Method for diagnosing early degradation of ultrasonic power supply power device

Abstract

The present application relates to an early degradation diagnosis method for ultrasonic power supply power devices, aiming at the problem of insufficient accuracy and timeliness of traditional insulation deterioration and local short circuit precursor detection, a multi-modal feature diagnosis method is proposed, which fuses cold-state resistance mutation, high-speed gray space evolution and near-infrared spectrum response. The method uses synchronous high-speed current sensing, image capture and near-infrared spectrum acquisition to construct the original multi-dimensional sensing data set, through time domain slope analysis, space brightness diffusion and spectrum peak correction, the insulation deterioration and heat accumulation characteristics are extracted, and through the lightweight decision tree model, the four-level degradation state classification is realized, the graded early warning and closed-loop risk control are completed. The system has the advantages of strong environmental adaptability, intuitive state criterion and high diagnosis precision, and can realize early identification of insulation problems such as gate oxide layer micro-breakdown, effectively improving the safety and reliability of power device operation.

Description

Technical Field

[0001] This invention relates to the field of early degradation diagnosis of power devices and dynamic detection of cold resistance of incandescent lamps, and in particular to a method for early degradation diagnosis of ultrasonic power devices. Background Technology

[0002] Currently, in the field of early degradation diagnosis of ultrasonic power supply PCBAs and their power devices, series-connected incandescent lamps are commonly used as current limiting and status indication elements. Traditional current limiting systems widely utilize the luminous intensity of incandescent lamps under steady-state power-on conditions to roughly distinguish between normal and abnormal currents, but this is limited to judging macroscopic phenomena such as single on / off cycles or decreased brightness. Current mainstream technologies typically infer power device faults by observing the light intensity of the incandescent lamp during its continuous illumination phase, component temperature rise, or circuit current jumps. Some solutions introduce photoelectric sensors to assist in automatic detection, enabling automatic identification of typical faults such as obvious short circuits and open circuits. This is widely used in actual production line maintenance and equipment screening, and has advantages such as simple structure, low cost, and intuitive response.

[0003] As research into the failure mechanisms of power semiconductors deepens, the industry trend is shifting from traditional equipment failure detection to early microscopic degradation diagnosis, with a particular focus on monitoring potential hazards in sub-optimal operating conditions such as gate oxide micro-breakdown, body diode leakage, and insulation aging. Some high-end diagnostic solutions are attempting to expand signal acquisition dimensions, such as using LED photoelectric sensors to replace incandescent lamps for wider dynamic range detection, or coupling multi-channel current, voltage, temperature, and light signals, using fuzzy comprehensive evaluation and neural network modeling to improve the ability to identify complex anomalies. Furthermore, advanced methods such as distributed fiber optic temperature measurement, remote dimming control, constant current source reconstruction, and standard fault feature database retrieval are also being explored in large-scale automated production lines. While these methods are technologically advanced, they generally suffer from limitations such as high cost, complex structure, insufficient environmental adaptability, and poor real-time performance, making it difficult to simultaneously meet the economic and response speed requirements of factory-level batch screening.

[0004] Existing technologies generally rely on qualitative signals such as steady-state brightness or lighting delay of incandescent lamps, and are severely inadequate in identifying early degradation of power devices, especially sub-healthy states such as microampere-level leakage and initial insulation performance deterioration. Specifically, the brightness-off threshold judgment can only detect obvious short circuits, open circuits, and overcurrents, lacking high-resolution quantitative analysis methods for the fine resistance and luminous characteristics of incandescent lamps during cold start-up. Therefore, when faced with power devices with gate oxide damage or slight insulation performance degradation, it is often impossible to distinguish the degradation level in a timely and accurate manner using traditional methods, making it difficult to detect microscopic hidden dangers early and seriously affecting preventive maintenance and safe and reliable operation of equipment.

[0005] Furthermore, for the identification of multi-level degradation states, there is currently no innovative method in the industry to deeply mine the nonlinear abrupt change signal of the cold-state resistance of incandescent lamps for multi-physical quantity coupling modeling. Existing technical approaches based on signal fusion, optoelectronic sensor networks, or AI discrimination models mostly rely on complex sensor arrays, precision optical filtering systems, or historical experience feature extraction, which not only increases the difficulty of system construction and the threshold for use, but also limits its promotion in embedded platforms and low-cost scenarios.

[0006] In summary, a key technical challenge in the early diagnosis of ultrasonic power supply PCBAs is how to move beyond the traditional reliance on the steady-state brightness of incandescent lamps and fully explore their microscopic physical characteristics, such as dynamic resistance, spatial grayscale diffusion, and infrared spectral drift during the initial cold-state power-on phase. This involves effectively mapping these characteristics to high-precision classification of power devices across multiple degradation states, from normal operation to mild insulation degradation, micro-breakdown, and local short circuits. A novel, high-resolution, low-cost, real-time, and embeddable multi-dimensional physical quantity joint criterion identification method is urgently needed to fill the gap in graded early warning technology from initial microscopic symptoms to critical faults, thereby promoting proactive equipment maintenance and intelligent manufacturing. Summary of the Invention

[0007] This application provides a method for diagnosing early degradation of ultrasonic power devices, aiming to solve one of the problems or issues of the prior art mentioned in the background section above.

[0008] The method for diagnosing early degradation of ultrasonic power devices provided in this application specifically includes: S1: Within the transient window of the ultrasonic power supply being connected to high voltage, the instantaneous current waveform data of the circuit is synchronously acquired using a high-speed current sensor, and a high frame rate image sensor is triggered to capture the grayscale sequence and near-infrared spectral response data of the incandescent lamp filament region, so as to obtain the original multidimensional sensing dataset containing information on the cold resistance jump.

[0009] S2: Based on the original multidimensional sensing dataset, the slope of the instantaneous current waveform data within the time window of zero to twenty milliseconds after power-on is calculated to extract the initial resistance change rate characteristic quantity that characterizes the resistance temperature coefficient response of the tungsten filament in the low-temperature region.

[0010] S3: Using the initial resistance change rate characteristic as an input variable, and combining the physical law of transient current disturbance caused by microampere leakage current, a cold resistance mutation response model is constructed to map the degree of insulation degradation of power devices, so as to generate a primary degradation criterion reflecting the risk of micro-breakdown of the gate oxide layer.

[0011] S4: For the pixel-level brightness evolution data in the grayscale sequence during the 20 to 100 millisecond steady-state pre-state stage, perform a spatial non-uniformity index calculation operation to generate a brightness diffusion rate feature vector that characterizes the degree of local Joule heat accumulation and the difference between single-point micro-short circuits and planar insulation aging.

[0012] S5: Based on the near-infrared spectral response data, the spectral peak position shift in the 700-nanometer to 950-nanometer band is calibrated to obtain the equivalent power consumption increment of the inversion loop and generate a correction factor for the average temperature rise rate of the tungsten wire that excludes environmental temperature drift and voltage fluctuation interference.

[0013] S6: The primary degradation criterion, the brightness diffusion rate feature vector, and the tungsten filament average temperature rise rate correction factor are fused into a three-dimensional joint state feature set and input into a lightweight decision tree model trained offline on classic failure samples to output early degradation diagnosis results of power devices containing four levels of degradation state labels.

[0014] S7: Based on the level values ​​in the four-level degradation status labels, perform a graded early warning strategy matching operation. If it is determined to be a near short circuit state, an emergency cut-off command is triggered. If it is determined to be a state of slight decrease in insulation performance or damage to the gate structure, a maintenance prediction log is generated to form a closed-loop risk control action.

[0015] S8: Based on the feedback of the deviation between the maintenance prediction log and the final diagnosis result, dynamically update the threshold parameter group of the lightweight decision tree model to optimize the subsequent identification accuracy of the multi-level degradation state of ultrasonic power devices from insulation deterioration to local short circuit precursors.

[0016] The method for diagnosing early degradation of ultrasonic power devices provided in this application has the following beneficial effects: (1) The method proposed in this application, by deeply exploring the dynamic response characteristics of electrothermal-optical coupling during the transient process of incandescent lamp power-on, overcomes the limitations of traditional techniques that rely solely on steady-state luminous intensity or average current for fault diagnosis, and significantly improves the sensitivity and diagnostic accuracy for the early degradation state of ultrasonic power supply PCBA. Because the tungsten filament has an extremely high temperature coefficient of resistance (approximately 4.5 × 10⁻⁶) in the cold state... -Even if power devices experience tiny leakage currents in the nA range due to gate oxide defects, this will still cause an observable step shift in the initial current slope. Combined with high-sampling-rate current and high-speed optical synchronous acquisition, non-invasive capture of leakage behavior below the microampere level can be achieved. Compared to the voltage threshold comparison or average power consumption monitoring methods commonly used in existing technologies, this solution improves fault detection sensitivity to the sub-microampere level without adding dedicated sensing elements. This effectively overcomes the problem of traditional methods failing to identify "hidden faults" such as early insulation degradation and the incipient stage of local breakdown, providing a more proactive and reliable basis for production line quality control and equipment life prediction.

[0017] (2) By constructing a multi-dimensional dynamic feature fusion mechanism—including the cold resistance change rate, the gray diffusion rate and spatial non-uniformity index of the filament region, and the temperature rise rate characterized by the near-infrared spectral peak position shift—this scheme achieves a refined ability to distinguish fault types and severity. Among them, the spatial evolution mode of brightness diffusion can effectively identify the Joule heat accumulation pattern, thereby distinguishing between single-point micro-short circuits and planar aging; while the spectral shift is used to invert the equivalent power consumption increment of the loop, enhancing diagnostic robustness while eliminating interference from environmental temperature drift and input voltage fluctuations. The above features are integrated and analyzed by a lightweight decision tree model, outputting a four-level graded status label, covering the complete degradation path from normal to near short circuit, with good interpretability and engineering practicality. Compared with complex algorithm systems that rely on fuzzy comprehensive evaluation or multi-sensor data fusion modeling, this method does not require remote communication, database matching, or manual parameter tuning. All calculations can be completed in real time in the embedded MCU, with a diagnostic latency of less than 200ms, meeting the needs of rapid screening and immediate early warning in industrial sites, and greatly improving detection efficiency and system response sensitivity.

[0018] (3) This method innovatively reconstructs the incandescent lamp from a traditional indicator element into a passive physical sensor unit integrating "excitation-sensing-feedback", making full use of its inherent electrothermal-optical response mechanism and avoiding the cost increase and structural complexity problems caused by introducing LEDs to replace the light source, distributed fiber optic temperature measurement, constant current source reconstruction, or additional current sampling circuits. The entire diagnostic process does not require a complex optical filtering system or multi-source synchronous calibration. It can achieve high-precision sensing by relying solely on low-cost high-speed CMOS and current sensors, and has good deployability and production line compatibility. More importantly, this technical path completely avoids the external intervention dependence common in current mainstream solutions, such as remote dimming control, fault sample database dependence, and fuzzy inference rule setting, forming a native innovation closed loop based on physical essence. This not only enhances the autonomy and stability of the system, but also provides a transferable technical paradigm for the early fault diagnosis of other power electronic devices.

[0019] In summary, the method proposed in this application constructs a highly sensitive, low-latency, and robust early degradation diagnosis system by mining the multi-physics dynamic response information of incandescent lamps during the power-on transient process. It achieves effective capture and accurate classification of weak fault signals of ultrasonic power supply PCBA without increasing hardware costs. It has both technological advancement and engineering application value, and is especially suitable for industrial power supply maintenance scenarios with stringent reliability requirements, showing significant prospects for widespread application. Attached Figure Description

[0020] Figure 1 This is the main flowchart of the diagnostic method for early degradation of ultrasonic power devices; Figure 2 This is a sub-flowchart of the diagnostic method for early degradation of ultrasonic power devices. Figure 3 This is another sub-flowchart of the method for early degradation diagnosis of ultrasonic power devices. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0023] like Figure 1 As shown, this application provides a method for diagnosing early degradation of ultrasonic power devices, specifically including: S1: Within the transient window of the ultrasonic power supply being connected to high voltage, the instantaneous current waveform data of the circuit is synchronously acquired using a high-speed current sensor, and a high frame rate image sensor is triggered to capture the grayscale sequence and near-infrared spectral response data of the incandescent lamp filament region, so as to obtain the original multidimensional sensing dataset containing information on the cold resistance jump.

[0024] S2: Based on the original multidimensional sensing dataset, the slope of the instantaneous current waveform data within the time window of zero to twenty milliseconds after power-on is calculated to extract the initial resistance change rate characteristic quantity that characterizes the resistance temperature coefficient response of the tungsten filament in the low-temperature region.

[0025] S3: Using the initial resistance change rate characteristic as an input variable, and combining the physical law of transient current disturbance caused by microampere leakage current, a cold resistance mutation response model is constructed to map the degree of insulation degradation of power devices, so as to generate a primary degradation criterion reflecting the risk of micro-breakdown of the gate oxide layer.

[0026] S4: For the pixel-level brightness evolution data in the grayscale sequence during the 20 to 100 millisecond steady-state pre-state stage, perform a spatial non-uniformity index calculation operation to generate a brightness diffusion rate feature vector that characterizes the degree of local Joule heat accumulation and the difference between single-point micro-short circuits and planar insulation aging.

[0027] S5: Based on the near-infrared spectral response data, the spectral peak position shift in the 700-nanometer to 950-nanometer band is calibrated to obtain the equivalent power consumption increment of the inversion loop and generate a correction factor for the average temperature rise rate of the tungsten wire that excludes environmental temperature drift and voltage fluctuation interference.

[0028] S6: The primary degradation criterion, the brightness diffusion rate feature vector, and the tungsten filament average temperature rise rate correction factor are fused into a three-dimensional joint state feature set and input into a lightweight decision tree model trained offline on classic failure samples to output early degradation diagnosis results of power devices containing four levels of degradation state labels.

[0029] S7: Based on the level values ​​in the four-level degradation status labels, perform a graded early warning strategy matching operation. If it is determined to be a near short circuit state, an emergency cut-off command is triggered. If it is determined to be a state of slight decrease in insulation performance or damage to the gate structure, a maintenance prediction log is generated to form a closed-loop risk control action.

[0030] S8: Based on the feedback of the deviation between the maintenance prediction log and the final diagnosis result, dynamically update the threshold parameter group of the lightweight decision tree model to optimize the subsequent identification accuracy of the multi-level degradation state of ultrasonic power devices from insulation deterioration to local short circuit precursors.

[0031] Step S1: Within the transient window of the ultrasonic power supply being connected to high voltage, a high-speed current sensor is used to synchronously acquire the instantaneous current waveform data of the circuit, and a high frame rate image sensor is triggered to capture the grayscale sequence and near-infrared spectral response data of the incandescent lamp filament region, in order to obtain the original multidimensional sensing dataset containing information on the cold-state resistance jump. Specifically, this includes: S1.1: Perform high-precision timing trigger control on the transient window of zero to one hundred milliseconds when the ultrasonic power supply is connected to high voltage, so as to generate a synchronous sampling clock signal containing microsecond-level alignment marks, ensuring a strict correspondence between the subsequent current waveform data and the optical response data on the time axis.

[0032] Within the transient window of zero to one hundred milliseconds when the ultrasonic power supply is connected to high voltage, a unified timing control reference signal is established for the synchronous acquisition requirements of high-speed current sensor and high frame rate image sensor, using the high voltage connection control signal as the initial trigger condition.

[0033] The reference signal is divided into microsecond-level time resolution segments, and a reference clock sequence with a period error of less than ±0.5μs is generated using a high-stability crystal oscillator.

[0034] The reference clock sequence is input to the timing synchronization control module, and the phase lock-in loop (PLL) operation is used to achieve phase locking between the high-speed current sampling channel and the optical sampling channel, so that the sampling time of the two channels has strict synchronization with a phase deviation of no more than one sampling period.

[0035] Based on the synchronous clock output by the PLL, a timestamp embedding operation is further performed to encode the microsecond-level alignment mark into the clock signal stream. The mark format adopts a fixed-length binary bit to ensure that subsequent acquired data can be directly registered on the time axis based on the mark.

[0036] To address potential electromagnetic interference pulses within the transient window, a digital filter is used to suppress interference in the synchronization clock signal, filtering out interference bits with abnormal amplitude within the period to ensure the timing accuracy of synchronous triggering.

[0037] Through the above processing method, the transient conditions of high-voltage access are transformed into a synchronous sampling clock signal containing microsecond-level alignment marks, thereby realizing the correspondence between high-speed current waveform data and optical response data on the time axis.

[0038] For example, in an early degradation diagnostic system for ultrasonic power supply PCBAs, the high-voltage input control signal is optically isolated before entering the timing control unit, which is configured with a crystal oscillator frequency of [frequency value missing]. Hz, output period is A reference clock of μs. The phase-locked loop parameters are set to phase detection sensitivity. Control voltage range V, ensuring that the synchronization error between the current sensing channel and the image acquisition channel is less than 100%. μs. In a synchronous clock stream, each An 8-bit alignment marker is embedded once per cycle, where the first 4 bits are the time window number and the last 4 bits are the cycle count. After the system is connected to 220V, it completes the generation and output of the synchronization marker within 0 to 100 milliseconds. Subsequently, a high-speed current sensor acquires analog current signals at a sampling rate of 1MHz, and an image sensor captures the filament grayscale sequence at 1000fps. Finally, it is verified that the synchronization error is maintained within ±0.8μs, which effectively improves the data registration accuracy of cold resistance jump characteristics.

[0039] S1.2: Based on the synchronous sampling clock signal, a high-speed current sensor with a sampling rate of not less than one megahertz is used to discretize and process the instantaneous analog current input into the incandescent lamp circuit, so as to obtain the original instantaneous current waveform data sequence characterizing the cold resistance jump process of the tungsten filament.

[0040] Based on the synchronous sampling clock signal, the acquisition channel of the high-speed current sensor is initialized to a high-impedance input state to avoid transient impacts causing saturation of the sampling front end, thereby ensuring the linear response range of subsequent acquisitions. The sampling rate is set to no less than... Hertz, based on the microsecond-level alignment mark of the synchronous clock, performs point-by-point triggering, ensuring that the sampling time of each discrete sample corresponds to the absolute time position of the actual loop current waveform. The instantaneous analog current input to the incandescent lamp circuit is quantized by an analog-to-digital converter (ADC), converting the voltage value of each sampling point into the corresponding current value. The conversion formula is as follows: in, This is the instantaneous current value. For sampling voltage, This is the ADC zero bias voltage correction value. The equivalent sampling resistance of the current sensor is used. A dynamic buffer is used to store the acquired instantaneous current values ​​in the sampling order, forming a time series matrix. An amplitude limiting algorithm is executed within the buffer to eliminate abnormal sampling points exceeding the physically possible range. The corrected and time-sequentially arranged instantaneous current sequence is output as raw waveform data, serving as the basic input for subsequent cold-state resistance jump process analysis. Through high-speed discretization acquisition processing, the analog quantity of the loop corresponding to the synchronous sampling clock signal is converted into a rigorous and usable raw instantaneous current waveform data sequence, achieving millisecond-level capture of the tungsten filament cold-state resistance jump process.

[0041] For example, within a transient window of 0 to 100 milliseconds after the ultrasonic power supply is connected to high voltage, the synchronous sampling clock signal is configured with a sampling rate of 1MHz, the ADC resolution is set to 12 bits, the equivalent sampling resistance is 0.1Ω, and the zero-bias voltage correction value is 2.5V. Under this configuration, the analog voltage captured by the sampling channel varies between 2.45V and 2.55V, and the instantaneous current value can be calculated using the formula to be within a certain range. Between 0.5A and 0.5A. For this current value sequence, amplitude limiting is applied within the buffer, and points with an absolute value greater than 2A are marked as abnormal and discarded. Under different degradation states, the current waveform of the tungsten filament in the cold state exhibits different initial slopes. For example, the instantaneous current corresponding to a slight decrease in insulation performance rises at a rate of approximately [missing value] within the first 5ms. The rate corresponding to gate structure damage can reach This sequence is directly output to the cold-state resistance change rate extraction module, significantly improving the accuracy of subsequent degradation state identification.

[0042] S1.3: According to the synchronous sampling clock signal, a high frame rate image sensor with a frame rate of not less than one thousand frames per second is triggered to continuously capture the light radiation field of the incandescent lamp filament area to generate an original grayscale sequence image set that records the evolution process of the filament from a cold state to a hot state.

[0043] Under the condition of a synchronous sampling clock signal input with microsecond-level alignment marks, timing locking processing is performed on the row and column scanning control logic of the high frame rate image sensor to ensure that the frame interval of the incandescent lamp filament area is stable at a standard frame rate of no less than one thousand frames per second. The locked scanning control signal undergoes precise exposure time setting processing, adjusting the exposure time of each frame to the millisecond level to fully capture the rapid changes in the filament light radiation field during the transition from cold to hot states. Pixel-level pre-readout calibration processing is performed on the scanning control signal with the set exposure time to eliminate grayscale shifts caused by sensor dark current and readout noise, and to generate a dynamic gain matrix adapted to the low brightness characteristics of cold-state emission. The dynamic gain matrix is ​​applied to the sensor's real-time grayscale data stream, performing frame-by-frame brightness correction processing, effectively enhancing the weak cold-state emission signal of the filament in the grayscale sequence. Continuous capture operations are performed based on the grayscale data stream after brightness correction. The filament light radiation field frame images arranged in chronological order are packaged into an original grayscale sequence image set. Through this processing method, the high-speed current acquisition results of the previous step are transformed into visual data with optical response dimensions, thereby achieving multimodal capture of the cold resistance transition process.

[0044] For example, within a transient window of 0–100 milliseconds after the ultrasonic power supply is connected to high voltage, a CMOS high frame rate image sensor is triggered at a frame rate of 1250fps to continuously capture images of the incandescent lamp filament area. The exposure time is set to 0.8 milliseconds. Before acquisition, an offset matrix for each pixel is generated by dark current measurement and applied to real-time grayscale stream correction. The dynamic gain matrix is ​​set to 1.5 in the cold phase and smoothly transitions to 1.0 in the hot phase to avoid overexposure. Continuous capture generates 125 frames of grayscale images, each with a resolution of 640×480 pixels. The brightness curve of the filament light radiation center region increases from an initial grayscale value of 120 to a grayscale value of 210. Grayscale histogram analysis of this sequence shows that the brightness increase rate is approximately... The grayscale unit per millisecond significantly improves the signal discernibility during the cold emission stage, providing high-quality input data for subsequent near-infrared spectral extraction and cold resistance rate calculation.

[0045] S1.4: For each frame of the original grayscale sequence image set, the spectral components in the 700-nanometer to 950-nanometer band are separated and extracted using a narrowband filter component integrated in the front end of the image sensor to obtain the original near-infrared spectral response data matrix characterizing the temperature change of the tungsten filament.

[0046] For each frame in the raw grayscale sequence image set generated by the high frame rate image sensor, the pixel matrix containing the incandescent lamp filament region is locked as the target for spectral separation processing.

[0047] Within the locked pixel matrix range, the narrowband filter component integrated into the front end of the image sensor is invoked to separate the radiation energy in the 700-950 nanometer band into independent narrowband spectral signal channels.

[0048] Bandpass filtering is performed on the pixel brightness values ​​in each narrowband spectral signal channel to remove interfering components outside the band and retain the radiation information of the target band, forming a spectral intensity baseline matrix that can be used for temperature inversion.

[0049] Pixel-level normalization is performed based on the spectral intensity baseline matrix to eliminate brightness drift caused by exposure changes between different frames and to unify the spectral intensity dimensions, so as to ensure the reliability of subsequent temperature mapping calculations.

[0050] Using the formula relating peak position to temperature in the blackbody radiation law: in The position of the spectral peak (unit: μm) The absolute temperature of the tungsten filament (in K) is used to calculate the peak position of the main peak pixel in the normalized spectral intensity matrix, convert the peak position result into a temperature value, and construct the original near-infrared spectral response data matrix.

[0051] Through the above processing method, the pixel information of the grayscale sequence image in the previous step is converted into near-infrared spectral response data that can directly characterize the dynamic temperature change of the tungsten filament, thereby realizing the capture of temperature information during the cold resistance transition process.

[0052] For example, in a grayscale sequence output by a high frame rate CMOS image sensor (1200fps), the center wavelength of the narrowband filter is pre-set to cover three channels: 760nm, 830nm, and 900nm, with a bandwidth of ±10nm for each channel. The three-channel spectral intensities obtained after bandpass filtering of the filament region pixel matrix in each frame are normalized to a 0–1 brightness scale. In the 830nm channel of the 15th frame, the position of the peak pixel column intensity maximum corresponds to an actual wavelength of 0.83μm. Substituting this into the formula... Solution =3492K. This value is consistent with the tungsten filament temperature result obtained from the inversion of current data in the same frame, indicating that the spectral response data matrix can accurately reflect the temperature rise characteristics of the tungsten filament in the cold state. In the comparison of different channels, the peak position shift of the 900nm channel is more obvious. The peak position change increases when there is a micro-short circuit in the power device, and the temperature calculation result shows a significant upward trend, verifying the sensitivity of this step in identifying early degradation states.

[0053] S1.5: Perform multimodal spatiotemporal fusion encapsulation on the original instantaneous current waveform data sequence, the original grayscale image set, and the original near-infrared spectral response data matrix to construct an original multidimensional sensing dataset containing complete cold-state resistance jump information, which serves as the unified input object for the subsequent feature extraction module.

[0054] Step S2: Based on the original multidimensional sensing dataset, the slope of the instantaneous current waveform data within a time window of 0 to 20 milliseconds after power-on is calculated to extract the initial resistance change rate characteristic quantity characterizing the resistance temperature coefficient response of the tungsten filament in the low-temperature region. Specifically, this includes: S2.1: Perform timestamp alignment processing on the instantaneous current waveform data collected by the high-speed current sensor in the original multidimensional sensing dataset to lock the zero-time reference point at the moment when the ultrasonic power supply is connected to the high voltage, and generate a synchronized transient current sequence with a unified time reference system.

[0055] The process begins by filtering the instantaneous current waveform data acquired by a high-speed current sensor in the original multidimensional sensing dataset. This filtering removes spurious signals caused by embedded noise or abnormal pulses during sampling, ensuring the stability of the time reference locking process. The filtered instantaneous current waveform data is then processed by a microsecond-precision time synchronization marker parsing module. This module parses the synchronous sampling clock signal generated in the previous main step and establishes an index mapping between sampling points and clock markers. Interpolation correction is performed on the sampling point indices in the mapping table to compensate for time deviations caused by sensor clock drift or sampling jitter, resulting in a corrected sampling time sequence. A zero-time locking algorithm is then applied to the corrected sampling time sequence. By detecting the amplitude jump point at the instant of energization in the current waveform and combining it with the clock marker position, the trigger moment of the physical event of the ultrasonic power supply being connected to high voltage is determined as the zero-time reference point. Finally, a time reference system reconstruction operation is performed on the sampling time sequence after the zero-time reference point. This transforms the timestamps of all sampling points into relative time values ​​with zero time as the origin, ultimately generating a synchronized transient current sequence with a unified time reference system. Through the above chain processing method, the input instantaneous current waveform data is converted into time-aligned data that can be used for dynamic feature extraction in the cold phase, so as to achieve accurate matching between the current signal and the optical signal on the physical event time axis.

[0056] For example, in a scenario for detecting the degradation of an ultrasonic power supply PCBA, the sampling rate of the high-speed current sensor is set to... The period of the synchronous sampling clock signal is A 5th-order Butterworth low-pass filter was selected for filtering, with a cutoff frequency of [missing information]. The interpolation correction uses cubic spline interpolation to address clock drift. Time offset compensation is performed on sampling points within the range. In the zero-time locking algorithm, time offset compensation is performed by detecting changes in the amplitude of the current waveform exceeding a certain threshold. The first sampling point of the noise threshold, combined with the synchronization marker position, determines the reference time at the instant of power-on. Reference frame reconstruction converts all sampling times into relative times starting from time zero, for example, the time after time zero. The relative time of each sampling point is The obtained synchronized transient current sequence is used in conjunction with optical data in subsequent step S2.2 to achieve multimodal feature extraction in the cold state phase, and the output time alignment error is controlled within a certain range. Internally, it can significantly improve the accuracy of cold resistance change rate calculation and the ability to distinguish degradation state.

[0057] S2.2: Based on the synchronized transient current sequence, perform dynamic truncation operation within a time window of 0 to 20 milliseconds to eliminate steady-state thermal effect interference and retain the cold-state stage data with the most significant temperature coefficient response of the tungsten filament in the low-temperature region, thereby generating a cold-state transient current subsequence.

[0058] Based on the time reference frame of the synchronized transient current sequence, a window limit setting operation is performed on the sequence, setting the starting point as the zero-time reference point and limiting the ending point to the twenty-millisecond position to form a dynamically intercepted time boundary condition.

[0059] Under the aforementioned time boundary conditions, the synchronized transient current sequence is indexed and mapped to remove all data segments exceeding twenty milliseconds, thereby eliminating redundant information about resistance changes caused by the steady-state thermal effect of the incandescent lamp tungsten filament.

[0060] Continuous correction calculations are performed on the remaining segments after removing steady-state data. A combination of missing value interpolation and data smoothing filtering is used to ensure the temporal integrity of the current waveform in the cold state phase within the window and the noise suppression effect.

[0061] In the corrected cold-state waveform, dynamic amplitude normalization is performed to eliminate the influence of grid voltage fluctuations on the current amplitude at the moment of power supply connection, and to maintain the true physical proportional relationship of cold-state current changes.

[0062] The amplitude-normalized cold-state stage data is encapsulated in chronological order into a cold-state transient current subsequence, which contains the initial microsecond-level waveform details of the resistance temperature coefficient response in the low-temperature region of the tungsten filament.

[0063] By using the above dynamic interception processing method, the synchronized transient current sequence result of the previous step is transformed into a cold transient current subsequence without steady-state interference, thereby achieving independent data extraction for the cold resistance jump process of tungsten wire.

[0064] For example, in one embodiment, the reference point for the zero-time connection of the ultrasonic power supply to the high-voltage grid is provided by a synchronous sampling system with a timestamp accuracy of 1 μs. Based on this reference point, the starting point of the truncation window is set to 0 ms, and the ending point to 20 ms. During the truncation process, approximately 80% of the steady-state interval data is removed. The truncation data is then smoothed using a five-point moving average filter with a filter coefficient of 0.2 to ensure that the complete waveform slope is preserved even under microampere-level leakage current disturbances. In the amplitude normalization stage, the normalization coefficient is based on the instantaneous voltage of the power grid. The ratio is adjusted, where 220 is the current average sampled voltage and 230 is the standard rated voltage value. The normalization formula is: , The average voltage within the sampling window. The processed cold transient current subsequence has a length of 20,000 points (sampling rate 1MHz). In subsequent S2.3, the derivative calculation of this subsequence can significantly improve the resolution of leakage current slope detection, avoid feature masking caused by steady-state thermal effects, and achieve high sensitivity capture of insulation degradation state.

[0065] S2.3: The cold transient current subsequence is processed by discrete derivative calculation using the sliding difference algorithm to capture the slope change of transient current disturbance caused by microampere leakage current and generate an instantaneous current slope vector containing resistance jump information.

[0066] Using the cold transient current subsequence generated in the previous step as input, a sliding difference algorithm is employed to calculate the discrete derivative. Specifically, a sliding window with a width of N sampling points (N being a positive integer greater than 1, such as 3 or 5) is set, and this window slides sequentially along the time axis starting from the beginning of the sequence with a single-point step size. At each position of the window, the difference between the current value of the last sampling point and the current value of the first sampling point within the sliding window is calculated, and then divided by the corresponding time interval to obtain the instantaneous current change rate at the center of the window (i.e., the first-order forward difference approximation derivative). By traversing the entire subsequence, the instantaneous change rates calculated from all windows are arranged in chronological order, thus generating an instantaneous current slope vector. This vector, by focusing on the rate of change of current over time rather than the absolute amplitude, significantly amplifies the weak slope anomaly hidden in the original waveform caused by microampere-level leakage current disturbances within the power device, thereby highlighting the differential information of resistance jumps characterizing early insulation degradation, providing a crucial input for subsequent accurate physical inversion.

[0067] S2.4: Based on the instantaneous current slope vector and the known physical constants of the tungsten wire material, perform the inverse operation of Ohm's law and temperature coefficient compensation processing to eliminate the influence of ambient temperature drift and inversely retrieve the dynamic evolution trajectory of the tungsten wire resistance over time, generating a cold resistance dynamic evolution curve.

[0068] Using the instantaneous current slope vector generated in the previous step and the preset physical constants of the tungsten filament material (including its reference resistance value, temperature coefficient of resistance, etc.) as input, physical inversion and compensation calculations are performed. First, the instantaneous rate of change of current represented by the instantaneous current slope vector, combined with the known circuit supply voltage, is used to perform a physical inversion and compensation calculation. In the formula, The instantaneous rate of change of the dynamic resistance of the tungsten wire is given. The instantaneous rate of change of the supply voltage to the circuit. A constant supply voltage applied to the circuit, This represents the instantaneous current flowing through the tungsten filament at the time of calculation. denoted as , where is the instantaneous rate of change of the instantaneous current in the tungsten

[0069] The instantaneous rate of change of the dynamic resistance of the tungsten filament is calculated. Then, temperature coefficient compensation is introduced. Using the temperature coefficient model of tungsten filament resistance, the theoretical resistance change component caused by the ambient temperature drift monitored by the sensor is subtracted from the calculated resistance change rate, thereby separating the resistance change caused purely by the insulation degradation of the device itself.

[0070] The tungsten wire resistance temperature coefficient model is expressed by the following formula: In the formula, For tungsten wire at a temperature of The resistance value at that time, For tungsten filament at reference temperature The reference resistance value below, These are known physical constants or device factory calibration values. This represents the current absolute temperature of the tungsten filament. The reference temperature is usually standard room temperature. is the temperature coefficient of resistance of tungsten wire.

[0071] By integrating this compensated resistance change rate sequence over time and superimposing it with the initial resistance value based on physical constants, the dynamic evolution curve of the cold-state resistance, reflecting the evolution of the intrinsic state of the tungsten filament after eliminating environmental thermal interference, can be derived. This curve clearly depicts the true trajectory of the resistance of the core conductive component of the power device changing over time during the initial cold start phase, providing direct physical time-series data for the final quantification of performance degradation.

[0072] S2.5: Perform linear regression fitting on the dynamic evolution curve of the cold resistance to extract the rate index of resistance change over time in the initial stage of power-on, and generate the initial resistance change rate characteristic quantity characterizing the degree of insulation degradation of the power device.

[0073] like Figure 2 As shown, step S3 involves using the initial resistance change rate characteristic as an input variable, combined with the physical law of transient current disturbance caused by microampere-level leakage current, to construct a cold-state resistance mutation response model to map the degree of insulation degradation of power devices, thereby generating a primary degradation criterion reflecting the risk of micro-breakdown of the gate oxide layer. Specifically, this includes: S3.1: Perform time-domain differential operation on the initial resistance change rate characteristic to obtain the instantaneous slope fluctuation sequence characterizing the dynamic response of the resistance temperature coefficient in the low-temperature region of the tungsten wire. This instantaneous slope fluctuation sequence will be used as the direct input for subsequent noise suppression processing.

[0074] When performing time-domain differential operations on the initial resistance change rate characteristic generated by step S2.5, this characteristic is loaded as an input object into the high-precision time series analysis module.

[0075] A unified time reference system is established with the initial stage of power-on at time zero as the benchmark. The instantaneous change is obtained by calculating the difference between adjacent samples for each sampling point in the feature quantity sequence.

[0076] By using normalization with the sampling interval Δt as the denominator, the difference value is converted into the rate of change per unit time, thus obtaining the instantaneous slope of the resistance change rate.

[0077] A sliding window operation mechanism is introduced to average and remove trends of multiple consecutive instantaneous slope values ​​according to the window length L, so as to eliminate low-frequency drift and retain the dynamic response of the resistance temperature coefficient in the low-temperature region of tungsten filament.

[0078] The instantaneous slope fluctuation value is calculated using the following formula: in, The slope represents the instantaneous rate of change of resistance. It is a function of resistance value. The current time sampling point, This represents the sampling time interval.

[0079] The calculated instantaneous slope sequence is stored as a fluctuation sequence dataset, and a time index is added for direct input to subsequent noise suppression algorithms.

[0080] By using the time-domain differentiation and windowing processing methods described above, the initial resistance change rate characteristic quantity of the previous step is transformed into an instantaneous slope fluctuation sequence that can accurately characterize the dynamic response of the resistance temperature coefficient in the low-temperature region of the tungsten filament, thereby amplifying and capturing the physical change trajectory under microampere-level leakage current interference.

[0081] For example, in the testing of an ultrasonic power supply PCBA with a rated power of 150W, the sampling rate was set to 1MHz, the time window length was set to 20ms, and the sampling interval Δt was set to... The time interval is seconds, and the sliding window length L is 100 sampling points. The resistance value of a certain segment on the initial resistance change rate characteristic curve is determined by... Ω increased to Ω, corresponding to the instantaneous slope calculated as ≈ Ω / s. This fluctuation value remains approximately higher than the baseline even after window averaging. The signal exhibits a significant instantaneous slope fluctuation, measured in Ω / s. Verification has shown that this slope fluctuation sequence can preserve the transient current disturbance characteristics caused by insulation degradation in subsequent wavelet threshold denoising, significantly improving the sensitivity of the degradation risk criterion.

[0082] S3.2: Based on the instantaneous slope fluctuation sequence, perform adaptive wavelet threshold denoising algorithm processing to remove high-frequency electromagnetic interference and extract pure transient current disturbance baseline signal. This pure transient current disturbance baseline signal will be used as the reference data for constructing physical mapping relationship.

[0083] S3.3: Using the pure transient current disturbance baseline signal combined with the physical law of transient current disturbance caused by microampere leakage current, a nonlinear fitting operation is performed to construct a cold-state resistance sudden change response model that can quantify the relationship between the degree of insulation degradation and the magnitude of resistance jump. This cold-state resistance sudden change response model will serve as the calculation engine for generating state criteria.

[0084] Under the input conditions, the target is the clean transient current disturbance baseline signal obtained by S3.2 adaptive wavelet threshold denoising. Combining this with the known physical response law of microampere-level leakage current to the cold-state resistance jump of the incandescent lamp tungsten filament, a nonlinear fitting modeling operation is performed. The baseline signal is segmented according to the time series, and the peak amplitude and corresponding time delay of each segment are extracted to form a physical response feature reference table. Using the amplitude information in the feature reference table as the ordinate and the time delay as the abscissa, a candidate set of combined functions containing polynomial and exponential terms is constructed to fit the correspondence between the degree of insulation degradation and the resistance jump amplitude. Based on the candidate set, the function form with the highest fitting degree is selected, and its coefficient parameters are optimized using the least squares method to establish the final cold-state resistance jump response model. The core calculation formula of this model is: in, This represents the amplitude of the cold-state resistance jump. This represents the amplitude of the transient current disturbance. To correspond to the delay time, , , These are the material and structural response coefficients obtained through fitting optimization. The order is a polynomial. By optimizing the coefficients, a fitting curve is formed that can cover the cold-state response characteristics under different leakage current levels. This curve is then solidified into a calculation engine for generating state criteria, realizing a mapping transformation from baseline signal to quantification of degradation degree.

[0085] The observed phenomenon originates from the transient electro-thermal-resistance coupling process induced by microampere-level leakage current at microscopic defects in a cold-state tungsten filament. Specifically, when leakage current flows through a localized region containing lattice distortion, impurities, or early cracks, localized Joule heating occurs. In the cold state, the tungsten filament has a high temperature coefficient of resistance; this localized temperature rise causes a sharp increase in resistance, leading to increased power dissipation and further temperature rise at that point, forming a rapid positive feedback loop, macroscopically manifested as a momentary "jump" in resistance. Subsequently, the heat diffuses to the surrounding low-temperature bulk, and the resistance at that point decays and recovers. Therefore, this physical law is mathematically characterized as: the magnitude of resistance change... With disturbance current The relationship is power function (reflecting nonlinear heating intensity), while the recovery process shows an exponential decay trend with time (t) (reflecting thermal diffusion).

[0086] Through model building and optimization, the pure transient current disturbance baseline from the previous step is transformed into a mathematical model characterizing the relationship between insulation degradation and resistance jump amplitude, thus realizing a core physical mapping platform for subsequent gate oxide micro-breakdown risk calculation.

[0087] For example, in the scenario of ultrasonic power device diagnostics, the peak value of the collected pure transient current disturbance baseline is Ampere, delay time is milliseconds, model order set to Fitting optimization yielded , , Substitute the above parameters into the formula: Calculated The output shows that the insulation degradation quantification value is in the high range. The criterion module can identify the warning state of an impending short circuit based on this and drive the downstream emergency disconnection logic, verifying that the model has significantly improved the accuracy of judgment and rapid response capability under high leakage risk.

[0088] S3.4: Input the initial resistance change rate characteristic into the constructed cold resistance mutation response model for mapping transformation processing to calculate the insulation degradation quantification index characterizing the probability of micro-breakdown risk of the gate oxide layer of the power device. This insulation degradation quantification index will be used as an intermediate variable for the generation of the final criterion.

[0089] When loading the initial resistance change rate feature into the cold-state resistance mutation response model, feature channels corresponding one-to-one with the model input nodes are constructed in vector form to ensure strict matching between the feature values ​​and the model fitting parameters in the numerical domain. A normalization transformation is performed on the input vector to ensure its distribution range is within the standardized range of model training. Scale differences between different test batches are eliminated by subtracting the mean and dividing by the standard deviation. Combining the nonlinear mapping kernel function defined in the cold-state resistance mutation response model, a mapping operation is performed on the normalized initial resistance change rate. Values ​​are passed in each hidden layer through a weight matrix product, and bias terms are accumulated to form an intermediate variable matrix for risk probability calculation. A probability density function is fitted to the elements of this matrix, selecting the function form that satisfies the minimum error criterion, and the insulation degradation quantification index of the gate oxide layer micro-breakdown risk of power devices is calculated. The time-series resistance trajectory obtained by the inverse operation of Ohm's law and temperature coefficient compensation is used as the physical constraint of the model, and the risk probability is solved using the following formula: in, This represents the initial change in resistance. This is the initial resistance value in the cold state. The temperature coefficient constant of resistance in the low-temperature region. The leakage current response coefficient is... This represents the peak value of the transient current disturbance. Through the above processing method, the result of the previous step is transformed into a quantitative indicator of insulation degradation characterizing the probability of micro-breakdown of the gate oxide layer, thus providing the quantitative physical input required for subsequent criterion generation.

[0090] For example, in the ultrasonic power supply PCBA diagnostic platform, the initial resistance change rate characteristic is set to 0.008 Ω / ms, the cold-state initial resistance is 15.2 Ω, and the resistance temperature coefficient constant α in the low-temperature region is set to... / ℃, leakage current response coefficient β is 0.12 A. - ¹, The peak value of the transient current disturbance was 0.035 A. Substituting the parameters into the risk probability formula, we get: Calculations show that the quantitative indicator of risk probability is approximately 2.36 × 10⁻⁶. -5 The model output indicates a Level-1 degradation state, suggesting a slight decrease in insulation performance. This result, demonstrated in continuous batch testing, shows stable risk indicators with minimal fluctuations, validating the significant effectiveness of mapping transformation and parameter normalization in eliminating batch-to-batch differences and improving the accuracy of micro-breakdown risk assessment.

[0091] S3.5: Perform multi-level threshold comparison logic judgment operation based on the insulation degradation quantification index to generate a primary degradation criterion that clearly identifies the micro-breakdown risk level of the gate oxide layer. This primary degradation criterion will be output to the downstream decision module as a core component of the three-dimensional joint state feature set.

[0092] Based on the insulation degradation quantification index data vector, a multi-level risk threshold setting operation is performed, using a set of boundary values ​​derived from offline sample statistical analysis as a comparison benchmark. The insulation degradation quantification index is input into the multi-level threshold comparison unit, where conditional matching operations are performed one by one to form a judgment matrix corresponding to the risk level. During the judgment process at each level, relying on a preset probability interval, a logical judgment operator is used to perform differential judgment between the measured value and the threshold set, generating a temporary judgment identifier. This temporary judgment identifier is mapped to a status label representing the gate oxide layer micro-breakdown risk level through encoding conversion, achieving a unification of qualitative discrimination and quantitative evaluation. During label generation, consistency verification is performed on the status labels to ensure that the output results conform to the physical laws of the cold resistance mutation response model. Risk level labels that pass verification are written into the primary degradation criterion register structure. Through multi-level threshold comparison logical judgment processing, the insulation degradation quantification index from the previous step is transformed into a clearly defined primary degradation criterion, achieving rapid hierarchical identification of gate oxide layer micro-breakdown risk.

[0093] For example, in the ultrasonic power supply PCBA testing process, the insulation degradation quantification index is calculated and set to a value in step S3.4. The multi-level risk threshold group is set as follows: Level-1 threshold range. Level-2 threshold range Level-3 threshold range The rest are in the Level-0 normal state. During the comparison process, conditional operators are used. Logical matching is performed on the insulation degradation quantification index, and the returned true value is mapped to a Level-1 label. This label undergoes consistency verification and is consistent with the risk level predicted by the cold resistance mutation response model. If it is deemed qualified, it is written into the primary degradation criterion register. The above processing was verified in test samples. When the insulation degradation quantification index falls into different ranges, the corresponding level label can be stably output. The application effect shows a significant improvement in the judgment of micro-breakdown risk level, which can meet the needs of rapid classification and early warning in production lines.

[0094] like Figure 3 As shown, step S4 involves performing a spatial non-uniformity index calculation operation on the pixel-level brightness evolution data in the grayscale sequence during the 20-100 ms steady-state pre-state stage to generate a brightness diffusion rate feature vector characterizing the degree of local Joule heat accumulation and the difference between single-point micro-short circuits and planar insulation aging. Specifically, this includes: S4.1: Perform region of interest extraction processing on the continuous frame images within a time window of 20 to 100 milliseconds in the grayscale sequence to obtain a dynamic filament grayscale matrix sequence containing the complete structure of the incandescent lamp filament, which serves as the basic input object for subsequent spatial feature analysis.

[0095] Based on the grayscale sequence, all consecutive frame images of the evolution stage after the incandescent lamp is lit and before the brightness reaches a stable state (specifically, within a time window of 20 to 100 milliseconds after power-on) are extracted. Then, for each frame image, according to a preset filament shape template or brightness threshold, the pixel region containing the complete filament structure is automatically identified and selected, and the grayscale values ​​of all pixels in the region are extracted and constructed into a two-dimensional matrix. This operation is repeated for each frame image within the time window, and finally a dynamic filament grayscale matrix sequence composed of multiple two-dimensional matrices is generated in chronological order. This sequence completely records the spatial distribution change of the filament's surface brightness over time in the early stage of hot start-up, which serves as the basic input object for subsequent spatial feature analysis.

[0096] S4.2: Based on the dynamic filament grayscale matrix sequence, the centroid tracking algorithm is used to perform coordinate calculation on the bright areas in each frame of the image to generate a set of filament hot spot displacement vectors that characterize the trajectory of the luminous center position as it migrates over time.

[0097] Using the dynamic filament grayscale matrix sequence generated in step S4.1 as input, centroid coordinate calculation is performed on each frame image (i.e., each two-dimensional grayscale matrix) in the sequence. Specifically, the operation is as follows: First, a global or adaptive threshold is applied to the single-frame image for binarization to separate high-brightness (e.g., filament illumination) regions. Then, the grayscale value of each pixel within this region is used as a weight to calculate its weighted average horizontal and vertical coordinates. This coordinate point is the centroid coordinate of the filament illumination region in the current frame image. By repeatedly performing the above centroid calculation on each frame image in the sequence in chronological order, a series of two-dimensional coordinate points arranged by timestamps are obtained. ,in This is the frame number. Finally, the centroid coordinate difference between adjacent frames (i.e., Arranged in sequence, they form a two-dimensional vector sequence, namely the filament hot spot displacement vector set. This vector set precisely describes, in a quantitative form, the migration trajectory and instantaneous direction of the luminous center of the filament in the image plane over time during the initial stage of filament heating, providing crucial spatiotemporal location information for subsequent analysis of the dynamic process of heat diffusion.

[0098] S4.3: Based on the filament hot spot displacement vector set, perform radial gray-scale gradient distribution statistical operation to calculate the instantaneous brightness diffusion rate scalar that reflects the speed of heat diffusion from the center to the edge.

[0099] For each time sampling point in the filament hot spot displacement vector set, a polar coordinate system with the light emission center as the origin is constructed, and the RGB grayscale matrix is ​​converted into a radial grayscale distribution vector in order to analyze the attenuation of brightness along different radial directions.

[0100] The central difference operator is executed in the radial gray vector to calculate the change in gray value of adjacent radii, and the squares of the changes in all radial directions are accumulated to form a set of radial gradient intensities.

[0101] Perform a radial integral operation on the radial gradient intensity set to obtain the average gradient value reflecting the global heat diffusion rate along the radius, and represent the dynamic change of this index in the form of a time series.

[0102] A sliding window is used to calculate the time mean and standard deviation of the average radial gradient value, and local peaks caused by random noise are removed to generate a stable instantaneous brightness diffusion rate scalar.

[0103] By using this instantaneous brightness diffusion rate scalar as input for subsequent spatial non-uniformity index calculations, a quantitative description of the local Joule thermal accumulation process can be achieved.

[0104] By statistically processing the radial gray-scale gradient distribution, the filament hot spot displacement vector set is transformed into an instantaneous brightness diffusion rate scalar that characterizes the heat diffusion rate, thereby realizing the quantitative determination of the heat diffusion characteristics before steady state.

[0105] For example, in an ultrasonic power supply connection test, the filament hot spot displacement vector set acquired by the high frame rate image sensor has a light emission center coordinate of (128, 128) pixels, a radial sampling step size of 2 pixels, and a maximum radius of 64 pixels. After polar coordinate transformation, the grayscale matrix has a grayscale value of 220 at a radius of 10 pixels and a grayscale value of 215 at a radius of 12 pixels. The radial difference is -5, which is squared to obtain 25 and accumulated into the radial gradient intensity set. The total intensity value of 5000 is obtained by summing the squared differences over all radial directions. The average gradient value is obtained by integrating over the radial direction and dividing by the number of sampling directions. The result was 13.89. In the time series, the sliding window length was set to 5 frames. After calculating the average value and standard deviation and removing abnormal peaks, the final instantaneous brightness diffusion rate scalar was 13.5. This value was used in the subsequent calculation of the spatial non-uniformity index. The verification experiment showed that the scalar significantly improved the performance of distinguishing between single-point micro short circuits and planar insulation aging, and maintained stable output under different ambient temperature conditions.

[0106] S4.4: Using the instantaneous brightness diffusion rate scalar, combined with the variance analysis algorithm, the spatial dispersion of pixel gray values ​​in the filament region is quantitatively evaluated to generate a spatial non-uniformity index characterizing the non-uniformity of local Joule heat accumulation.

[0107] Using the instantaneous brightness diffusion rate scalar and the dynamic filament grayscale matrix sequence generated in the previous steps as input, a spatial uniformity quantification assessment based on analysis of variance is performed. Specifically, for a specific image frame in the dynamic sequence corresponding to the current instantaneous rate, a pixel grayscale matrix containing the complete filament structure is extracted. Subsequently, a one-way analysis of variance is performed on the grayscale values ​​of all effective pixels in the matrix to calculate its overall variance, thereby quantifying the dispersion of brightness among pixels. To eliminate the influence of absolute brightness levels, this variance value is divided by the mean pixel grayscale value of the image frame to obtain a normalized dispersion coefficient, which is the spatial non-uniformity index. The larger this index, the more non-uniform the spatial distribution of brightness (i.e., temperature) on the filament surface, directly characterizing the strength of local Joule heat accumulation caused by material aging or microscopic defects. This provides a key quantitative basis for distinguishing between concentrated overheating caused by single-point micro-short circuits and diffuse thermal non-uniformity caused by planar insulation aging.

[0108] S4.5: Based on the spatial non-uniformity index and the filament hot spot displacement vector set, perform a multi-dimensional feature fusion mapping operation to output a brightness diffusion rate feature vector that ultimately characterizes the difference between single-point micro-short circuits and planar insulation aging.

[0109] Based on the spatial non-uniformity index and the filament hot spot displacement vector set, numerical scaling is performed on the two types of physical features. The dimensions of the spatial non-uniformity index are standardized and mapped to the coordinate components of the hot spot displacement vector into a unified feature space.

[0110] In a unified feature space, a two-dimensional feature matrix is ​​constructed, with the spatial non-uniformity index as the first column vector and the instantaneous amplitude sequence of the hot spot displacement vector as the second column vector. The correlation coefficient between the two is calculated using the covariance matrix to ensure that the calculation of the brightness diffusion rate can simultaneously reflect the coupling effect of Joule thermal accumulation and luminous center migration.

[0111] Based on the estimation results of the correlation coefficient, a weighted fusion algorithm is used to combine the two types of features, and the spatial non-uniformity index is assigned as a weight. Assign weights to the hot spot displacement vector And perform a linear combination to generate candidate vectors for brightness diffusion velocity.

[0112] Normalize the candidate vectors of brightness diffusion velocity and map them to the interval Furthermore, a threshold segmentation function is used to eliminate abnormally high amplitude interference, ensuring the numerical stability of the feature vector under different working conditions.

[0113] Through the above-mentioned multi-dimensional feature fusion mapping processing method, the spatial non-uniformity index and filament hot spot displacement vector set results of the previous step are transformed into brightness diffusion velocity feature vectors that can be directly used to distinguish between single-point micro short circuits and planar insulation aging, thereby achieving the expected technical effect of spatial thermal distribution diagnosis in the pre-steady-state stage.

[0114] For example, in the pre-steady-state stage after a certain ultrasonic power supply PCBA is connected to 220V, its grayscale matrix sequence, after processing, yields a spatial non-uniformity index of: The mean of the instantaneous amplitude sequence of the hot spot displacement vector is The two are then constructed into a two-dimensional feature matrix. The correlation coefficient was obtained from the analysis of covariance. According to weighting coefficient and Perform a linear combination, the calculation formula is as follows: The result is Normalize the results to an interval. After applying threshold segmentation to filter outout values, the final brightness diffusion rate characteristic value is obtained. In the validation of multiple batches of samples, this feature value is concentrated in single-point micro-short-circuit samples. The range, while in the planar insulation aging samples, it is concentrated in The interval indicates that the fusion feature can significantly improve the ability to distinguish between the two types of failure modes.

[0115] Step S5: Based on the near-infrared spectral response data, the spectral peak shift in the 700-nanometer to 950-nanometer band is calibrated to retrieve the equivalent power consumption increment of the inversion loop and generate a correction factor for the average temperature rise rate of the tungsten filament, excluding interference from environmental temperature drift and voltage fluctuations. Specifically, this includes: S5.1: Perform spectrum truncation processing on the raw spectral response data of the incandescent lamp filament region captured by the high frame rate image sensor in the 700-950 nanometer band to obtain narrowband near-infrared spectral sequence data containing the thermal radiation characteristics of tungsten filament.

[0116] For the raw spectral response data of the incandescent lamp filament region captured by the high frame rate image sensor, the input data source is determined to be a multi-dimensional spectral matrix containing multi-band information of visible light and near-infrared light, and it relies on the time stamp synchronization and spatial region extraction results completed in the previous steps.

[0117] Before performing the spectrum interception process, the narrowband filter parameter configuration module is called to set the physical filter bandwidth to a center wavelength of 825nm, covering the range of 700nm to 950nm, to ensure that the interception range completely covers the near-infrared characteristic segment of tungsten filament thermal radiation.

[0118] Wavelength axis resampling is performed on the original spectral response data matrix. An interpolation algorithm is used to convert non-uniform sampling points into an equally spaced wavelength array for accurate calibration of peak position shift in the future.

[0119] By using a frequency domain window function to perform bandpass filtering on the wavelength array, spectral components below 700nm and above 950nm are filtered out, retaining only the near-infrared energy distribution within the target band.

[0120] Within the retained band data, energy normalization is performed to standardize the near-infrared spectral intensity values ​​of all pixels to a uniform dimension, facilitating comparison of data from different time frames.

[0121] The encapsulated, filtered, and normalized near-infrared spectral data sequence serves as a narrowband near-infrared spectral sequence, providing a stable and reliable input for subsequent peak detection and temperature rise rate derivation.

[0122] By using the above-mentioned spectral interception and processing methods, the original spectral response data from the previous step is transformed into a narrow-band near-infrared spectral sequence containing only the thermal radiation characteristics of tungsten filaments, thereby achieving the effects of ambient light interference isolation and purification of temperature-sensitive bands.

[0123] For example, in an online inspection system for an ultrasonic power supply PCBA, the pixel array size of the high frame rate image sensor is 2048×2048, the sampling frame rate is 1200fps, and the raw spectral response data includes the 400nm to 1000nm band. The light intensity value of each wavelength channel is recorded with 12-bit precision. During this step, the physical bandwidth of the narrowband filter is set to ±125nm, the center wavelength is 825nm, and the wavelength axis resampling step size is 1nm. After bandpass filtering, the retained near-infrared spectral sequence length is 251 wavelength sampling points. Energy normalization scaling scales the spectral intensity value of each frame to the range of 0 to 1. For example, the original intensity value of a frame at a wavelength of 830nm is 3120, and the normalized intensity value is... The value is approximately 0.76. After all processing is completed, the output narrowband near-infrared spectral sequence is used for peak detection in S5.2, which can significantly improve the stability of peak position shift calculation. Even in the presence of ambient light intensity fluctuations and voltage transient disturbances, the peak position calibration error is kept within ±0.5nm, ensuring a significant improvement in the calculation accuracy of the temperature rise rate correction factor.

[0124] S5.2: Perform peak detection algorithm processing based on the narrowband near-infrared spectral sequence data to extract the near-infrared spectral peak position shift characteristic parameter characterizing the instantaneous temperature change of the tungsten filament.

[0125] Using the narrowband near-infrared spectral sequence data generated in step S5.1 as input, feature parameter extraction based on a peak detection algorithm is performed. Specifically: First, for the spectral data of each moment (or each frame) in the sequence, within the band range of 700 nm to 950 nm, the main spectral peak characterizing the thermal radiation intensity of the tungsten filament is accurately identified through local maximum search or Gaussian fitting, and its corresponding center wavelength position (unit: nanometers) is recorded, thus obtaining a spectral peak position sequence arranged in chronological order. Next, the difference in peak center wavelength between adjacent moments (or relative to a stable reference moment) in the sequence is calculated to obtain a series of instantaneous peak position shifts. Finally, by statistically analyzing these instantaneous shifts (such as calculating the mean, root mean square, or trend slope within a calculation window), a stable and quantifiable near-infrared spectral peak position shift feature parameter is generated. The near-infrared spectral peak shift characteristic parameter directly and sensitively reflects the instantaneous temperature change of the tungsten filament caused by Joule heating due to electric current. This is because, according to Wien's displacement law of blackbody radiation, the wavelength of the radiation peak is inversely proportional to the temperature of the radiating body. Therefore, any systematic shift in the peak position is a direct optical measure of the temperature change of the tungsten filament.

[0126] S5.3: Using the near-infrared spectral peak shift characteristic parameters combined with Planck's blackbody radiation law mapping relationship, nonlinear fitting calculations are performed to generate an estimated value of the loop equivalent power consumption increment that reflects the real-time energy dissipation level of the loop.

[0127] Based on the near-infrared spectral peak shift characteristic parameters obtained through S5.2 processing, a pre-defined physical mapping algorithm module is invoked, using these characteristic parameters as input variables to establish a functional relationship between the instantaneous surface temperature of the tungsten filament and the change in radiation peak position. Applying Planck's blackbody radiation law, the peak shift is mapped to the corresponding radiation intensity distribution curve, and wavelength range constraints are set during curve fitting to eliminate interference from unrelated frequency bands. A nonlinear regression model conforming to the characteristics of the near-infrared band is constructed, and the least squares fitting method is used to minimize the error between the observed peak position data and the theoretical blackbody radiation curve, generating an approximate sequence of instantaneous tungsten filament temperature values. Using the instantaneous temperature and time series of the tungsten filament, the instantaneous energy dissipation value of the loop under the reference condition is calculated using the heat-work conversion formula, and the energy consumption change is iteratively solved based on the temperature increment per unit time. Combining the temperature coefficient of resistance and instantaneous current data, the temperature change is converted into an estimated value of the loop's equivalent power consumption increment, forming a heat-work index that can be used for downstream differential compensation.

[0128] in, The radiation power density per unit wavelength, is Planck's constant. The speed of light in a vacuum. The wavelengths corresponding to the spectral peaks are given. Boltzmann's constant, The absolute temperature of the tungsten filament is obtained through fitting. The peak shift is converted into an estimated value of the loop's equivalent power consumption increment through nonlinear fitting calculations, thus achieving accurate quantification of heat dissipation.

[0129] By using nonlinear fitting and heat-work conversion formula processing, the peak position shift characteristic parameter of the previous step is transformed into an estimated value of the loop equivalent power consumption increment that can characterize the real-time energy dissipation level of the loop, thus realizing the generation of high-precision heat-work index based on spectral signals.

[0130] For example, in a scenario where an ultrasonic power supply PCBA is connected in series with an incandescent lamp, the near-infrared peak shift is detected as a shift of +4nm at 780nm. When performing nonlinear fitting, the fitting wavelength range is set to 780–950nm, and Planck's constant is... speed of light Boltzmann constant The absolute temperature of the tungsten wire obtained by fitting is... The temperature rise is 20K. Substituting this into the heat-work conversion formula, the instantaneous radiant power change is 2.15W. Comparing this to the baseline power condition, the equivalent power consumption increment of the circuit is calculated to be 1.8W. This value can eliminate the influence of environmental temperature drift in subsequent differential compensation and is ultimately used to calculate the tungsten filament average temperature rise rate correction factor, improving the thermal characteristic resolution for identifying the degradation state of power devices.

[0131] S5.4: Perform differential compensation calculation based on the estimated value of the equivalent power consumption increment of the circuit and the environmental reference temperature data collected by the environmental temperature sensor to eliminate the influence of environmental temperature fluctuations and generate a net power temperature rise driving amount to eliminate environmental interference.

[0132] Based on the estimated equivalent power consumption increment of the loop and the ambient reference temperature data collected by the ambient temperature sensor, differential compensation calculation is performed to eliminate the influence of ambient temperature fluctuations. The estimated equivalent power consumption increment of the loop is input into the differential calculation unit and matched one-to-one with the ambient reference temperature data according to timestamps. An initial value of the thermal driving quantity is constructed using a mapping relationship. The power consumption data and temperature reference are differentially processed to remove redundant heat components introduced by external temperature changes. The initial value is filtered by a time constant to smooth abrupt changes and improve the stability of the thermal data. A median correction is performed on the smoothed driving quantity using a thermal conduction compensation coefficient to match its thermal inertia characteristics with those of a tungsten filament. Through the above processing method, the equivalent power consumption increment of the loop is transformed into a net power temperature rise driving quantity free from environmental interference, achieving the elimination of ambient temperature drift and providing a high-precision thermal reference for subsequent voltage interference correction stages.

[0133] For example, in the ultrasonic power supply PCBA testing scenario, the estimated equivalent power consumption increment of the loop is 2.35 W, and the ambient temperature reference data collected by the ambient temperature sensor is 27.4 ℃. The differential compensation calculation formula is as follows: in, This is an estimated value for the incremental equivalent power consumption of the loop. This is the ambient temperature compensation coefficient. This is the current ambient temperature data. This is the environmental reference temperature. In this example, it is taken as... It is 0.12 W / ℃. The temperature is 25.0 ℃. The net power temperature rise drive after removing environmental interference is calculated as follows: After calculation, the net power temperature rise drive was found to be 2.05 W. This drive, after being processed by a first-order low-pass filter with a time constant τ = 18 ms, had its output fluctuation range converge to ±0.02 W, significantly improving the stability of the thermal properties and successfully eliminating the power offset caused by ambient temperature fluctuations. This provides stable input data for the subsequent calculation of the tungsten filament average temperature rise rate correction factor.

[0134] S5.5: Based on the net power temperature rise driving amount to eliminate environmental interference and the real-time voltage fluctuation coefficient provided by the power grid voltage monitoring module, normalization correction processing is performed to finally generate the tungsten filament average temperature rise rate correction factor that eliminates voltage fluctuation interference.

[0135] Step S6: The primary degradation criterion, brightness diffusion rate feature vector, and tungsten filament average temperature rise rate correction factor are fused into a three-dimensional joint state feature set, and input into a lightweight decision tree model trained offline using classic failure samples, to output early degradation diagnosis results for power devices containing four levels of degradation state labels. Specifically, this includes: S6.1: Perform spatiotemporal alignment and dimensional normalization on the primary degradation criterion, brightness diffusion rate feature vector, and tungsten filament average temperature rise rate correction factor to eliminate numerical scale differences between different physical quantities and generate a standardized multidimensional feature component sequence with a unified metric standard.

[0136] S6.2: Perform tensor splicing and weighted fusion operations based on the standardized multidimensional feature component sequence to construct a three-dimensional joint state feature set that can fully characterize the insulation degradation, local heat accumulation and equivalent power consumption increment of power devices.

[0137] The input standardized multidimensional feature component sequence is fed into the tensor construction unit via the feature processing module. Within this unit, feature components from different physical quantities are first grouped and encoded. Insulation degradation-related components, thermal accumulation-related components, and power consumption increment-related components are mapped to basic blocks along different axes of the tensor to clarify the spatial arrangement logic of the data. For each basic block, a tensor dimension expansion operation is performed, introducing empty dimensions and zero-padding to ensure consistent axis lengths, providing dimension matching conditions for subsequent stitching. Tensor stitching calculations are then performed on the matched basic blocks, placing the insulation degradation block on the first axis, the thermal accumulation block on the second axis, and the power consumption increment block on the third axis. During stitching, the temporal index of the original component sequence remains unchanged to maintain the spatiotemporal continuity of the features. The stitched three-dimensional tensor enters the weighted fusion module. This module performs linear weighted calculations based on the statistical weights of each feature component's contribution to the failure mode classification. The weighting coefficients are obtained through historical sample analysis and stored in the weight register. The fusion calculation formula is as follows: in, For the fused 3D joint state feature set tensor, For the first Class features in the first Weighting coefficients for time steps For the first Class features in the first Standardized feature components of time step, This represents the index of three feature dimensions. This represents the time series index. After the fusion calculation is completed, a normalization filter operation is performed on the tensor to remove extreme value anomalies introduced by the weighting bias, and a three-dimensional joint state feature set with good continuity is generated. Through the above processing method, the result of the previous step is transformed into a three-dimensional joint state feature set that can fully characterize the insulation degradation, local heat accumulation and equivalent power consumption increment of power devices, thus achieving the high consistency input required by the downstream decision tree inference engine.

[0138] For example, in a degradation detection application of an ultrasonic power supply PCBA, the primary degradation criterion component matrix has a size of 10×20 (feature count × time index), the brightness diffusion velocity feature vector matrix has a size of 8×20, and the tungsten filament average temperature rise rate correction factor matrix has a size of 5×20. After spatiotemporal alignment and dimensional normalization, these components are mapped to the first, second, and third axes of a three-dimensional tensor, respectively, and then filled into a 10×20×10 structure of the same size. After tensor splicing, weighted fusion is performed using weight coefficients of 0.5, 0.3, and 0.2, corresponding to the formula in... The values ​​are 0.5, 0.3, and 0.2 respectively. The values ​​are taken as the normalized eigenvalues ​​of each matrix, and a fusion operation is performed to obtain a tensor. Each element is a weighted sum of the corresponding elements of the three feature matrices. During the normalization filtering process, the threshold range is set to [-1.0, 1.0], outliers outside the range are removed and replaced with the axial mean. The final output three-dimensional joint state feature set shows a significant improvement in the correlation between features in the validation set. The classification model's accuracy in identifying degradation states under this input condition is significantly improved compared to the baseline input, and it can accurately distinguish between three early failure modes: slight degradation of insulation performance, gate structure damage, and proximity short circuit.

[0139] S6.3: Utilize the lightweight decision tree model structure parameters pre-stored in the embedded memory to perform root node initialization loading processing on the three-dimensional joint state feature set, so as to generate an active classification inference engine containing current decision boundary information and node splitting threshold configuration.

[0140] The lightweight decision tree model structure parameters in the embedded memory are called for the three-dimensional joint state feature set. The static parameter file containing the model root node configuration and the full tree split threshold is read as the original data object for classification inference initialization.

[0141] The parameter file is parsed into node threshold vectors, a list of split feature indices, and a leaf node category mapping table, and a corresponding parameter cache area is established in the MCU memory to establish the basic data structure required for the inference engine to run.

[0142] The feature index mapping process is performed on the three-dimensional joint state feature set, matching its dimension labels with the split feature index list one by one, and generating a feature value index matrix with node reference paths to ensure that the features are called in the preset order during subsequent classification traversal.

[0143] By using a memory pointer loading method, the threshold parameters and split feature index of the root node are written into the runtime structure of the inference engine, and an active inference node object containing the current decision boundary information and the node split threshold configuration is constructed.

[0144] The validity of the root node threshold is verified by threshold loading and validation operations. The determination formula is as follows: The difference between the storage threshold and the standard threshold, To ensure that the model initialization accuracy meets diagnostic requirements, an upper limit for allowable error is set.

[0145] Through the above loading and verification processes, the standardized three-dimensional joint state feature set is transformed into a feature input matrix that can be directly called by the active classification inference engine, thereby achieving the effect of early degenerate state classification initialization based on preset decision boundaries.

[0146] For example, in an embedded MCU system used for production line screening, the stored lightweight decision tree model contains 16 internal nodes and 4 leaf nodes. The node threshold range is set from 0.05 to 1.20, and the corresponding input features are, in order, the primary degradation criterion, the brightness diffusion rate, and the tungsten filament average temperature rise rate correction factor. After reading the model parameter file in the memory, the split feature index list [0,1,2,0,2,1] and the split threshold vector [0.12,0.35,0.55,0.08,0.90,1.15] are parsed. The three-dimensional joint state feature set [0.14,0.38,0.93] is subjected to index mapping processing to generate the feature value index matrix [[0.14],[0.38],[0.93],[0.14],[0.93],[0.38]], and loaded into the root node and subsequent node structures of the inference engine. The threshold verification formula is applied as follows. The calculated difference is 0.02, which meets the upper limit of error of 0.05, confirming that the root node initialization is valid. After executing this initialization process, the inference engine enters an active state, enabling it to quickly perform layer-by-layer decisions when the input feature set arrives, achieving high-precision classification startup in the early degenerate state.

[0147] S6.4: Input the three-dimensional joint state feature set into the active classification inference engine to perform layer-by-layer feature matching and node path traversal operations, so as to generate a leaf node index sequence pointing to a specific failure mode based on the combination logic of cold resistance abrupt response, spatial non-uniformity index and spectral peak position shift.

[0148] A layer-by-layer node processing chain is established in the active classification inference engine for the loaded 3D joint state feature set.

[0149] The cold resistance abrupt response parameters, spatial non-uniformity index, and spectral peak position shift are calculated using the node splitting threshold to perform the first layer of feature matching, generating initial matching results for selecting sub-paths.

[0150] Based on the first-level matching results, a target threshold comparison operation is performed, and the feature selection mechanism of a specific subtree branch is invoked to send the data stream that meets the splitting conditions into the feature matching module of the corresponding child node.

[0151] Combinatorial logic is performed on the input features within the intermediate layer nodes. A multi-condition cross-operation method is adopted to match the product relationship between the cold resistance mutation response and the spatial non-uniformity index with the threshold condition of the spectral peak position shift, thereby obtaining a joint matching criterion for path traversal.

[0152] By calling the path traversal control unit through the joint matching criteria, the node sequence scan is performed sequentially, and the current node index is recorded when the threshold condition is met, and it is used as a candidate leaf node.

[0153] Perform index optimization sorting on all candidate leaf nodes and generate the final leaf node index sequence according to matching degree and risk level.

[0154] By using the above-mentioned layer-by-layer feature matching and node path traversal processing method, the three-dimensional joint state feature set of the previous step is transformed into a leaf node index sequence pointing to a specific failure mode, thereby realizing the logical location of the early degradation state of the power device.

[0155] For example, in ultrasonic power supply PCBA testing, the three-dimensional joint state feature set includes a normalized cold resistance change rate of 0.0025 Ω / ms, a spatial non-uniformity index of 0.75, and a spectral peak shift of 5.2 nm. The threshold for the first-layer node of the active classification inference engine is set to a cold resistance change rate > 0.0020 Ω / ms. The first-layer matching operation imports this feature set into the right subtree that satisfies the splitting condition. The second-layer node combinational logic condition is set to (spatial non-uniformity index > 0.70) and (spectral peak shift > 4.0 nm), and the matching value is calculated using a product relationship. The result is 0.001875, which meets the threshold condition for spectral peak shift. The path traversal control unit marks this node as a candidate leaf node. In the discrimination of the third and last layer nodes, the matching degree sorting algorithm prioritizes the Index-3 node according to the node risk level. The corresponding failure mode is gate structure damage. The final leaf node index sequence output is [3]. The diagnostic reasoning process is completed within 200ms, which significantly improves the accuracy of early degradation state positioning.

[0156] S6.5: Perform final category mapping and confidence verification processing based on the leaf node index sequence to output a four-level degradation state label that clearly identifies slight degradation of insulation performance, gate structure damage, or near short circuit state as the early degradation diagnosis result of the power device.

[0157] After the leaf node index sequence is determined, a category label mapping table retrieval operation is performed on the decision tree leaf node corresponding to each index to calculate the state category code pointed to by the node. The mapped state category codes are then validated for consistency within a preset four-level degradation label set to ensure that the code strictly corresponds to one of the four categories defined in the label set: "slightly degraded insulation performance," "gate structure damage," "near short circuit," and "normal." For each state category code, a confidence value is calculated based on the sample distribution information of the decision tree training set. The confidence value is calculated using the proportion of the category's sample count to the total number of samples in the leaf node, and is smoothed to avoid numerical abrupt changes caused by small sample sizes. A threshold comparison logic is applied to the calculated confidence value. If the confidence value is lower than a set lower limit, the upper-level node path is backtracked to rematch the category label, thereby improving decision stability. All validated state category codes with satisfactory confidence are converted into four-level degradation state labels and encapsulated into a power device early degradation diagnosis result structure. In this step, the leaf node index sequence from the previous step is transformed into a four-level degradation state label that can be directly used for risk control strategy matching by combining confidence verification and category mapping, thereby achieving categorical clarity and decision reliability in the diagnostic results.

[0158] For example, in the diagnostic process of an ultrasonic power supply PCBA, the three-dimensional joint state feature set is output as a leaf node index sequence {15,27,42} by a lightweight decision tree inference engine. A category mapping table retrieval is performed on index 15, yielding a state category code of 2, corresponding to the label "gate structure damage". Leaf node 15 contains 120 training samples, and the total number of samples is 150. The proportional confidence score is calculated using the formula... The unsmoothed confidence score was 0.8. After smoothing, the confidence score was adjusted to... The result is 0.8026. Mapping index 27 yields category code 1, corresponding to the label "slight decrease in insulation performance," with a sample count of 80 / 100. After the same smoothing process, the confidence score is... The result is 0.7788. Mapping index 42 yields category code 3, corresponding to the label "near short circuit," with a sample count of 50 / 60. The smoothed confidence score is... The result was 0.7969. All three confidence levels were higher than the set lower threshold of 0.75, passing consistency verification. The final output diagnostic result {Label 2, Label 1, Label 3} corresponds to different operational strategies in the risk control module, achieving highly reliable multi-level degradation state determination.

[0159] Step S7: Based on the level values ​​in the four-level degradation state labels, a graded early warning strategy matching operation is performed. If a near-short-circuit state is determined, an emergency disconnection command is triggered. If a state of slight insulation performance degradation or gate structure damage is determined, a maintenance prediction log is generated to form a closed-loop risk control action. Specifically, this includes: S7.1: Perform threshold comparison processing on the level values ​​in the four-level degradation state labels to generate a graded early warning strategy matching result that includes emergency cut-off triggering conditions, maintenance prediction generation conditions, and normal maintenance conditions.

[0160] The input object is the four-level degradation status label output in step S6.5. This label is an integer level value, and the mapping range includes Level-0 to Level-3, corresponding to the states of normal, slight decrease in insulation performance, gate structure damage, and proximity to short circuit, respectively. According to the preset graded early warning strategy matching algorithm, this level value needs to be compared with the graded threshold group to generate strategy matching results for emergency disconnection, maintenance prediction, and normal maintenance.

[0161] The degradation status label is loaded into the internal register of the early warning strategy matching module, and a matching initialization operation with the threshold parameter group is performed to obtain the correspondence matrix between the current level value and all early warning levels.

[0162] The threshold comparison unit is invoked to perform a one-to-one comparison operation between the level value and the emergency cutoff trigger upper limit threshold, generating a Boolean trigger flag bit, which is the direct input variable for subsequent emergency cutoff conditions.

[0163] The threshold comparison unit is invoked to perform interval member judgment on the level value and the threshold range of the maintenance prediction generation interval, and a maintenance prediction condition flag is generated. This flag is associated with the historical monitoring data encapsulation strategy.

[0164] The threshold comparison unit is invoked to perform an equality judgment between the level value and the normal maintenance condition threshold, and a normal operation flag is generated. This flag is used to maintain the current working state and not trigger any control actions.

[0165] The three flag bits are combined in order of policy priority to generate a policy matching result tuple, and stored in the early warning policy register as a unified driving signal for steps S7.2 to S7.4.

[0166] By comparing and contrasting the tiered thresholds, the four-level degradation status labels from the previous step are transformed into tiered early warning strategy matching results that include emergency cut-off triggering conditions, maintenance prediction generation conditions, and normal maintenance conditions, thereby enabling proactive decision-making for risk control actions.

[0167] For example, a power supply PCBA used in ultrasonic welding equipment outputs a level 4 degradation status label value of 3 during production line testing, representing a near-short circuit state. The threshold parameter group inside the early warning strategy matching module is configured as follows: emergency cut-off trigger upper limit threshold. Maintenance prediction interval threshold range Normal maintenance condition threshold The rating value is compared with the emergency cutoff trigger threshold. The output trigger flag is set to TRUE; the level value is compared with the maintenance prediction interval threshold range, and the interval membership operation is performed, resulting in FALSE; the level value is compared with the normal maintenance condition, and the equality judgment is performed, resulting in FALSE. The strategy matching result tuple is (TRUE, FALSE, FALSE), which drives the emergency cut-off action to block the large current in the main circuit and prevent the fault from escalating; if the level value is 2, the emergency cut-off flag is FALSE, and the maintenance prediction flag is TRUE, then the maintenance log generation strategy will be triggered, and the characteristic parameters of the gate structure damage will be recorded for subsequent maintenance optimization reference. This embodiment verifies the response accuracy and action decision rationality of the graded early warning strategy matching module under different level states, and achieves a significantly improved risk prevention and control capability.

[0168] S7.2: Based on the emergency cut-off triggering conditions in the matching results of the graded early warning strategy, a forced shutdown operation is performed on the drive signal of the main circuit of the ultrasonic power supply to generate an emergency cut-off command to block the flow of large current and prevent the precursor of local short circuit from evolving into a catastrophic failure.

[0169] S7.3: Using the maintenance prediction generation conditions in the matching results of the graded early warning strategy, the historical monitoring data corresponding to the state of slight decrease in insulation performance or damage to the gate structure are encapsulated and processed to generate a maintenance prediction log that records degradation characteristics and maintenance suggestions.

[0170] S7.4: Based on the execution status of the emergency cut-off command and the storage status of the maintenance prediction log, perform integrity verification processing on the action feedback signal of the risk control module to generate a status confirmation signal indicating that the closed-loop risk control action has been completed.

[0171] Step S8: Based on the deviation feedback between the maintenance prediction log and the final diagnosis result, dynamically update the threshold parameter group of the lightweight decision tree model to optimize the accuracy of subsequent identification of multi-level degradation states of ultrasonic power devices, from insulation deterioration to precursors of local short circuits. Specifically, this includes: S8.1: Compare and analyze the Level 4 degradation status labels recorded in the maintenance prediction log with the final diagnosis results to generate a bias feedback vector characterizing the classification error of the current lightweight decision tree model. The bias feedback vector includes the difference in misjudgment level and the confidence level deviation. The final diagnosis result is a definitive conclusion obtained after verifying the labeled power devices through offline testing (such as high-precision electrical testing, microscopic observation, or laboratory analysis) during subsequent downtime maintenance cycles.

[0172] S8.2: Perform sensitivity backtracking calculation based on the deviation feedback vector to locate the critical feature dimension that causes the fourth-level degradation state label to be judged incorrectly, and generate an error attribution index pointing to the cold resistance sudden change response model or the brightness diffusion rate feature vector.

[0173] S8.3: The error attribution index is used to perform gradient correction processing on the corresponding node splitting threshold in the lightweight decision tree model to generate a dynamic adjustment step size that adapts to the aging characteristics of the current ultrasonic power device, thereby eliminating threshold drift caused by environmental temperature drift.

[0174] S8.4: Perform an iterative update operation on the threshold parameter set according to the dynamically adjusted step size to reconstruct the decision boundary of the lightweight decision tree model, generate an optimized threshold parameter set with self-learning ability, and ensure that the insulation performance slightly decreases to the level of four-level discrimination near the short circuit state.

[0175] S8.5: Write the optimized threshold parameter group into the non-volatile storage area of ​​the embedded MCU and replace the original configuration to complete the online version iteration of the lightweight decision tree model and achieve continuous improvement in the recognition accuracy of multi-level degradation state of ultrasonic power devices.

[0176] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0177] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0178] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for diagnosing early degradation of ultrasonic power devices, characterized in that, Specifically, it includes: S1: Within the transient window when the ultrasonic power supply is connected to high voltage, the instantaneous current waveform data of the circuit, the grayscale sequence of the incandescent lamp filament area and the near-infrared spectral response data are simultaneously acquired. S2: Perform slope calculation processing on the instantaneous current waveform data within the initial time window after power-on cold start, and extract the initial resistance change rate characteristic quantity that characterizes the resistance temperature coefficient response of the tungsten wire in the low temperature region. S3: Based on the initial resistance change rate characteristic quantity, combined with the physical law of transient current disturbance caused by microampere leakage current, construct a cold resistance sudden change response model and generate a primary degradation criterion; S4: For the pixel-level brightness evolution data in the grayscale sequence before the steady state, perform a spatial non-uniformity index calculation operation to generate a brightness diffusion rate feature vector. S5: Based on the near-infrared spectral response data, the spectral peak shift in the specified band is calibrated, the equivalent power consumption increment of the inversion loop is retrieved, and a correction factor for the average temperature rise rate of the tungsten wire is generated. S6: The primary degradation criterion, the brightness diffusion rate feature vector, and the tungsten filament average temperature rise rate correction factor are fused into a three-dimensional joint state feature set and input into the decision tree model to output the early degradation diagnosis results of the power device containing degradation state labels. S7: Based on the level value in the degradation status label, perform a graded early warning strategy matching operation. If it is determined to be a near short circuit state, trigger an emergency cut-off command. If it is determined to be a state of slight decrease in insulation performance or damage to the gate structure, generate a maintenance prediction log.

2. The method for diagnosing early degradation of ultrasonic power devices according to claim 1, characterized in that, After step S7, the method further includes: S8: Based on the feedback of the deviation between the maintenance prediction log and the final diagnosis result, dynamically update the threshold parameter group of the decision tree model.

3. The method for diagnosing early degradation of ultrasonic power devices according to claim 1, characterized in that, The instantaneous current waveform data is acquired synchronously using a high-speed current sensor; the grayscale sequence and near-infrared spectral response data of the incandescent lamp filament region are captured using a high frame rate image sensor.

4. The method for diagnosing early degradation of ultrasonic power devices according to claim 1, characterized in that, The degradation status labels include: normal, slightly decreased insulation performance, gate structure damage, and near short circuit.

5. The method for diagnosing early degradation of ultrasonic power devices according to claim 1, characterized in that, Step S3 specifically includes: The initial resistance change rate characteristic quantity is processed by time domain differentiation to obtain the instantaneous slope fluctuation sequence characterizing the dynamic response of the resistance temperature coefficient in the low-temperature region of the tungsten wire. Based on the instantaneous slope fluctuation sequence, an adaptive wavelet threshold denoising algorithm is performed to remove high-frequency electromagnetic interference and extract the pure transient current disturbance baseline signal. By using the pure transient current disturbance baseline signal and the physical law of transient current disturbance induced by microampere leakage current, a nonlinear fitting operation is performed to construct a cold resistance change response model that can quantify the relationship between the degree of insulation degradation and the magnitude of resistance jump. The initial resistance change rate characteristic is input into the constructed cold resistance abrupt change response model for mapping transformation processing, and the insulation degradation quantification index characterizing the probability of micro-breakdown risk of the gate oxide layer of the power device is calculated. Based on the aforementioned insulation degradation quantification index, a multi-level threshold comparison logic judgment operation is performed to generate a primary degradation criterion that clearly identifies the micro-breakdown risk level of the gate oxide layer.

6. The method for diagnosing early degradation of ultrasonic power devices according to claim 5, characterized in that, The initial cold start period of the filament is a time window of 0 to 20 milliseconds after the filament is energized.

7. The method for diagnosing early degradation of ultrasonic power devices according to claim 1, characterized in that, Step S4 specifically includes: The region of interest is extracted from the continuous frame images within the time window before the steady state in the grayscale sequence to obtain a dynamic filament grayscale matrix sequence containing the complete structure of the incandescent lamp filament. Based on the dynamic filament grayscale matrix sequence, coordinate calculation is performed on the bright areas in each frame of the image to generate a set of filament hot spot displacement vectors that characterize the trajectory of the luminous center position as it migrates over time. Based on the filament hot spot displacement vector set, a radial gray-scale gradient distribution statistical operation is performed to calculate the instantaneous brightness diffusion rate scalar, which reflects the speed at which heat diffuses from the center to the edge. Using the instantaneous brightness diffusion rate scalar, combined with the variance analysis algorithm, the spatial dispersion of pixel gray values ​​in the incandescent lamp filament region is quantitatively evaluated to generate a spatial non-uniformity index characterizing the non-uniformity of local Joule heat accumulation. Based on the spatial non-uniformity index and the filament hot spot displacement vector set, a multi-dimensional feature fusion mapping operation is performed to output the brightness diffusion rate feature vector that ultimately characterizes the difference between single-point micro-short circuits and planar insulation aging.

8. The method for diagnosing early degradation of ultrasonic power devices according to claim 1, characterized in that, The time window for the pre-steady-state phase is between twenty and one hundred milliseconds.

9. The method for diagnosing early degradation of ultrasonic power devices according to claim 1, characterized in that, Step S5 specifically includes: From the near-infrared spectral response data, spectral segments containing the characteristic thermal radiation spectrum of tungsten filaments are extracted to obtain narrowband near-infrared spectral sequence data containing the thermal radiation characteristics of tungsten filaments; Based on the narrowband near-infrared spectral sequence data, a peak detection algorithm is executed to extract the characteristic parameter of near-infrared spectral peak position shift that characterizes the instantaneous temperature change of the tungsten filament; By using the near-infrared spectral peak shift characteristic parameters and the mapping relationship of Planck's blackbody radiation law, a nonlinear fitting calculation is performed to generate an estimated value of the loop equivalent power consumption increment that reflects the real-time energy dissipation level of the loop. The net power temperature rise driving amount to remove environmental interference is generated by performing differential compensation calculation based on the estimated value of the equivalent power consumption increment of the circuit and the environmental reference temperature data collected by the environmental temperature sensor. Based on the net power temperature rise driving amount and the real-time voltage fluctuation coefficient provided by the power grid voltage monitoring module, a normalization correction process is performed to generate a tungsten wire average temperature rise rate correction factor that eliminates voltage fluctuation interference.

10. The method for diagnosing early degradation of ultrasonic power devices according to claim 1, characterized in that, The characteristic thermal radiation spectrum of the tungsten filament is in the 700-950 nanometer band.

Images