elements and nonlinear cascade operations to achieve high signal-to-noise ratio extraction of IGBT aging features under strong noise conditions.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: an IGBT aging feature extraction method based on multi-scale geometric edge sharpening operators, comprising the following steps: S1. Data Acquisition and Physical Parameter Initialization: The discrete sequence of collector-emitter voltage of the IGBT module during the turn-off transient process is obtained using a high-speed data acquisition system. V ce Simultaneously initialize the physical parameters required by the algorithm, including the sampling frequency. f sample Switching frequency f sw and the typical value of Miller plateau duration in the device under test datasheet. τ ; S2. Construct physically adaptive structural elements: Establish a mapping relationship from the device physical time domain to the discrete sampling domain, and calculate the reference scale of the structural elements based on the physical parameters in step S1 above. L base Based on this, a set of multi-scale flat linear structure elements with linearly increasing scale is generated. B ; The multi-scale flat linear structural element group is represented as: B ={ B 1, B 2,..., BN}, where, N Let be the total number of scales for decomposition; where the th is the th . s Structural elements at various scales B s length L s satisfy: L s =(2 s -1)⋅ L base ; S3. Calculate the multi-scale regularized morphological gradient spectrum: Using the first step S2 above s Structural elements at various scales B s The discrete sequence of collector-emitter voltages acquired in step S1 above V ce Multiscale expansion and erosion calculations are performed, and a scale regularization factor is introduced. λ s The calculation results are weighted and corrected to generate a multi-scale regularized morphological gradient spectrum; No. s Regularized morphological gradients at various scales G s ( t The expression for ) is: In the formula, the scaling regularization factor is... λ s =1 / L s , D s ( t The result of the expansion operation is denoted as . E s ( t The result is the corrosion calculation result. S4. Generate geometric edge sharpening trajectory: Introducing numerical stability factor ϵ The multi-scale regularized morphological gradient spectrum generated in step S3 is subjected to nonlinear cascaded product operation and amplitude normalization, and the multiplication gating effect is used to suppress background noise, generating a high signal-to-noise ratio geometrically sharpened trajectory MGES. t ); Geometric Edge Sharpening Trajectory (MGES) t The expression for ) is: , , In the formula, The geometric mean of the multi-scale regularized morphological gradient. and Sequences The minimum and maximum values ​​in; This indicates a series of multiplication operations. N The total number of scales for decomposition. ϵ This is the numerical stability factor, with a value of 10. -6 ; S5. Quantification of aging characteristics: S51, Peak Positioning: The geometric edge sharpening trajectory MGES generated in step S4 above ( t Perform a local modulus maxima search in ) S52. Temporal calibration: The time corresponding to the first significant local modulus maxima in the geometric edge sharpening trajectory is marked as the starting time of the Miller plateau. t start This characterizes the IGBT exiting the saturation region; the time corresponding to the second significant local modulus maxima or gradient fallback point in the subsequent geometric edge sharpening trajectory is marked as the Miller plateau termination time. t end This indicates the end of Miller capacitor charging; S53. Timing Feature Calculation: Calculate the time difference between the two to obtain the Miller platform duration. t gp : ; S54. Calculation of the rate of change characteristic: Index of original collector-emitter voltage discrete sequence V ce At the start of the Miller platform t start And Miller platform end time t end The voltage amplitude is used to calculate the collector-emitter voltage change rate using the differential ratio method. dv ce / dt : , In the formula, for V ce At the start of the Miller platform t end voltage amplitude, for V ce At the start of the Miller platform t start The voltage amplitude.

[0010] In step S2, the reference scale L base The calculation formula is: In the formula, η This is an adaptive adjustment coefficient, and its value range is 0.1 ≤ η ≤0.3; ⌊⋅⌋ indicates the floor function.

[0011] In step S3, the expansion operation result D s ( t The formula for calculating ) is: In the formula, This represents the morphological dilation operator; the result of the erosion operation. E s ( t The formula for calculating ) is: In the formula, This represents the morphological erosion operator.

[0012] The beneficial effects of this invention are: 1) The method of this invention addresses the problem that the transient aging characteristics of IGBT turn-off under strong electromagnetic interference are easily submerged by noise. By using physical adaptive structural elements and nonlinear cascade operations, it achieves high signal-to-noise ratio extraction of IGBT aging characteristics under strong noise environment. It effectively overcomes the defects of traditional differential method amplifying noise and linear filtering blurring edges, and significantly improves the signal-to-noise ratio and accuracy of feature extraction.

[0013] 2) In the multi-scale expansion and erosion operation, the method of the present invention introduces a scale regularization factor to eliminate the gradient amplitude imbalance caused by the difference in the geometric size of the structuring elements, thereby achieving the purpose of eliminating amplitude differences and ensuring that the gradient response at different scales depends only on the waveform characteristics of the signal itself, providing a unified amplitude benchmark for subsequent nonlinear fusion.

[0014] 3) When performing nonlinear cascaded product operations, the method of this invention utilizes the "gating effect" of multiplication. Since random noise usually only responds at small scales, while the gradient value approaches zero at large scales, the multiplication operation will use the "zero value" at large scales to suppress the entire noise term. Conversely, the true rising edge of the IGBT voltage exhibits a significant gradient peak at all scales, and the multiplication operation will exponentially enhance its energy, thereby greatly improving the signal-to-noise ratio. Attached Figure Description

[0015] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a schematic diagram of the physically adaptive structural elements constructed in an embodiment of the present invention; Figure 3 This is a comparison chart showing the effects of the method of the present invention on processing noisy waveforms with the traditional differential method; Figure 4This is a diagram illustrating the multi-scale decomposition and fusion process in an embodiment of the present invention. Figure 5 This is a schematic diagram illustrating the principle of extracting Miller platform features based on sharpened trajectory in an embodiment of the present invention. Detailed Implementation

[0016] The present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments.

[0017] Example: Figure 1 The flowchart shown above is based on the principle of using physical parameter constraints to construct adaptive structural elements and using nonlinear operations in multi-scale space to achieve "denoising" and "sharpening" of weak aging features.

[0018] This invention provides a method for extracting IGBT aging features based on a multi-scale geometric edge sharpening operator, comprising the following steps: S1. Data Acquisition and Physical Parameter Initialization: Using a high-speed data acquisition system (such as a high-precision oscilloscope or DAQ card), the discrete sequence of collector-emitter voltage of the IGBT module during the turn-off transient process is acquired in real time. V ce [ n In order to capture transient changes at the microsecond level, the sampling frequency... f sample Set to 100MHz and above.

[0019] Simultaneously initialize the physical parameters required by the algorithm, including the sampling frequency. f sample Switching frequency f sw and the typical value of Miller plateau duration in the device under test datasheet. τ Typically 1 to 5 μs .

[0020] S2. Construct physically adaptive structural elements: This step aims to construct a set of geometric probes capable of "sensing" specific physical processes of IGBTs. The specific construction logic is as follows: Figure 2 As shown; Structural element definition: Unlike traditional two-dimensional image processing, this embodiment uses a flat linear structural element to address the one-dimensional time series characteristics of IGBT voltage. s Structural elements at various scales B s Defined as an amplitude of 0 and a length of L s A one-dimensional discrete set, mathematically represented as This "zero-height" characteristic ensures that the operation only performs sliding window statistics in the time dimension without changing the original amplitude of the signal.

[0021] like Figure 2 As shown in figure a, a mapping relationship is established from the physical time domain to the discrete sampling domain, and the scale is calculated using the physical parameter constraint formula. L base ; benchmark scale L base The calculation formula is: In the formula, η This is an adaptive adjustment coefficient, and its value range is 0.1 ≤ η ≤0.3; ⌊⋅⌋ indicates floor function. This formula ensures that the structuring element size is both smaller than the Miller plateau width to preserve features and larger than the high-frequency noise pulse width to achieve filtering.

[0022] like Figure 2 As shown in b, the reference size of the structural element is calculated based on the physical parameters in step S1 above. L base Based on this, a set of multi-scale flat linear structure elements with linearly increasing scale is generated. B ; The multi-scale flat linear structural element group is represented as: B ={ B 1, B 2,..., B N}, where, N Let be the total number of scales for decomposition; where the th is the th . s Structural elements at various scales B s length L s satisfy: L s =(2 s -1)⋅ L base A set of detection sequences, from narrow to wide, is formed to cover different geometric features ranging from local details to macroscopic trends.

[0023] S3. Calculate the multi-scale regularized morphological gradient spectrum: Using the first step S2 above s Structural elements at various scales B s The discrete sequence of collector-emitter voltages acquired in step S1 above V ce Perform parallel multi-scale morphological processing.

[0024] Basic operations: For the first s Multi-scale dilation and erosion operations are performed at each scale. The dilation operation results are... Ds ( t The formula for calculating ) is: In the formula, The morphological dilation operator, in the formula, represents the use of structuring elements. B s Discrete sequence of collector-emitter voltage V ce Perform dilation operation; result of erosion operation. E s ( t The formula for calculating ) is: In the formula, This represents the morphological erosion operator; in the formula, it signifies the use of structuring elements. B s Discrete sequence of collector-emitter voltage V ce Perform the erosion operation.

[0025] Scale regularization: To eliminate the difference in gradient magnitude gain caused by structuring elements of different lengths (i.e., large-scale structuring elements produce larger gradient responses, thus masking small-scale details), this invention introduces a scale regularization factor. λ s Scale regularization factor λ s =1 / L s .

[0026] Gradient spectrum generation: The computation results are weighted and corrected to generate a multi-scale regularized morphological gradient spectrum; s Regularized morphological gradients at various scales G s ( t The expression for ) is: , By using regularization, the "total gradient energy" is converted into "gradient density per unit scale", ensuring that the gradient response at different scales depends only on the waveform characteristics of the signal itself.

[0027] S4. Generate geometric edge sharpening trajectory: By leveraging the noncorrelation of noise at different scales and the strong correlation of signals at different scales, a nonlinear cascaded product operation is performed on the gradient spectrum. The principle is as follows: Figure 4 As shown, the specific calculations are as follows: Cascaded products (gated denoising): Introducing a numerical stability factor ϵ The value is 10. -6 ; Calculate the product of gradients at each scale: In the formula, The geometric mean of the multi-scale regularized morphological gradient. This indicates a series of multiplication operations. N The total number of scales for decomposition.

[0028] This step utilizes the "gating effect" of multiplication. For example... Figure 4 As shown, for regions with random noise distribution (i.e., regions represented by disordered pulses in the figure), it typically only occurs at small scales (such as...). s There is a response at =1), but at large scales (e.g., s =3) The downgradient value approaches zero. The multiplication operation uses the large-scale "zero value" to suppress the entire noise term (a veto, as shown by the cross mark "×" in the figure); conversely, for the true rising edge region of the IGBT voltage (i.e., the region where the voltage steps in the figure), it exhibits significant gradient peaks at all scales. This multi-scale consistent triggering gating is turned on (as shown by the circle mark "○" in the figure), and the multiplication operation exponentially increases its energy.

[0029] Amplitude normalization: To facilitate subsequent unified threshold processing, a normalization operator is used. Map the calculation results to QUOTE The interval is used to obtain the final geometric edge sharpening trajectory MGES. t ); Geometric Edge Sharpening Trajectory (MGES) t The expression for ) is: In the formula, and Sequences The minimum and maximum values ​​in.

[0030] like Figure 3 The comparison images shown are for reference only. Figure 3 (a) shows the original noisy IGBT turn-off waveform. V ce ; Figure 3 (b) demonstrates the extraction results of the existing technique (traditional differential method). In noisy environments, the noise is significantly amplified, forming dense high-amplitude pseudo-peaks, which completely obscure the true features; while Figure 3 (c) shows the extraction results of the MGES operator of the present invention. The background noise is deeply suppressed to the zero basis, and the start and end points of the Miller plateau are two clear and independent sharp main peaks, which significantly improves the signal-to-noise ratio of feature recognition.

[0031] S5. Quantification of aging characteristics: Based on the generated MGES ( tThe trajectory can be accurately located using a simple peak search algorithm, the principle of which is as follows: Figure 5 As shown: S51, Peak Positioning: The geometric edge sharpening trajectory MGES generated in step S4 above ( t Perform a local modulus maxima search in ).

[0032] S52. Temporal calibration: The time corresponding to the first significant local modulus maxima in the geometric edge sharpening trajectory is marked as the starting time of the Miller plateau. t start This characterizes the IGBT exiting the saturation region; the time corresponding to the second significant local modulus maxima or gradient fallback point in the subsequent geometric edge sharpening trajectory is marked as the Miller plateau termination time. t end This indicates the end of Miller capacitor charging; Locating the start time t start ( Figure 5 b point A): Search MGES ( t The first significant local modulus maxima (main peak) in the trajectory, which physically corresponds to the collector-emitter voltage. V ce The rapid rise indicates that the IGBT has exited the saturation conduction region and entered the amplification region (e.g., Figure 5 (as shown in a).

[0033] Location end time t end ( Figure 5 b point B): Search MGES ( t The second significant local maximum of the modulus in the trajectory (or the moment when the gradient rapidly falls back to zero) physically corresponds to... V ce Reaching the bus voltage signifies the end of Miller capacitor charging, and the IGBT enters the cutoff region.

[0034] S53. Timing Feature Calculation: Calculate the time difference between the two to obtain the Miller platform duration. t gp : .

[0035] S54. Calculation of the rate of change characteristic: Index of original collector-emitter voltage discrete sequence V ce At the start of the Miller platform t start And Miller platform end time t end The voltage amplitude is calculated, the slope is determined, and the collector-emitter voltage change rate is calculated using the differential ratio method. dvce / dt : In the formula, for V ce At the start of the Miller platform t end voltage amplitude, for V ce At the start of the Miller platform t start The voltage amplitude.

[0036] This method uses geometric features to lock the computation interval, making dv / dt The calculation focuses on the actual voltage rise time, avoiding the noise amplification problem caused by the differentiation throughout the process.

[0037] This invention addresses the problem that the transient aging features of IGBTs under strong electromagnetic interference are easily submerged by noise. By utilizing physical adaptive structural elements and nonlinear cascade operations, it achieves high signal-to-noise ratio extraction of IGBT aging features under strong noise conditions. This effectively overcomes the shortcomings of traditional differential methods that amplify noise and linear filtering that blurs edges, and significantly improves the signal-to-noise ratio and accuracy of feature extraction.

[0038] The above description is only used to illustrate the technical solution of the present invention and is not intended to limit it. Any other modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention, as long as they do not depart from the spirit and scope of the technical solution of the present invention, should be covered within the scope of the claims of the present invention.

Claims

1. A method for extracting IGBT aging features based on multi-scale geometric edge sharpening operators, characterized in that: Includes the following steps: S1. Data Acquisition and Physical Parameter Initialization: The discrete sequence of collector-emitter voltage of the IGBT module during the turn-off transient process is obtained using a high-speed data acquisition system. V ce Simultaneously initialize the physical parameters required by the algorithm, including the sampling frequency. f sample Switching frequency f sw and the typical value of Miller plateau duration in the device under test datasheet. τ ; S2. Construct physically adaptive structural elements: Establish a mapping relationship from the device physical time domain to the discrete sampling domain, and calculate the reference scale of the structural elements based on the physical parameters in step S1 above. L base Based on this, a set of multi-scale flat linear structure elements with linearly increasing scale is generated. B ; The multi-scale flat linear structural element group is represented as: B ={ B 1, B 2,..., B N }, where, N The total number of scales for decomposition; Among them, the first s Structural elements at various scales B s length L s satisfy: L s =(2 s -1)⋅ L base ; S3. Calculate the multi-scale regularized morphological gradient spectrum: Using the first step S2 above s Structural elements at various scales B s The discrete sequence of collector-emitter voltages acquired in step S1 above V ce Multiscale expansion and erosion calculations are performed, and a scale regularization factor is introduced. λ s The calculation results are weighted and corrected to generate a multi-scale regularized morphological gradient spectrum; No. s Regularized morphological gradients at various scales G s ( t The expression for ) is: In the formula, the scaling regularization factor is... λ s =1 / L s , D s ( t The result of the expansion operation is denoted as . E s ( t The result is the corrosion calculation result. S4. Generate geometric edge sharpening trajectory: Introducing numerical stability factor ϵ The multi-scale regularized morphological gradient spectrum generated in step S3 is subjected to nonlinear cascaded product operation and amplitude normalization, and the multiplication gating effect is used to suppress background noise, generating a high signal-to-noise ratio geometrically sharpened trajectory MGES. t ); Geometric Edge Sharpening Trajectory (MGES) t The expression for ) is: ， ， In the formula, The geometric mean of the multi-scale regularized morphological gradient. and Sequences The minimum and maximum values ​​in; This indicates a series of multiplication operations. N The total number of scales for decomposition. ϵ It is the numerical stability factor; S5. Quantitatively extract aging characteristics: S51, Peak Positioning: The geometric edge sharpening trajectory MGES generated in step S4 above ( t Perform a local modulus maxima search in ) S52. Temporal calibration: The time corresponding to the first significant local modulus maxima in the geometric edge sharpening trajectory is marked as the starting time of the Miller plateau. t start This characterizes the IGBT exiting the saturation region; the time corresponding to the second significant local modulus maxima or gradient fallback point in the subsequent geometric edge sharpening trajectory is marked as the Miller plateau termination time. t end This indicates the end of Miller capacitor charging; S53. Timing Feature Calculation: Calculate the time difference between the two to obtain the Miller platform duration. t gp : ; S54. Calculation of the rate of change characteristic: Index of original collector-emitter voltage discrete sequence V ce At the start of the Miller platform t start And Miller platform end time t end The voltage amplitude is used to calculate the collector-emitter voltage change rate using the differential ratio method. dv ce / dt : ， In the formula, for V ce At the start of the Miller platform t end voltage amplitude, for V ce At the start of the Miller platform t start The voltage amplitude.

2. The IGBT aging feature extraction method based on multi-scale geometric edge sharpening operator according to claim 1, characterized in that: In step S2, the reference scale L base The calculation formula is: In the formula, η For adaptive adjustment coefficients; ⌊⋅⌋ indicates floor function.

3. The IGBT aging feature extraction method based on multi-scale geometric edge sharpening operators according to claim 1, characterized in that: In step S3, the expansion operation result D s ( t The formula for calculating ) is: In the formula, This represents the morphological dilation operator; the result of the erosion operation. E s ( t The formula for calculating ) is: In the formula, This represents the morphological erosion operator.

Images