Power supply device aging detection method, aging detection model training method, and device

Abstract

The application discloses a power supply equipment aging detection method, an aging detection model training method and equipment, relates to the technical field of power electronics, and is aimed at a power supply equipment including a three-phase six-pulse rectifier and a direct-current bus capacitor group. The direct-current bus voltage ripple is subjected to time-frequency conversion and ripple form features are extracted. Frequency attention is obtained according to frequency statistical features of a current sample at multiple frequency positions. A physical frequency band prior is established with three times, six times and nine times of a current power grid fundamental frequency as the center. After the frequency attention and the physical frequency band prior are fused, the ripple form features are subjected to frequency band reweighting. The aging detection can not only represent morphological changes of the ripple with time and frequency, but also combine ripple change features reflected in different key frequency positions in the three-phase six-pulse rectification process, enhance feature distinguishability between different aging states, solve the problem that different aging states are difficult to accurately distinguish, and improve the aging detection accuracy of the power supply equipment.

Description

Technical Field

[0001] This invention relates to the field of power electronics technology, and in particular to a method for aging detection of power supply equipment, a training method for aging detection models, and equipment. Background Technology

[0002] With the widespread application of power electronic equipment in industrial power supply, energy storage, and special power supply scenarios, components such as rectifiers and DC bus capacitors may experience parameter drift and gradual aging during long-term operation, manifesting as corresponding changes in electrical signals such as bus voltage ripple. Therefore, monitoring the operating status of power supply equipment to promptly identify component aging is crucial for ensuring reliable equipment operation. Related technologies can assess the status of power supply equipment by collecting operating parameters such as voltage, current, and temperature, combined with spectrum analysis, time-frequency analysis, or data-driven models. However, the aging of different components in power supply equipment can lead to complex changes in operating signals, and different aging states may exhibit similar or overlapping behaviors. Existing detection methods still have limitations in distinguishing and identifying different aging states.

[0003] Therefore, how to improve the accuracy of identifying the aging status of power supply equipment is a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention provides a power supply equipment aging detection method, an aging detection model training method, and an apparatus, to at least solve the problem in related technologies where it is difficult to accurately distinguish between different aging states, resulting in low accuracy of power supply equipment aging detection.

[0005] To solve the above-mentioned technical problems, the present invention provides a method for aging detection of power supply equipment. The power supply equipment includes a three-phase six-pulse rectifier and a DC bus capacitor bank connected to the DC side of the three-phase six-pulse rectifier. The method for aging detection of power supply equipment includes: The DC bus voltage of the power supply equipment is collected, the bus voltage ripple is obtained based on the DC bus voltage, and the bus voltage ripple is transformed by time and frequency to obtain the ripple time and frequency diagram. Extract the ripple morphology features from the ripple time-frequency plot; Based on the frequency statistical features corresponding to multiple frequency positions extracted from the ripple time-frequency diagram, the frequency attention values ​​corresponding to the multiple frequency positions are determined; physical frequency band prior values ​​are established with the third, sixth, and ninth harmonics of the current power grid base frequency as the center; the frequency attention values ​​are fused with the physical frequency band prior values, and the ripple morphology features are reweighted based on the fusion result to obtain the reweighted ripple morphology features; The aging test results of the power supply equipment are determined based on the ripple morphology characteristics after frequency band reweighting.

[0006] To address the aforementioned technical problems, this invention also provides a training method for an aging detection model. The aging detection model is used for aging detection of power supply equipment, which includes a three-phase six-pulse rectifier and a DC bus capacitor bank connected to the DC side of the three-phase six-pulse rectifier. The training method includes: Obtain a training sample set, wherein the training samples in the training sample set include samples when the power supply device is in a healthy state and samples when it is in an aging state; The bus voltage ripple is obtained based on the DC bus voltage of the training sample, and the bus voltage ripple of the training sample is transformed by time and frequency to obtain the ripple time and frequency diagram corresponding to the training sample. The ripple time-frequency map corresponding to the training sample is input into the initial aging detection model to extract the ripple morphology features of the ripple time-frequency map corresponding to the training sample. Based on the frequency statistical features extracted from the ripple time-frequency map corresponding to the training sample, frequency attention values ​​corresponding to the multiple frequency positions are determined. Physical frequency band prior values ​​are established centered on the third, sixth, and ninth harmonics of the current power grid base frequency. The frequency attention values ​​are fused with the physical frequency band prior values, and the ripple morphology features are reweighted based on the fusion result to obtain the reweighted ripple morphology features corresponding to the training sample. The aging detection result corresponding to the training sample is determined based on the reweighted ripple morphology features corresponding to the training sample. The loss value is determined based on the difference between the aging detection results corresponding to the training samples and the labeled data of the training samples; The parameters of the initial aging detection model are adjusted based on the loss value and iteratively trained until the preset iteration termination condition is met, thus obtaining the aging detection model.

[0007] To address the aforementioned technical problems, the present invention also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the power supply device aging detection method or the training method of the aging detection model described above.

[0008] This invention addresses the challenges of accurately distinguishing between different aging states in power supply equipment, including three-phase six-pulse rectifiers and DC bus capacitor banks. It involves time-frequency transformation of the DC bus voltage ripple and extraction of ripple morphology features. Simultaneously, it obtains frequency attention based on the frequency statistical characteristics of the current sample at multiple frequency positions. Furthermore, it establishes physical frequency band priors centered on the third, sixth, and ninth harmonics of the current grid base frequency. By fusing the frequency attention with these physical frequency band priors, the ripple morphology features are reweighted. This allows aging detection to not only characterize the morphological changes of ripple over time and frequency but also to incorporate the ripple change characteristics reflected at different key harmonic positions during three-phase six-pulse rectification. This enhances the distinguishability between different aging states, solving the problem of inaccurate differentiation between different aging states in related technologies and improving the targeting and accuracy of power supply equipment aging detection. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A flowchart of a power supply equipment aging test method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a power supply equipment aging detection system provided in an embodiment of the present invention; Figure 3 This is a diagram illustrating the bus voltage ripple formation mechanism of a three-phase six-pulse rectifier provided in an embodiment of the present invention. Figure 4 A comparison diagram of the time-domain waveforms of bus voltage ripple under different aging conditions provided in an embodiment of the present invention; Figure 5 A comparison diagram of the frequency domain characteristics of bus voltage ripple under different aging conditions provided in the embodiments of the present invention; Figure 6 This is a comparison diagram of ripple time-frequency diagrams under different aging states provided in the embodiments of the present invention; Figure 7 A flowchart for constructing a ripple time-frequency diagram is provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a ripple morphology feature extraction module provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of a frequency attention module guided by physical frequency band prior, provided in an embodiment of the present invention. Figure 10This is a schematic diagram of the structure of an aging detection model provided in an embodiment of the present invention; Figure 11 A flowchart of another power supply device aging detection method provided in an embodiment of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0012] It should be noted that, in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0013] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] Here, we will first explain some key terms used in the embodiments of the present invention.

[0015] Three-phase six-pulse rectification refers to a working mode in which three-phase alternating current is rectified by a rectifier bridge composed of six rectifier diodes. Under ideal operating conditions, six sequentially occurring charging pulses are generated within one power grid cycle. Therefore, when the power grid fundamental frequency is... At that time, under healthy conditions, the main stable ripple on the DC side is concentrated in and its adjacent frequency bands.

[0016] Continuous Wavelet Transform (CWT) is a method for time-frequency analysis of non-stationary signals by changing the scale of the mother wavelet and shifting it along the time axis. CWT can simultaneously describe the changes in a signal at both time and frequency positions. This invention preferably employs complex Morlet continuous wavelet transform to balance frequency resolution at low frequencies and time resolution at higher frequencies.

[0017] The Short-Time Fourier Transform (STFT) is a time-frequency analysis method that uses a finite-length window function to extract local signals and performs Fourier transforms on each local signal. For the bus voltage ripple involved in this application, a fixed-window-length STFT can be used to form a time-frequency representation. However, when simultaneously representing stable main ridges and short-time spikes, gaps, and tails, a trade-off between time resolution and frequency resolution is necessary. Therefore, the embodiments of this invention preferably employ continuous wavelet transform.

[0018] Equivalent series resistance (ESR) is an equivalent parameter used to characterize actual capacitor losses. As bus capacitors age, their effective capacitance may decrease, and their ESR may increase. A decrease in effective capacitance increases the discharge ripple between adjacent rectified charging pulses; an increase in ESR increases the transient voltage drop during pulse charging.

[0019] An analog-to-digital converter (ADC) is used to convert the analog signal output from a bus voltage sensor into a discrete digital signal. A digital signal processor (DSP) can be used to perform processing such as sampling buffering, validity checking, mean removal, and detrending.

[0020] The Ripple Morphology Enhancement Module (RMEM) is a module used in this embodiment of the invention to extract ripple morphology features. The ripple morphology features in this embodiment may include temporal continuity features, frequency broadening features, local defect features, and edge abrupt change features. The RMEM and Ripple Morphology Enhancement Module in the accompanying drawings correspond to the ripple morphology feature extraction module in this embodiment of the invention.

[0021] The Frequency Prior-Guided Attention Module (FPGAM) is a module in this embodiment of the invention used to generate frequency attention values, introduce physical frequency band prior values, and perform frequency band reweighting on ripple morphology features. The physical frequency band prior, learnable weight fusion, feature recalibration, and output features shown in the accompanying figures correspond to the physical frequency band prior value, the fusion of the frequency attention value and the physical frequency band prior value, the frequency band reweighting, and the frequency band reweighted ripple morphology features in this embodiment of the invention, respectively.

[0022] In three-phase rectified power supply equipment, a bus capacitor is typically installed on the DC bus to store energy and reduce fluctuations in the rectified output voltage. The bus capacitor and rectifier diodes may experience gradual parameter changes during long-term operation. Because these devices are located on the rectification charging and bus energy storage paths, these parameter changes can be reflected in the DC bus voltage ripple.

[0023] Taking a three-phase six-pulse rectifier as an example, in an ideal diode bridge, at any given moment, the upper bridge arm diode corresponding to the highest voltage and the lower bridge arm diode corresponding to the lowest voltage are both conducting. Let the three-phase voltages be as follows: , and Then the instantaneous voltage on the DC side of the rectifier bridge can be expressed as: ; In the formula, This represents the instantaneous voltage on the DC side of the rectifier bridge. For the effective value of the line voltage... , power grid base frequency In a system operating under continuous current conditions, the average DC voltage across six pulses is approximately... The upper limit of the capacitor input filter's charging capacity is close to... .

[0024] One grid cycle consists of six charging pulses. The main ripple frequency in a healthy state can be expressed as: .

[0025] Between adjacent charging pulses, when the load can be equivalent to a steady DC bus current At that time, the peak-to-peak value of the discharge ripple of the bus capacitor can be approximately expressed as: ; In the formula, This represents the equivalent capacitance of the bus capacitor bank. This represents the approximate ripple peak-to-peak value obtained by considering only the discharge between adjacent charging pulses. , , For example, the theoretical order of magnitude is approximately Actual ripple is also affected by the rectifier conduction angle, line impedance, diode voltage drop, and equivalent series resistance of the capacitor.

[0026] Bus capacitor aging typically manifests as a decrease in effective capacitance and equivalent series resistance. An increase. A decrease in capacitance will increase the capacitor discharge slope. This increases the 300Hz main ridge line, trough depth, and ripple envelope. Equivalent series resistance The rise amplifies the transient voltage drop generated by the charging current pulse, making the time-domain waveform more prominent with spikes, steep edges and tails, and widening the harmonic energy band in the time-frequency diagram.

[0027] Solder fatigue, bonding wire aging, junction temperature cycling, or increased contact resistance in rectifier diodes can cause an increase in the forward voltage drop and dynamic resistance of a certain conduction path. At this time, the six charging pulses are no longer of equal amplitude, and the time-domain waveform shows periodic gaps, unequal peak heights, and local tails; in the frequency domain, in addition to the 300Hz main component, the grid fundamental frequency and its lower-order and odd-order combination components may also be generated.

[0028] Existing detection methods can determine device status through amplitude at a few fixed frequency points, fixed thresholds, or manual observation. However, a small number of frequency points cannot fully describe pulse gaps, spikes, tails, bandwidth broadening, and non-stationary changes. Changes in input voltage, load current, and temperature can also alter ripple amplitude, causing deviations in fixed thresholds under different devices or operating conditions. Aging of different devices may also simultaneously change the energy near the sixth harmonic, making it difficult to distinguish the source of aging when relying solely on a single amplitude change rate.

[0029] In addition, when the sampling rate Number of sampling points in the window At that time, the frequency interval of the Fast Fourier Transform (FFT) is 150Hz, 300Hz, and 450Hz typically do not strictly coincide with discrete frequency points. Directly reading a single frequency point results in a picket fence effect. Manually observing waveforms also relies on the experience of maintenance personnel, making it difficult to generate repeatable aging type and aging level outputs.

[0030] To address the problem of low accuracy in power supply equipment aging detection due to the difficulty in accurately distinguishing between different aging states in related technologies, this invention provides a power supply equipment aging detection method, apparatus, electronic device, and storage medium. For power supply equipment including a three-phase six-pulse rectifier and a DC bus capacitor bank, the invention performs time-frequency transformation on the DC bus voltage ripple and extracts ripple morphological features. Simultaneously, it obtains frequency attention based on the frequency statistical characteristics of the current sample at multiple frequency positions and establishes physical frequency band priors centered on the third, sixth, and ninth harmonics of the current grid base frequency. After fusing the frequency attention with the physical frequency band priors, the ripple morphological features are reweighted by frequency band. This allows aging detection to not only characterize the morphological changes of ripple over time and frequency but also to incorporate the ripple change characteristics reflected at different key harmonic positions during three-phase six-pulse rectification, thereby enhancing the feature differentiation between different aging states. This solves the problem of inaccurate distinction between different aging states in related technologies and improves the targeting and accuracy of power supply equipment aging detection.

[0031] The present invention provides a power supply equipment aging test method. The method is described in detail below in conjunction with the execution flow of the power supply equipment aging test method.

[0032] Figure 1 This is a flowchart of a power supply equipment aging detection method provided in an embodiment of the present invention.

[0033] The power supply equipment targeted in this embodiment of the invention includes a three-phase six-pulse rectifier and a DC bus capacitor bank connected to the DC side of the three-phase six-pulse rectifier. For example... Figure 1 As shown, the power supply equipment aging detection method provided in this embodiment of the invention may include steps S101 to S104.

[0034] S101: Collects the DC bus voltage of the power supply equipment, obtains the bus voltage ripple based on the DC bus voltage, and performs time-frequency transformation on the bus voltage ripple to obtain the ripple time-frequency diagram.

[0035] The ripple time-frequency diagram, obtained by time-frequency transformation of the bus voltage ripple in this embodiment of the invention, is a time-frequency input representation used for aging detection. During online detection, the ripple time-frequency diagram is formed from the DC bus voltage corresponding to the current detection window of the power supply device under test; during model training, the corresponding ripple time-frequency diagram can also be formed from the DC bus voltage of the training samples using the same processing method. The three-channel grayscale matrix in the attached figure is a specific implementation of the ripple time-frequency diagram, which can be composed of multiple channels formed by normalized time-frequency amplitude, time gradient, and frequency gradient.

[0036] Figure 2 This is a schematic diagram of a power supply equipment aging detection system provided in an embodiment of the present invention. Figure 2 As shown, the main power circuit under test can sequentially include a three-phase AC power supply, a three-phase six-pulse rectifier, a DC bus capacitor bank, an H-bridge inverter, and a superconducting coil load. The DC bus capacitor bank is connected to the DC side of the three-phase six-pulse rectifier.

[0037] In this embodiment of the invention, the superconducting coil load can be approximated as having a stable power requirement within the diagnostic frequency band of 50 Hz to 1000 Hz. The higher frequency switching components generated by the H-bridge inverter are suppressed by the bus voltage sensor bandwidth and the anti-aliasing low-pass filter. Therefore, in the low-frequency range of interest, the acquired bus voltage ripple is mainly affected by the six-pulse rectification charging process, the bus capacitor discharge process, rectifier asymmetry, and changes in device parameters.

[0038] Figure 2 In this configuration, a bus voltage sensor is connected to the DC bus side to acquire the DC bus voltage. The output of the bus voltage sensor can be converted into a digital signal by an ADC, and sampling and waveform buffering are performed by a DSP. Figure 2The preprocessing and time-frequency transformation in the middle form a ripple time-frequency diagram, which is to obtain the bus voltage ripple based on the DC bus voltage and perform time-frequency transformation. Figure 2 The intermediate aging detection network is a network implementation of the aging detection model, corresponding to subsequent ripple morphology feature extraction, frequency attention value determination, physical frequency band prior value establishment, frequency band reweighting, and aging detection result determination. Figure 2 The results of the aging test are only used as operation and maintenance information output and do not involve the real-time control, interlocking and protection of the H-bridge inverter.

[0039] On the field side, existing bus voltage sensors, ADCs, and DSPs can be reused for sampling and waveform buffering. Ripple time-frequency diagram construction and pre-trained aging detection model inference can run on a host computer or edge processing unit. This allows for obtaining input signals for aging detection without adding main circuit sensors and reduces modifications to the original control and protection channels.

[0040] Figure 3 This is a diagram illustrating the bus voltage ripple formation mechanism of a three-phase six-pulse rectifier provided in an embodiment of the present invention. Figure 3 As shown, Figure 3 (a) shows the three-phase phase voltages. , and , Figure 3 (b) shows the rectified voltage output of the three-phase six-pulse rectifier. Within a 20ms, 50Hz mains cycle, the rectified output generates six repetitive charging pulses. Therefore, in a healthy state, the main ripple frequency is 300Hz, i.e. . Figure 3 The physical relationships shown form the circuit basis for establishing the a priori values ​​of the physical frequency band, centered on the third, sixth, and ninth harmonics of the current power grid base frequency. Figure 3 (b) shows that one grid cycle consists of six charging pulses, and the main ripple frequency in a healthy state satisfies =6 When the power grid base frequency At 50Hz, It is 300Hz.

[0041] This embodiment of the invention can employ a three-phase 380V, 50Hz input, with a stable DC power of approximately 4.1kW. The rectifier diode module in the three-phase six-pulse rectifier can use devices with a 1200V reverse withstand voltage and a 50A rated current. 1200V is approximately 2.23 times the peak line voltage of 537.4V, which can cover positive grid deviations and surges. The 50A rated current provides a margin for the pulse charging current of the capacitor input rectification.

[0042] The DC bus capacitor bank can use four 6800 capacitors. 450V electrolytic capacitors, connected in a 2-series, 2-parallel configuration. Each series is equivalent to 3400kΩ. / 900V, equivalent capacitance after two series connected in parallel 6800 The withstand voltage is 900V. Each capacitor can be connected in parallel with 220kΩ capacitors. / 2W equalizing resistor. When each capacitor operates at approximately 270V, the power consumption of the equalizing resistor is approximately 0.33W; calculated based on the 450V limit voltage, the power consumption is approximately 0.92W, which is lower than the 2W rated power.

[0043] Under healthy conditions, the equivalent ESR of the bus capacitor bank can be taken as 25mΩ. Equivalent bus current 8A can be used. The equivalent incoming line parameters can be... , This represents the combined value of the two-phase cable, transformer leakage impedance, and optional incoming line reactor. The values ​​above are used to provide examples of feasible parameters. Equipment of different voltage and power ratings can be converted using the same circuit relationships.

[0044] The bus voltage sensor has a range of 0 to 800V and a bandwidth of at least 5kHz. The ADC sampling rate can be 10kHz, with a resolution of 16 bits, and 1024 points are collected per detection window. In this case, the Nyquist frequency is 5kHz, the window length is 102.4ms, and it contains approximately 30.7 300Hz ripple cycles. The anti-aliasing filter can be a fourth-order Butterworth low-pass filter with a cutoff frequency of 2kHz. This cutoff frequency is higher than the 450Hz critical frequency band and lower than the 5kHz Nyquist frequency, which can suppress aliasing of higher frequency switching components in the inverter.

[0045] S102: Extract the ripple morphology features from the ripple time-frequency plot.

[0046] Ripple morphology characteristics are used to describe the variations of bus voltage ripple in the time direction, frequency direction, and local time-frequency regions. Specifically, they can include time continuity characteristics, frequency broadening characteristics, local defect characteristics, and edge abrupt change characteristics. Specific implementation methods are described in the following examples.

[0047] S103: Based on the frequency statistical features corresponding to multiple frequency positions extracted from the ripple time-frequency diagram, determine the frequency attention values ​​corresponding to multiple frequency positions; establish physical frequency band prior values ​​centered on the third, sixth, and ninth harmonics of the current power grid base frequency; fuse the frequency attention values ​​with the physical frequency band prior values, and perform frequency band reweighting on the ripple morphology features based on the fusion result to obtain the frequency band reweighted ripple morphology features.

[0048] Among them, the frequency attention value reflects the relative importance of different frequency positions in the current ripple time-frequency diagram for aging detection. The physical band prior value reflects the physical information of the third, sixth, and ninth harmonics corresponding to the three-phase six-pulse rectification. After the two are fused, the frequency distribution of the current ripple time-frequency diagram itself and the band information of the six-pulse rectification can be utilized simultaneously. The specific implementation method is described in the following examples.

[0049] S104: Determine the aging test results of the power supply equipment based on the ripple morphology characteristics after frequency band reweighting.

[0050] The aging test results can include aging type and aging level. Aging type can include healthy, bus capacitor aging, rectifier diode aging, and mixed aging. Aging level can include normal, warning, and alarm. The aging type and aging level can be determined by the corresponding output head based on the ripple morphology characteristics after frequency band reweighting. Specific implementation methods are described in the following examples.

[0051] Through steps S101 to S104 described above, the morphological changes caused by aging can be extracted from the ripple time-frequency diagram of the power supply device under test. Then, based on the frequency statistical characteristics of the current ripple time-frequency diagram and the physical frequency band information of the six-pulse wave, these morphological changes are reweighted by frequency band. Therefore, the differences in ripple morphology and key octave positions under different aging states can be used for aging detection.

[0052] Based on the above embodiments, in this embodiment of the invention, S101, performing time-frequency transformation on the bus voltage ripple to obtain a ripple time-frequency diagram, may include: performing continuous wavelet transform on the bus voltage ripple to obtain wavelet coefficients corresponding to multiple frequency positions and multiple time positions, and performing logarithmic compression on the amplitude of the wavelet coefficients; normalizing the logarithmically compressed amplitude according to a fixed calibration scale determined when the power supply equipment is in a healthy state; and forming an amplitude channel, a time gradient channel, and a frequency gradient channel according to the normalized time-frequency amplitude, the gradient of the time-frequency amplitude along the time direction, and the gradient of the time-frequency amplitude along the frequency direction, respectively, to obtain a ripple time-frequency diagram.

[0053] Figure 4 This is a comparison diagram of the time-domain waveforms of bus voltage ripple under different aging conditions provided in the embodiments of the present invention. Figure 4 The diagram shows the bus voltage ripple corresponding to healthy condition, mild aging of bus capacitors, severe aging of bus capacitors, severe aging of rectifier diodes, and mixed aging. Among these, Figure 4 (a) indicates a healthy state. Figure 4 (b) shows the bus capacitor in a slightly aged state. Figure 4 (c) indicates that the bus capacitors are in a severely aged state. Figure 4 (d) represents the severely aged state of the rectifier diode. Figure 4 (e) represents the mixed aging state of the capacitor and rectifier diode.

[0054] Under healthy conditions, the conduction intervals of the six rectifier diodes are approximately equivalent, and the ripple exhibits a relatively stable six-pulse periodic characteristic. When the effective capacitance of the bus capacitor decreases, the discharge amplitude between adjacent charging pulses increases. Figure 4 The ripple amplitude of a severely aged bus capacitor is greater than that of a healthy state. When a single rectifier diode or a single conducting branch ages, some charging pulses may exhibit reduced amplitude, delay, gaps, or tails, causing the ripple to show periodic uneven heights. Mixed aging includes waveform characteristics resulting from changes in both bus capacitor parameters and rectifier branch parameters.

[0055] Figure 5 This is a comparison diagram of the frequency domain characteristics of bus voltage ripple under different aging conditions provided in the embodiments of the present invention. Figure 5 The full-band logarithmic spectrum and the locally amplified spectrum from 100Hz to 520Hz are shown. With a grid fundamental frequency of 50Hz, 150Hz, 300Hz, and 450Hz correspond to the third, sixth, and ninth harmonics, respectively. Bus capacitor aging primarily increases the sixth harmonic component; rectifier diode aging more easily causes changes in the third and ninth harmonics and their adjacent frequency bands, reflecting the disruption of the six-pulse symmetry. Hybrid aging can simultaneously exhibit sixth harmonic enhancement and asymmetric frequency band changes. Among these, Figure 5 (a) is the full-band logarithmic coordinate spectrum to ensure that the health status curve is visible; Figure 5 (b) is the locally amplified spectrum from 100 Hz to 520 Hz.

[0056] Figure 6 This is a comparison diagram of ripple time-frequency diagrams under different aging states provided in the embodiments of the present invention. Figure 6 The horizontal axis represents time, the vertical axis represents frequency, and the grayscale value represents the normalized time-frequency amplitude. Figure 6 The health status of the components, the severe aging of the bus capacitor, the severe aging of the rectifier diode, and the mixed aging exhibit different ridge amplitudes, ridge continuity, local textures, and bandwidth broadening, respectively. Figure 6 Severe aging of the bus capacitor, severe aging of the rectifier diode, and mixed aging of the capacitor and diode are specific states of bus capacitor aging, rectifier diode aging, and mixed aging, respectively, in the embodiments of this invention. Figure 6 (a) is the ripple time-frequency diagram under healthy conditions. Figure 6 (b) is the time-frequency diagram of ripple under severe aging of the bus capacitor. Figure 6 (c) is the time-frequency diagram of the ripple of the rectifier diode under severe aging conditions. Figure 6 (d) is the time-frequency diagram of ripple under the mixed aging state of capacitor and rectifier diode. Figure 6 (e) is the gray scale corresponding to the normalized time-frequency amplitude, with gray values ​​ranging from 0 to 1 indicating that the normalized time-frequency amplitude ranges from low to high.

[0057] Figure 7 This is a flowchart illustrating the construction of a ripple time-frequency diagram, as provided in an embodiment of the present invention. Figure 7 The sampling sequence corresponds to the acquired DC bus voltage digital sequence; mean removal and detrending to obtain ripple correspond to obtaining the bus voltage ripple based on the DC bus voltage; Morlet continuous wavelet transform is a type of continuous wavelet transform; logarithmic compression and fixed calibration correspond to logarithmically compressing the wavelet coefficient amplitude and normalizing it according to a fixed calibration scale based on the healthy state; the three-channel grayscale matrix corresponds to the amplitude channel, time gradient channel, and frequency gradient channel in this embodiment of the invention. The three channels together constitute the ripple time-frequency diagram. Figure 7 middle, Indicates the amplitude channel. Represents the time gradient channel. This represents the frequency gradient channel; both the training and inference phases use fixed grayscale upper and lower limits obtained from health calibration, without independently normalizing individual samples.

[0058] In practical implementation, the DSP can be configured according to the sampling rate. Collect bus voltage. Each detection window contains... There are multiple sampling points. Detection can be triggered during the steady-state phase of the equipment, at fixed times each day, or during the stable phase after shutdown.

[0059] Before the bus voltage ripple is generated, the sampling sequence can be checked for saturation, missing samples, and abnormal jumps. For the ADC-calibrated bus voltage samples that pass the validity check... The bus voltage ripple sequence can be obtained by removing the mean and linear detrending. : ; in, This indicates linear detrending processing, used to remove the linear trend term from the DC bus voltage sequence within the sampling window.

[0060] During the health calibration phase, a unified ripple amplitude scale and grayscale upper and lower limits can be determined. This fixed calibration scale can be used for both the ripple time-frequency plots generated during model training and those generated during actual inference. This process eliminates the need to determine normalization upper and lower limits for each ripple time-frequency plot individually, thus preserving information about the absolute ripple amplitude increase caused by bus capacitor aging.

[0061] The embodiments of this invention preferably employ complex Morlet continuous wavelet transform. The continuous wavelet transform of the bus voltage ripple can be expressed as: ; In the formula, The continuous wavelet transform coefficients represent the bus voltage ripple. This represents the bus voltage ripple sequence. Indicates scale. Represents a time variable. Indicates time shift, The symbol represents the complex Morlet mother wavelet, and the asterisk represents the complex conjugate. In one specific embodiment, the Morlet center parameter can be 6, and 96 frequency positions are set at logarithmic intervals in the range of 50Hz to 1000Hz, thereby obtaining a wavelet coefficient matrix of 96×1024.

[0062] Logarithmic compression of the wavelet coefficient amplitudes yields: ; in, This represents the time-frequency amplitude after logarithmic compression. Indicates the actual frequency. The wavelet coefficients representing the bus voltage ripple are written as functions of the corresponding frequency and time position when using frequency coordinates.

[0063] Fixed lower limit obtained using health calibration and fixed upper limit After normalization, we get: ; in, This represents the normalized time-frequency amplitude. This represents the fixed lower limit obtained from health calibration; This represents the fixed upper limit obtained from health calibration; This indicates a truncation operation that restricts the calculation result to a specified range of values.

[0064] Based on the normalized time-frequency amplitude, three channels can be formed: ; in, Indicates the amplitude channel; Indicates the time gradient channel; Indicates the frequency gradient channel; Indicates the sign of partial derivatives; Indicates time shift; Indicates the actual frequency. Amplitude channel. Used to preserve absolute time-frequency energy; time gradient channel Used to highlight the appearance, disappearance, and tailing of pulses; frequency gradient The channels are used to highlight the boundaries of harmonic ridges and broaden the bandwidth. The three channels can be scaled to a matrix size of 224×224 and together form a 224×224×3 ripple time-frequency diagram.

[0065] Through the above processing, stable sixth harmonic ridges and short-term spikes, gaps, and tails can be represented in the same ripple time-frequency plot, providing input for subsequent ripple morphology feature extraction.

[0066] The embodiments of the present invention further describe the ripple morphology features and their extraction methods.

[0067] In this embodiment of the invention, the ripple morphology features may include time continuity features, frequency broadening features, local defect features, and edge abrupt change features; wherein, the time continuity features are used to characterize the continuity, periodicity, and duration of the ripple features at the same frequency position over time; the frequency broadening features are used to characterize the energy diffusion relationship between adjacent frequency positions; the local defect features are used to characterize local abnormal textures within a preset time range and a preset frequency range; and the edge abrupt change features are used to characterize local changes in the time direction and the frequency direction.

[0068] Temporal continuity characteristics can be used to describe the response of the same frequency position over multiple consecutive charge-discharge cycles. After the effective capacitance of the bus capacitor decreases, the sixth-harmonic main ripple may synchronously increase over multiple consecutive cycles. After the rectifier diodes age, some pulses may be missing or of unequal height, causing periodic variations or local discontinuities in the intensity of the corresponding frequency ridge.

[0069] Frequency broadening characteristics can be used to describe energy diffusion between adjacent frequency positions. When the bus capacitance ESR increases, the charging pulse edges become steeper, allowing energy to diffuse to adjacent frequency positions. Rectifier diode asymmetry can also cause energy diffusion from the sixth harmonic band to the third, ninth harmonic, and adjacent frequency bands.

[0070] Local defect features can characterize local gaps, isolated high-energy regions, locally thickened ridges, and local variations between multiple frequency bands within a two-dimensional neighborhood defined by a preset time range and a preset frequency range. The preset time range and preset frequency range can be determined based on the coverage of the convolution kernel or the spatial scale of the preceding feature map. For example, in the following 3×3 local depthwise convolution implementation, a local neighborhood consists of the current feature location and its adjacent locations in the time and frequency directions.

[0071] Edge mutation characteristics can characterize the local rate of change in both the time and frequency directions. Peaks, steep edges, and tails caused by rising ESR, as well as ridge line breaks and notch edges caused by aging of rectifier diodes, can all produce significant gradient changes.

[0072] In this embodiment of the invention, extracting ripple morphology features from the ripple time-frequency map may include: performing channel mapping on the input features extracted from the ripple time-frequency map, and inputting the channel-mapped features into the time-direction ridge branch, the frequency-direction broadening branch, the local texture branch, and the bidirectional gradient branch, respectively; extracting temporal continuity features along the time direction through the time-direction ridge branch, extracting frequency broadening features along the frequency direction through the frequency-direction broadening branch, extracting local defect features in the two-dimensional neighborhood of the time and frequency directions through the local texture branch, and extracting edge abrupt change features based on the local gradients of the time and frequency directions through the bidirectional gradient branch; fusing the outputs of the time-direction ridge branch, the frequency-direction broadening branch, the local texture branch, and the bidirectional gradient branch, and adjusting the fused incremental features according to the channel gating results to obtain ripple morphology features.

[0073] Figure 8 This is a schematic diagram of a ripple morphology feature extraction module provided in an embodiment of the present invention. Figure 8 As shown, RMEM is used to extract ripple morphology features. Figure 8 In this embodiment of the invention, the temporal depthwise convolution corresponds to the temporal ridge branch, the frequency-oriented depthwise convolution corresponds to the frequency-oriented broadening branch, the local depthwise convolution corresponds to the local texture branch, and the Sobel gradient branch corresponds to the bidirectional gradient branch. Figure 8 Channel concatenation, fusion convolution, and channel gating are used to fuse the outputs of the four branches and adjust the incremental features formed after fusion. Figure 8 The residual summation is used to perform residual processing between the adjusted incremental features and the input features.

[0074] Let the input features be ,in, Indicates the number of network characteristic channels. Indicates the height of the frequency axis. This represents the width of the time axis. First, channel mapping is performed using a 1×1 convolution: ; in, Indicates channel mapping operation; This represents the features after channel mapping.

[0075] The channel-mapped features are then divided into four branches: ; ; ; ; In the formula, This indicates the output of the time-to-ridge branch. This indicates the output of the frequency-widening branch. This represents the output of a local texture branch. This represents the output of the bidirectional gradient branch. This indicates depthwise convolution (DWConv), and the subscript indicates the kernel size. and These represent the Sobel kernels in the time and frequency directions, respectively. This represents the convolution operation. This represents a tiny positive number used to prevent the square root from being zero.

[0076] The outputs of the four branches can be concatenated and merged: .

[0077] Incremental characteristics are adjusted via channel gating: ; ; In the formula, This represents the incremental feature formed by the fusion of the four branches. This indicates a branch merging operation. This indicates a channel splicing operation. This indicates the output of the time-to-ridge branch. This indicates the output of the frequency-widening branch. This represents the output of a local texture branch. This represents the output of the bidirectional gradient branch; Indicates the channel gating result. This represents the Sigmoid function. and Indicates the gating mapping parameters, This represents the SiLU activation function. This represents element-wise multiplication. This represents the output of the ripple morphology feature extraction module. Global Average Pooling (GAP) is used to form the channel descriptions required for channel gating.

[0078] Temporal ridge branching preserves the response of continuous harmonic ridges, frequency broadening branching identifies variations in energy diffusion to adjacent frequency positions, local texture branching extracts local gaps and high-energy regions, and bidirectional gradient branching enhances edge variations such as spikes, breaks, and tails. Using depthwise convolution reduces the number of parameters. Residual processing reduces the over-enhancement of healthy, stable main ridges.

[0079] In this embodiment of the invention, the determination of frequency attention values, the establishment of physical frequency band prior values, and the fusion of frequency attention values ​​and physical frequency band prior values ​​in S103 may include: averaging the input features extracted from the ripple time-frequency map along the channel direction and the time direction to obtain frequency statistical features corresponding to multiple frequency positions, and performing one-dimensional convolution processing on the frequency statistical features to obtain frequency attention values; determining the ridge center frequency corresponding to the maximum time-frequency energy in the ripple time-frequency map within a preset sixth harmonic search range, and determining the current power grid base frequency based on one-sixth of the ridge center frequency; determining the center frequencies corresponding to the third, sixth, and ninth harmonics based on the current power grid base frequency, setting learnable frequency band amplitude coefficients and bandwidths for the third, sixth, and ninth harmonics respectively, and determining physical frequency band prior values ​​corresponding to multiple frequency positions based on the actual frequency, center frequency, frequency band amplitude coefficient, and bandwidth of each frequency position; and fusing the frequency attention values ​​and physical frequency band prior values ​​using non-negative learnable fusion weights.

[0080] Figure 9 This is a schematic diagram of the structure of a frequency attention module guided by physical frequency band prior, provided in an embodiment of the present invention.

[0081] Figure 9 The frequency description obtained after averaging along the middle channel and time is the frequency statistical feature in this embodiment of the invention; the learned frequency attention after one-dimensional convolution is the frequency attention value in this embodiment of the invention. Figure 9 The physical frequency band prior in this invention refers to the physical frequency band prior value in the embodiments of this invention. Figure 9 The learnable weight fusion in this embodiment refers to the use of non-negative learnable fusion weights to fuse the frequency attention value and the physical frequency band prior value. Figure 9 Feature recalibration in this invention refers to frequency band reweighting in this embodiment. Figure 9 The output features are the ripple morphology features after frequency band reweighting in the embodiments of the present invention. Figure 9 Channel attention is also shown, which is a further implementation of feature reweighting in the embodiments of the present invention.

[0082] For input features Averaging along the channel direction and the time direction yields the [number]. Frequency statistical characteristics of each frequency position: ; in, Indicates the first Frequency statistical characteristics corresponding to each frequency position; Indicates the frequency position index; Indicates the network feature channel index; Indicates the time location index; Indicates the number of network feature channels; Indicates the width of the time axis.

[0083] Learning-based frequency attention can be obtained through one-dimensional convolution: ; in, This represents one-dimensional convolution processing; This represents the learning frequency attention value.

[0084] Based on the power grid base frequency estimate corresponding to the current window Physical frequency band prior values ​​can be established: ; In the formula, This represents the prior value of the physical frequency band. Indicates the first The actual frequency corresponding to each frequency position Indicates the number of prior frequency bands. Indicates the first The preferred values ​​for the harmonics of the prior frequency bands are 3, 6, and 9. Indicates the first The frequency band amplitude coefficient of each frequency band Indicates the first The bandwidth of each frequency band, in Hz; This represents the current estimated baseband frequency of the power grid, in Hz.

[0085] In one implementation, when there is severe rectification asymmetry, the doubling factor can be extended to 1, 2, 3, 6, and 9, thereby preserving lower-order asymmetric components.

[0086] The current grid fundamental frequency can be obtained using grid frequency measurements provided by the existing control system. Alternatively, it can be estimated from the sixth harmonic main ripple within the current bus voltage ripple window. Specifically, the center of the maximum time-frequency energy ridge can be extracted within a preset sixth harmonic search range. And calculate: .

[0087] After obtaining the current power grid base frequency, the three physical prior centers are updated as follows: .

[0088] The preset sixth harmonic search range can be determined based on the allowable range of grid frequency variation for the device. For example, for a device with a nominal grid base frequency of 50Hz and an allowable base frequency variation within the range of 49Hz to 51Hz, the sixth harmonic search range can be set to 294Hz to 306Hz. For other allowable frequency ranges, the upper and lower limits of the allowable base frequency range can be multiplied by 6 to obtain the corresponding search range.

[0089] The bandwidth of the three frequency bands can also be adjusted according to the fluctuation of the power grid frequency estimation results. When the fundamental frequency estimation values ​​of multiple consecutive detection windows fluctuate significantly, the bandwidth of the physical frequency band can be increased accordingly to reduce the possibility that slight frequency drift will cause the target ripple ridge to fall outside the prior frequency band. When the fundamental frequency estimation is stable, a narrower bandwidth can be used. The specific bandwidth can be determined through calibration using multiple sets of samples under healthy and known aging conditions.

[0090] The power equipment aging detection method provided in this embodiment of the invention may further include: calculating the current grid base frequency based on the sixth harmonic main ripple, and updating the third, sixth, and ninth harmonic physical frequency bands according to the current grid base frequency. Specifically, the energy ridge center of the sixth harmonic main ripple is determined in the time-frequency data corresponding to the current bus voltage ripple; the current grid base frequency is determined according to the correspondence between the energy ridge center and the sixth harmonic; and the center positions of the third, sixth, and ninth harmonic physical frequency bands are updated according to the current grid base frequency. Thus, the physical frequency band positions can change with the actual grid frequency without adding a dedicated grid frequency sensor.

[0091] The sixth harmonic is the main stable ripple under healthy three-phase six-pulse rectification. The third and ninth harmonics are more sensitive to asymmetric textures caused by disruption of the symmetry of the six conduction intervals. The third harmonic belongs to the lower frequency asymmetric modulation component, while the ninth harmonic is more sensitive to sharper gaps, pulse edge changes, and local distortions.

[0092] To further describe the relative contributions of different aging types in the three key frequency bands, we can define the first... Cumulative time-frequency energy of each octave band: ; In the formula, Indicates the first The cumulative time-frequency energy of each octave band within the entire detection window Indicates A set of frequencies centered on a certain bandwidth Indicates the actual frequency. Indicates time and location. The wavelet coefficients represent the bus voltage ripple at the corresponding time-frequency position.

[0093] The relative energy relationship between the third and sixth harmonics can be further calculated. And the relative energy relationship between the ninth harmonic and the sixth harmonic. . It can describe the relative contribution of the low-order asymmetric frequency band to the main ripple of the six-pulse wave. This can describe the relative contribution of the higher-frequency asymmetric band to the six-pulse main ripple. When the effective capacitance of the bus capacitor decreases, The main trend is increasing, while The changes are relatively limited. Aging of a single rectifier branch is more likely to cause… and Increase. Mixed aging can simultaneously manifest as... Increased frequency band size and asymmetric frequency band contribute more to the relative increase.

[0094] Table 1 shows a comparison of key octave bands and ripple time-frequency characteristics under different aging modes.

[0095] Table 1

[0096] As can be seen from Table 1, the aging of the bus capacitor and the aging of the rectifier diode show different performance in the three key frequency bands. Therefore, establishing the physical frequency band prior values ​​with the third, sixth and ninth harmonics as the center and reweighting the ripple morphology characteristics by frequency band can provide a basis for distinguishing different aging types.

[0097] The power supply equipment aging detection method provided in this embodiment of the invention may further include: determining the cumulative time-frequency energy of the corresponding frequency bands for third, sixth, and ninth harmonics, and forming auxiliary features to distinguish between bus capacitor aging, rectifier diode aging, and mixed aging based on the relative energy relationships between the third and sixth harmonics and between the ninth and sixth harmonics. These auxiliary features can be input into the aging detection network along with ripple morphology features, or used to interpret the network output. Thus, the difference between the symmetrical change of the six-pulse period and the uniform discharge process of the bus capacitor can be utilized to provide additional basis for subsequent aging type determination.

[0098] Learning-based frequency attention and physical frequency band priors can be fused using non-negative learnable fusion weights: ; In the formula, The non-negative learnable fusion weights represent learning-oriented frequency attention; Non-negative learnable fusion weights representing prior values ​​of physical frequency bands. This represents the frequency attention after fusion. The fusion weights can be optimized during training.

[0099] Simultaneously, channel attention can be formed from input features. The input features are then reweighted based on the fused frequency attention and channel attention. ; In the formula, This represents the retention factor used to prevent the channel from being over-suppressed; it can be set to 0.5. This indicates the ripple morphology characteristics after frequency band reweighting.

[0100] Figure 10 This is a schematic diagram of the structure of an aging detection model provided in an embodiment of the present invention.

[0101] like Figure 10 As shown, the input can be a 224×224×3 ripple time-frequency map. The input passes through a lightweight convolutional initial feature extraction stage (Stem) before entering the shallow and middle layers. The shallow and middle layers are configured with RMEM (Resource-based Feature Extraction Model) to extract ripple morphology features such as temporal continuity, frequency broadening, local defects, and edge abrupt changes. The RMEM output is then input into the FPGAM (Field-Package Functional Model) set in the higher and deeper layers. The FPGAM performs frequency band reweighting on the extracted ripple morphology features based on the third, sixth, and ninth harmonics physical frequency band information. Subsequently, global pooling is used to form shared features, which are then input into the aging type header and aging level header, respectively.

[0102] Figure 10 The shallow stage RMEM, mid-stage RMEM, high-stage FPGAM, and deep stage FPGAM represent a preferred configuration. The number of RMEMs and FPGAMs can be adjusted according to the number of layers in the lightweight classification backbone. Preferably, RMEMs are placed in the shallow or mid-stages that still have high time-frequency resolution, while FPGAMs are placed in the mid-to-high stages that have a larger receptive field. Figure 10 A lightweight, deep separable classification backbone can be selected using a YOLO26n-cls-like structure. This network is used for state classification of the entire ripple time-frequency map and is not used to output target boxes.

[0103] Besides the aforementioned lightweight, deep separable classification backbone, one-dimensional convolutional networks, larger visual Transformers, ResNet, or ConvNeXt can also be used to process the same input. One-dimensional convolutional networks can directly process temporal data, but they are not conducive to explicitly utilizing two-dimensional harmonic ridges and cross-band textures; larger visual Transformers have a large number of parameters and training sample requirements; general-purpose ResNet or ConvNeXt lacks specific physical frequency band information for the ripple time axis and frequency axis. Therefore, the embodiments of this invention preferably use a lightweight, deep separable classification backbone.

[0104] For the continuous sixth-harmonic enhancement characteristic caused by the decrease in effective capacitance of the bus capacitor, FPGAM mainly enhances the morphological response already extracted by RMEM near the sixth harmonic. For the notches, breaks, and asymmetric textures caused by aging of rectifier diodes, FPGAM mainly enhances the corresponding morphologies in the third harmonic, ninth harmonic, and adjacent frequency bands. For mixed aging, both sets of responses can be retained simultaneously. Thus, RMEM first determines the morphological changes of ripple occurrence, and FPGAM then determines the frequency positions of these morphological changes.

[0105] In this embodiment of the invention, determining the aging test result of the power supply equipment based on the ripple morphology characteristics after frequency band reweighting in S104 may include: when the power supply equipment is in a healthy state, establishing multiple healthy operating condition benchmarks according to different combinations of bus voltage, load current, and temperature, and saving the benchmark characteristics under the healthy state for each healthy operating condition benchmark. The benchmark characteristics include ripple amplitude, third harmonic band energy, sixth harmonic band energy, and ninth harmonic band energy; determining the target healthy operating condition benchmark from the multiple healthy operating condition benchmarks based on the current bus voltage, load current, and temperature, and saving the ripple amplitude, third harmonic band energy, sixth harmonic band energy, and ninth harmonic band energy of the current detection cycle. The frequency harmonic band energy and the ninth harmonic band energy are compared with the baseline characteristics corresponding to the target healthy operating condition baseline. The network output of the current detection cycle is obtained based on the ripple morphology characteristics after frequency band reweighting. The aging detection result is determined by combining the comparison results and the network output of multiple consecutive detection cycles. When the ripple abnormality occurs during changes in bus voltage, load current, or temperature and disappears after the corresponding operating condition is restored, the ripple abnormality is identified as an operating condition disturbance. When the ripple abnormality persists for multiple detection cycles or shows a monotonically increasing trend under the operating condition corresponding to the target healthy operating condition baseline, the ripple abnormality is used as aging evidence to determine the aging detection result.

[0106] During the health calibration phase after initial commissioning or major overhaul, multiple health condition benchmarks can be established based on different combinations of bus voltage, load current, and temperature. Each health condition benchmark can store the ripple amplitude, third-order frequency band energy, sixth-order frequency band energy, ninth-order frequency band energy, and network characteristic distribution under healthy conditions. During online comparison, the health condition benchmark uses the same defined network characteristic distribution as the current testing cycle.

[0107] Once the current testing cycle begins, the current bus voltage, load current, and temperature can be compared with parameters corresponding to multiple health condition benchmarks. The health condition benchmark with the smallest parameter difference is selected as the target health condition benchmark. In one specific implementation, the relative deviation of the average bus voltage in the current window from the health calibration value can not exceed ±3%, the relative deviation of the equivalent load current can not exceed ±5%, and the difference between the ambient temperature or bus capacitor case temperature and the health calibration temperature can not exceed ±5℃. If any of these allowable ranges are exceeded, the current window can be marked as a condition mismatch, and resampling will be performed after the condition is restored.

[0108] The ±3%, ±5%, and ±5℃ values ​​mentioned above are recommended initial values. For devices with different power ratings or different thermal characteristics, the permissible range can be determined during the health calibration phase using multiple sets of health samples. Specifically, under the condition that the device's health is confirmed, the input voltage, load, and temperature can be changed separately, and the corresponding ripple time-frequency characteristics recorded. When the deviation of the health sample characteristics caused by a change in a certain operating condition variable is still within the statistical distribution range of the health samples, that range can be included in the permissible operating condition interval. When it exceeds this range, direct comparisons across operating conditions are not performed.

[0109] Figure 11 A flowchart illustrating another power supply device aging detection method provided in an embodiment of the present invention. Figure 11 As shown, the process includes: S201: Determine whether the operating conditions meet the voltage, load, and temperature ranges; otherwise, execute S211. S202: Collect bus voltage samples at 1024 points; S203: Determine if the waveform is complete and without saturation; otherwise, execute S211. S204: Obtain from , , The constructed ripple time-frequency diagram; S205: Network Inference; S206: Continuous 5-window voting and confidence level screening; S207: Determine the aging test results; S208: When the aging level is 0 (normal), maintain periodic monitoring; S209: When the aging level is a Level 1 warning, retest and arrange for inspection; S210: When the aging level alarm is level 2, stop the machine for maintenance as soon as possible; S211: This result is invalid. Please resample after the sample has stabilized.

[0110] It should be noted that, Figure 11 The S205 network inference process involves inputting the ripple time-frequency map into the aging detection model and obtaining the network output. Figure 11S204 in the process of obtaining the ripple time-frequency diagram , , This refers to the ripple time-frequency diagram construction described in the above embodiments. Figure 11 The number of data points collected is N, which is equal to 1024. This refers to the DC bus voltage acquisition described in the above embodiments. Figure 11 The handling measures at each level are for operation and maintenance reference only and do not directly trigger control or protection actions.

[0111] In practical implementation, for S206, a single window outputs the probability of aging type and aging level. When the confidence level of the highest level is insufficient, it outputs "to be retested". It is preferable to continuously collect data from 5 windows, with at least 3 windows outputting the same level to form the final result. This processing is used to reduce false alarms caused by random noise. For level 0 (normal), periodic monitoring is maintained; for level 1 (early warning), retesting is performed under the same operating conditions, and capacitance, ESR, and diode voltage drop checks are arranged; for level 2 (alarm), it is recommended to arrange shutdown and maintenance as soon as possible. The detection triggering time is preferably once a day or after each shutdown, and preferably executed during steady state, at a fixed time each day, or during the stable phase after shutdown.

[0112] The power supply equipment aging detection method provided in this embodiment of the invention may further include: obtaining the aging type probability and aging level probability corresponding to multiple consecutive detection windows; performing confidence screening on the single-window output; and determining the current detection result based on the number of windows with the same aging level appearing in the multiple windows that have passed the confidence screening. In one embodiment, five consecutive windows may be used, and the final level is formed when at least three windows output the same level. This process can reduce the probability of a single detection window being affected by accidental noise and resulting in abnormal output.

[0113] S207 confirms the aging test results. In the attached diagram, if S207 is divided into Level 0 (Normal), Level 1 (Warning), and Level 2 (Alarm), these correspond to Normal, Warning, and Alarm, respectively. S208 maintains periodic monitoring; under normal conditions, periodic testing continues. S209 retests and arranges inspections; under warning conditions, retesting is performed under the same operating conditions, and checks on the effective capacitance of the bus capacitor, ESR, and forward voltage drop of the rectifier diodes can be arranged. S210 prompt shutdown for maintenance; under alarm conditions, equipment maintenance is recommended. The above actions are maintenance information outputs and do not automatically trigger H-bridge inverter control, interlocking, or protection actions.

[0114] In historical trend analysis, key ripple characteristics and network outputs across multiple consecutive testing cycles can be stored. When ripple anomalies occur only during load, voltage, or temperature changes and disappear synchronously after the corresponding operating conditions recover, the anomaly can be identified as an operating condition disturbance. When ripple anomalies persist or show a monotonically increasing trend across multiple testing cycles under the same target health baseline, the anomaly can be used as evidence of aging. This reduces the likelihood of input voltage, load, and temperature changes being mistakenly identified as device aging.

[0115] Subsequent testing can be preferably performed once daily or triggered after each shutdown. Under healthy equipment conditions, a health calibration standard can be established near approximately 80% of rated load, rated voltage, and normal temperature. The actual trigger time and baseline operating conditions can be adjusted according to the equipment operation plan.

[0116] In this embodiment of the invention, the extraction of ripple morphology features, the acquisition of ripple morphology features after frequency band reweighting, and the determination of aging detection results are all performed by a pre-trained aging detection model.

[0117] The training steps for the aging detection model may include: acquiring a training sample set, which includes measured data of the power supply equipment in a healthy state, controlled aging test data, and simulation data generated based on the main circuit state equation; among which, the controlled aging test data includes: bus capacitor aging data obtained by replacing capacitor groups with different measured effective capacitance or measured equivalent series resistance, rectifier diode aging data obtained by replacing rectifier diodes with different measured forward voltage drops or measured dynamic resistances, and mixed aging data obtained by simultaneously setting changes in bus capacitor parameters and single rectifier branch parameters; and based on the measured capacitance, equivalent series resistance, diode forward voltage drop, and equivalent resistance of the conduction path... The process involves: determining the aging type and aging level labels for training samples in the training sample set; obtaining the bus voltage ripple based on the DC bus voltage of the training samples, and performing time-frequency transformation on the bus voltage ripple corresponding to the training samples to obtain the ripple time-frequency map of the training samples; inputting the ripple time-frequency map of the training samples into the initial aging detection model to obtain the aging detection result output by the initial aging detection model, and determining the loss value based on the difference between the aging detection result corresponding to the training samples and the aging type and aging level labels corresponding to the training samples; adjusting the parameters of the initial aging detection model based on the loss value and performing iterative training until the preset iteration termination condition is met to obtain the aging detection model.

[0118] In this embodiment of the invention, the network structure and input construction method of the initial aging detection model are the same as those of the aging detection model. During the model training phase, training samples with aging type and aging level labels are used to form a ripple time-frequency map, and the aging detection model parameters are adjusted accordingly. During the actual inference phase, the model parameters are not adjusted; instead, the ripple time-frequency map formed by the current detection window of the power supply device under test is input into the pre-trained aging detection model. The pre-trained aging detection model extracts ripple morphological features, determines frequency attention values ​​and physical frequency band prior values ​​according to the processing procedure described in the above embodiment, and performs frequency band reweighting, ultimately outputting the aging detection result.

[0119] In this embodiment of the invention, the preset iteration termination condition may include at least one of reaching a preset number of iterations or the loss value being less than a preset loss value.

[0120] To ensure that the aging level training reflects the order of normal, warning, and alarm states, three losses can be set: type classification loss, level classification loss, and level consistency loss. The total loss can be expressed as: ; The expected value of the aging level probability can be expressed as: ; The loss of level consistency can be expressed as: ; In the formula, Indicates the total loss. This represents the type classification cross-entropy loss calculated based on the aging type label; This represents the cross-entropy loss for grading based on aging level labels. and These represent the weights of the classification loss and the consistency loss, respectively. This represents the expected rank calculated from the rank probability. Indicates the first The probability of each aging level. Choose 0, 1, or 2; This indicates a loss of consistency in the level. Represents the smoothing L1 loss function. This indicates the true aging level. The loss of consistency in aging levels utilizes the ordered relationship between normal, warning, and alarm states to reduce the likelihood of healthy samples being directly identified as alarms.

[0121] The power supply equipment aging detection method provided in this embodiment of the invention may further include: forming shared features from the frequency band reweighted ripple morphology features; inputting the shared features into an aging type head and an aging level head respectively; the aging type head outputs the type probability of different aging types, and the aging level head outputs the level probability of different aging levels; when training the aging level head, setting a level consistency loss according to the order relationship between each aging level. Thus, while maintaining the aging type judgment, the level output can reflect the maintenance severity relationship between normal, warning, and alarm states.

[0122] The training sample set can consist of measured equipment health status data, controlled aging test data, and auxiliary simulation data generated based on the main circuit state equation. Among them, the controlled aging test data is used to construct known state data corresponding to bus capacitor aging, rectifier diode aging, and mixed aging.

[0123] Bus capacitor aging tests can be primarily achieved by replacing capacitor banks with different effective capacitances and ESRs, as confirmed by measurements using a bridge or inductance, capacitance, and resistance meter (LCR meter). Labels should prioritize measured parameters rather than solely using nominal device values. Alternatively, while maintaining a constant capacitance, calibrated small-resistance components can be connected in series in the capacitor branches to create an equivalent increase in ESR, thus studying the impact of ESR variations on ripple characteristics.

[0124] Rectifier diode aging tests can be implemented by replacing diodes or rectifier modules of the same type with different measured forward voltage drops and dynamic resistances. Alternatively, a calibrated low-resistance component can be added to a single conducting branch to controllably increase the resistance of the conduction path. Mixed aging samples can be obtained by simultaneously changing the bus capacitance parameter and the parameters of a single rectifier branch. During the test, the actual capacitance, ESR, diode forward voltage drop, and equivalent resistance of the conduction path are recorded, and these measured parameters are used as the basis for training labels and simulation parameters.

[0125] In this embodiment of the invention, the training sample set may also include data augmentation samples obtained by augmenting measured data, controlled aging test data and simulation data. Data augmentation may include at least one of time shift, sensor noise and disturbances of input voltage and load, but does not include up-down flipping, frequency axis flipping and color transformation.

[0126] Specifically, data augmentation avoids vertical flipping, frequency axis flipping, and color transformation to prevent the generation of training data inconsistent with the physical meaning of actual electrical signals. Training, validation, and test sets can be divided according to device number or acquisition date to prevent adjacent windows from the same continuous waveform from simultaneously entering the training and test sets.

[0127] The aging level indicates the severity of maintenance and does not require bus capacitors and rectifier diodes to share the same physical threshold. For bus capacitor aging, the aging level can be determined based on the degree of decrease in measured effective capacitance relative to a healthy baseline and the degree of increase in ESR relative to a healthy value. For rectifier diode aging, the aging level can be determined based on measured forward voltage drop, dynamic resistance, conduction path imbalance, and device test results. A normal rating indicates that device parameters are within the acceptable healthy range; a warning rating indicates that parameters have shown a stable deviation but have not yet reached the maintenance or replacement criteria; and an alarm rating indicates that parameters have reached or exceeded the preset maintenance, replacement, or severe aging criteria.

[0128] In other implementations, when it is necessary to identify abnormalities such as loose bus connections, poor rectifier contacts, or other abnormalities that can cause changes in bus voltage ripple, training samples and type labels of the corresponding aging or abnormality types can be added, and the aging detection model can be further trained.

[0129] In a simulation embodiment used to create a comparison between light and heavy aging, light aging of the bus capacitor can be defined as an effective capacitance decreasing to 80% of the healthy value and an ESR increasing to 150% of the healthy value; heavy aging of the bus capacitor can be defined as an effective capacitance decreasing to 60% of the healthy value and an ESR increasing to 300% of the healthy value. These values ​​are used as data construction points in this embodiment and are not intended to limit all equipment models to using the same thresholds. For actual equipment, the grade boundaries should be determined primarily based on the manufacturer's maintenance standards, measured health benchmarks, and on-site inspection standards.

[0130] To generate interpretable demonstration waveforms and auxiliary training samples, the relationship between the maximum and minimum values ​​of the three-phase instantaneous voltages and the parameters of the two-phase conduction loops can be used. and Simulations were performed using a piecewise linear model of the diodes and a bus capacitor model including ESR. The rectifier bridge conduction period can be represented as: ; ; ; In the formula, This represents the charging current from the rectifier bridge to the bus capacitor. This represents the equivalent threshold voltage drop of the currently conducting diode pair. This represents the dynamic resistance corresponding to the current conduction path. This represents the internal voltage of an ideal capacitor. Indicates the DC bus voltage. This represents the equivalent series resistance of the DC bus capacitor bank. When the rectifier bridge is reverse biased and... When it decays to zero, it can be made And cause the bus capacitor to discharge according to the following formula: .

[0131] The internal integration frequency in the simulation can be set to 200kHz. After reaching steady state, a fourth-order 2kHz low-pass filter can be used for filtering, and the frequency can be downsampled to 10kHz to match the aforementioned sampling parameters.

[0132] In one simulation example, under healthy conditions, the average bus voltage is 521.40V, the peak-to-peak ripple is 2.13V, the sixth harmonic amplitude is 0.906Vpk (Vpk represents the peak voltage unit), and the third harmonic amplitude is 0.0005Vpk. When the bus capacitors are slightly aged, the average bus voltage is 521.48V, the peak-to-peak ripple is 2.73V, the sixth harmonic amplitude is 1.157Vpk, and the third harmonic amplitude is 0.0003Vpk. When the bus capacitors are severely aged, the average bus voltage is 521.54V, the peak-to-peak ripple is 3.98V, the sixth harmonic amplitude is 1.662Vpk, and the third harmonic amplitude is 0.0003Vpk.

[0133] Under severe aging of the rectifier diode, the average bus voltage is 521.09V, the peak-to-peak ripple value is 2.72V, the sixth harmonic amplitude is 0.905Vpk, and the third harmonic amplitude is 0.0019Vpk. Under mixed aging, the average bus voltage is 521.25V, the peak-to-peak ripple value is 3.88V, the sixth harmonic amplitude is 1.413Vpk, and the third harmonic amplitude is 0.0014Vpk. The above simulation results are used to illustrate the correspondence between device parameter changes and ripple morphology. Actual network training should prioritize the use of measured data.

[0134] comprehensive Figures 4 to 6 As can be seen from the simulation data above, the decrease in the effective capacitance of the bus capacitor mainly leads to a stable increase in the peak-to-peak value of the ripple and the sixth harmonic main component; the changes in the parameters of a single rectifier diode are mainly manifested as pulse unevenness, local gaps, and low-order asymmetric textures, while the third harmonic amplitude does not necessarily increase significantly in the milder stages. Therefore, relying solely on a single 150Hz frequency threshold is insufficient to fully characterize the aging state, while the ripple time-frequency plot can retain a more complete time-domain morphology and multi-band texture.

[0135] This invention provides a training method for an aging detection model. The aging detection model is used for aging detection of power supply equipment. The power supply equipment includes a three-phase six-pulse rectifier and a DC bus capacitor bank connected to the DC side of the three-phase six-pulse rectifier. The training method may include: acquiring a training sample set, where the training samples include samples of the power supply equipment in a healthy state and samples of the power supply equipment in an aging state; obtaining the bus voltage ripple based on the DC bus voltage of the training samples, and performing a time-frequency transformation on the bus voltage ripple of the training samples to obtain a ripple time-frequency diagram corresponding to the training samples; inputting the ripple time-frequency diagram corresponding to the training samples into an initial aging detection model to extract the ripple morphology features of the ripple time-frequency diagram corresponding to the training samples; and performing a time-frequency transformation based on the ripple time-frequency diagram corresponding to the training samples. Frequency statistical features corresponding to multiple frequency locations are extracted, and frequency attention values ​​corresponding to multiple frequency locations are determined. Physical frequency band priors are established centered on the third, sixth, and ninth harmonics of the current power grid base frequency. The frequency attention values ​​and physical frequency band priors are fused, and the ripple morphology features are reweighted based on the fusion result to obtain the reweighted ripple morphology features corresponding to the training samples. The aging detection result corresponding to the training samples is determined based on the reweighted ripple morphology features. The loss value is determined based on the difference between the aging detection result corresponding to the training samples and the labeled data of the training samples. The parameters of the initial aging detection model are adjusted based on the loss value, and iterative training is performed until the preset iteration termination condition is met, thus obtaining the aging detection model.

[0136] The specific implementation methods of each step in the above training method are the same as the corresponding processing in the embodiment of the power supply equipment aging test method, and will not be repeated here.

[0137] The foregoing detailed various embodiments of the power supply equipment aging test method and the training method of the aging test model. Based on this, the present invention also discloses a power supply equipment aging test device, electronic device and computer-readable storage medium corresponding to the above method.

[0138] This invention also provides a power supply equipment aging test device. The power supply equipment includes a three-phase six-pulse rectifier and a DC bus capacitor bank connected to the DC side of the three-phase six-pulse rectifier. The power supply equipment aging test device may include: The acquisition module is used to collect the DC bus voltage of the power supply equipment, obtain the bus voltage ripple based on the DC bus voltage, and perform time-frequency transformation on the bus voltage ripple to obtain the ripple time-frequency diagram. The morphological feature extraction module is used to extract the morphological features of ripple in the ripple time-frequency plot. The frequency band reweighting module is used to determine the frequency attention value corresponding to multiple frequency positions based on the frequency statistical characteristics of multiple frequency positions extracted from the ripple time-frequency map; establish physical frequency band prior values ​​with the third, sixth, and ninth harmonics of the current power grid base frequency as the center; fuse the frequency attention value with the physical frequency band prior value, and reweight the ripple morphology features based on the fusion result to obtain the frequency band reweighted ripple morphology features; The aging detection module is used to determine the aging detection results of the power supply equipment based on the ripple morphology characteristics after frequency band reweighting.

[0139] In one specific implementation, the morphological feature extraction module is further used to perform channel mapping on the input features extracted from the ripple time-frequency map, and input the channel-mapped features into the time-direction ridge branch, the frequency-direction broadening branch, the local texture branch, and the bidirectional gradient branch, respectively; extract time continuity features along the time direction through the time-direction ridge branch, extract frequency broadening features along the frequency direction through the frequency-direction broadening branch, extract local defect features in the two-dimensional neighborhood of the time and frequency directions through the local texture branch, and extract edge abrupt change features based on the local gradients of the time and frequency directions through the bidirectional gradient branch; fuse the outputs of the four branches, and adjust the fused incremental features according to the channel gating results to obtain the ripple morphological features.

[0140] In one specific implementation, the frequency band reweighting module is further used to average the input features extracted from the ripple time-frequency map along the channel direction and the time direction to obtain frequency statistical features corresponding to multiple frequency positions, and to perform one-dimensional convolution processing on the frequency statistical features to obtain frequency attention values; to determine the ridge center frequency corresponding to the maximum time-frequency energy in the ripple time-frequency map within a preset sixth harmonic search range, and to determine the current power grid base frequency based on one-sixth of the ridge center frequency; to determine the center frequencies corresponding to the third, sixth, and ninth harmonics based on the current power grid base frequency, to set learnable frequency band amplitude coefficients and bandwidths for the third, sixth, and ninth harmonics respectively, and to determine the physical frequency band prior values ​​corresponding to multiple frequency positions based on the actual frequency, center frequency, frequency band amplitude coefficient, and bandwidth corresponding to each frequency position; and to fuse the frequency attention values ​​and physical frequency band prior values ​​using non-negative learnable fusion weights.

[0141] In one specific implementation, the sample acquisition module is further configured to perform continuous wavelet transform on the bus voltage ripple to obtain wavelet coefficients corresponding to multiple frequency positions and multiple time positions, and to perform logarithmic compression on the amplitude of the wavelet coefficients; to normalize the logarithmically compressed amplitude according to a fixed calibration scale determined when the power supply equipment is in a healthy state; and to form an amplitude channel, a time gradient channel, and a frequency gradient channel according to the normalized time-frequency amplitude, the gradient of the time-frequency amplitude along the time direction, and the gradient of the time-frequency amplitude along the frequency direction, respectively, to obtain a ripple time-frequency diagram.

[0142] In one specific implementation, the aging detection module is further configured to establish multiple health condition references according to different combinations of bus voltage, load current, and temperature when the power supply equipment is in a healthy state, and save the reference characteristics under the healthy state for each health condition reference; determine the target healthy condition reference from the multiple healthy condition references based on the current bus voltage, load current, and temperature, and compare the corresponding characteristics of the current detection cycle with the reference characteristics corresponding to the target healthy condition reference; obtain the network output of the current detection cycle based on the ripple morphology characteristics after frequency band reweighting, and determine the aging detection result by combining the comparison results and the network output of multiple consecutive detection cycles.

[0143] This invention also provides a training device for an aging detection model. The aging detection model is used for aging detection of power supply equipment. The power supply equipment includes a three-phase six-pulse rectifier and a DC bus capacitor bank connected to the DC side of the three-phase six-pulse rectifier. The training device for the aging detection model includes: The training sample acquisition module is used to acquire the training sample set; The training sample transformation module is used to obtain the bus voltage ripple based on the DC bus voltage of the training sample and perform time-frequency transformation to obtain the ripple time-frequency diagram corresponding to the training sample. The model processing module is used to input the ripple time-frequency map corresponding to the training sample into the initial aging detection model and obtain the aging detection result corresponding to the training sample. The parameter adjustment module is used to determine the loss value based on the difference between the aging detection result corresponding to the training sample and the labeled data of the training sample, and adjust the parameters of the initial aging detection model according to the loss value and perform iterative training until the preset iteration termination condition is met to obtain the aging detection model.

[0144] It should be noted that in the various embodiments of the power supply equipment aging detection device provided in this invention, the division of units is only a logical functional division, and other division methods can be used. The connection between different units can be electrical, mechanical, or other connection methods. Separate units can be located in the same physical location or distributed across multiple network nodes. Each unit can be implemented in hardware or as a software functional unit. That is, according to actual needs, some or all of the units provided in this invention can be selected and corresponding connection or integration methods can be used to achieve the purpose of the solution in this invention.

[0145] Since the embodiments of the apparatus and the embodiments of the method correspond to each other, please refer to the description of the embodiments of the method for the embodiments of the apparatus, which will not be repeated here.

[0146] This invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above-described embodiments of the power supply device aging detection method or the above-described embodiments of the aging detection model training method.

[0147] This invention provides a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps of any of the above-described embodiments of the power supply device aging detection method or the above-described embodiments of the aging detection model training method when it is run.

[0148] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0149] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0150] The aging detection method for power supply equipment, the training method for the aging detection model, and the equipment provided by this invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only intended to help understand the method and core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of this invention.

Claims

1. A method for aging detection of power supply equipment, characterized in that, The power supply equipment includes a three-phase six-pulse rectifier and a DC bus capacitor bank connected to the DC side of the three-phase six-pulse rectifier. The aging test method for the power supply equipment includes: The DC bus voltage of the power supply equipment is collected, the bus voltage ripple is obtained based on the DC bus voltage, and the bus voltage ripple is transformed by time and frequency to obtain the ripple time and frequency diagram. Extract the ripple morphology features from the ripple time-frequency plot; Based on the frequency statistical features corresponding to multiple frequency positions extracted from the ripple time-frequency diagram, the frequency attention values ​​corresponding to the multiple frequency positions are determined; physical frequency band prior values ​​are established with the third, sixth, and ninth harmonics of the current power grid base frequency as the center; the frequency attention values ​​are fused with the physical frequency band prior values, and the ripple morphology features are reweighted based on the fusion result to obtain the reweighted ripple morphology features; The aging test results of the power supply equipment are determined based on the ripple morphology characteristics after frequency band reweighting.

2. The power supply equipment aging test method according to claim 1, characterized in that, The ripple morphological characteristics include temporal continuity characteristics, frequency broadening characteristics, local defect characteristics, and edge abrupt change characteristics; The time continuity feature is used to characterize the continuity, periodicity, and duration of ripple features at the same frequency position over time. The frequency broadening feature is used to characterize the energy diffusion relationship between adjacent frequency positions; The local defect features are used to characterize local abnormal textures within a preset time range and a preset frequency range; The edge mutation feature is used to characterize local changes in the time and frequency directions.

3. The power supply equipment aging test method according to claim 2, characterized in that, Extracting the ripple morphology features from the ripple time-frequency plot includes: The input features extracted from the ripple time-frequency map are channel-mapped, and the channel-mapped features are respectively input into the time-directed ridge branch, the frequency-directed broadening branch, the local texture branch, and the bidirectional gradient branch. The temporal continuity feature is extracted along the temporal direction by the temporal ridge branch, the frequency broadening feature is extracted along the frequency direction by the frequency broadening branch, the local defect feature is extracted in the two-dimensional neighborhood of the temporal and frequency directions by the local texture branch, and the edge abruptness feature is extracted according to the local gradient of the temporal and frequency directions by the bidirectional gradient branch. The outputs of the temporal ridge branch, the frequency broadening branch, the local texture branch, and the bidirectional gradient branch are fused, and the incremental features obtained by fusion are adjusted according to the channel gating results to obtain the ripple morphology features.

4. The power supply equipment aging test method according to claim 1, characterized in that, The determination of the frequency attention value, the establishment of the physical frequency band prior value, and the fusion of the frequency attention value and the physical frequency band prior value include: The input features extracted from the ripple time-frequency map are averaged along the channel direction and the time direction to obtain the frequency statistical features corresponding to the multiple frequency positions. The frequency statistical features are then subjected to one-dimensional convolution processing to obtain the frequency attention value. Within a preset sixth harmonic search range, the ridge center frequency corresponding to the maximum time-frequency energy in the ripple time-frequency graph is determined, and the current power grid base frequency is determined based on one-sixth of the ridge center frequency. The center frequencies corresponding to the third, sixth, and ninth harmonics are determined based on the current power grid base frequency. Learnable frequency band amplitude coefficients and bandwidths are set for the third, sixth, and ninth harmonics respectively. Based on the actual frequency corresponding to each frequency position, the center frequency, the frequency band amplitude coefficient, and the bandwidth, the physical frequency band prior values ​​corresponding to the multiple frequency positions are determined. The frequency attention value and the physical frequency band prior value are fused using non-negative learnable fusion weights.

5. The power supply equipment aging test method according to claim 1, characterized in that, Perform time-frequency transformation on the bus voltage ripple to obtain the ripple time-frequency diagram, including: A continuous wavelet transform is performed on the bus voltage ripple to obtain wavelet coefficients corresponding to multiple frequency positions and multiple time positions, and the amplitude of the wavelet coefficients is logarithmically compressed. The amplitude after logarithmic compression is normalized based on the fixed calibration scale determined when the power supply equipment is in a healthy state. Based on the normalized time-frequency amplitude, the gradient of the time-frequency amplitude along the time direction, and the gradient of the time-frequency amplitude along the frequency direction, an amplitude channel, a time gradient channel, and a frequency gradient channel are formed respectively to obtain the ripple time-frequency diagram.

6. The power supply equipment aging test method according to claim 1, characterized in that, The aging test results of the power supply equipment are determined based on the ripple morphology characteristics after frequency band reweighting, including: When the power supply equipment is in a healthy state, multiple healthy operating condition references are established according to different combinations of bus voltage, load current and temperature, and the reference characteristics under the healthy state are saved for each healthy operating condition reference. The reference characteristics include ripple amplitude, third-degree frequency band energy, sixth-degree frequency band energy and ninth-degree frequency band energy. Based on the current bus voltage, load current and temperature, a target health condition benchmark is determined from the plurality of health condition benchmarks, and the ripple amplitude, third-octave band energy, sixth-octave band energy and ninth-octave band energy of the current detection cycle are compared with the benchmark characteristics corresponding to the target health condition benchmark. The network output for the current detection cycle is obtained based on the ripple morphology characteristics after frequency band reweighting. The aging detection result is determined by combining the comparison results and the network outputs of multiple consecutive detection cycles. Specifically, when the ripple anomaly occurs during changes in bus voltage, load current, or temperature and disappears after the corresponding operating condition is restored, the ripple anomaly is identified as an operating condition disturbance. When the ripple anomaly persists or shows a monotonically increasing trend in multiple detection cycles under the operating condition corresponding to the target healthy operating condition benchmark, the ripple anomaly is used as aging evidence to determine the aging detection result.

7. The power supply equipment aging test method according to claim 1, characterized in that, The extraction of the ripple morphology features, the acquisition of the frequency band reweighted ripple morphology features, and the determination of the aging detection results are all performed through a pre-trained aging detection model. The training steps of the aging detection model include: A training sample set is obtained, which includes measured data of the power supply equipment in a healthy state, controllable aging test data, and simulation data generated based on the main circuit state equation; wherein, the controllable aging test data includes: bus capacitor aging data obtained by replacing capacitor groups with different measured effective capacitance or measured equivalent series resistance, rectifier diode aging data obtained by replacing rectifier diodes with different measured forward voltage drop or measured dynamic resistance, and mixed aging data obtained by simultaneously setting changes in bus capacitor parameters and changes in single rectifier branch parameters; Based on the measured capacitance, equivalent series resistance, diode forward voltage drop, and conduction path equivalent resistance, the aging type label and aging level label of the training samples in the training sample set are determined. The bus voltage ripple is obtained based on the DC bus voltage of the training sample, and the bus voltage ripple corresponding to the training sample is transformed by time and frequency to obtain the ripple time and frequency diagram corresponding to the training sample. Input the ripple time-frequency map corresponding to the training sample into the initial aging detection model to obtain the aging detection result output by the initial aging detection model, and determine the loss value based on the difference between the aging detection result corresponding to the training sample and the aging type label and aging level label corresponding to the training sample. The parameters of the initial aging detection model are adjusted based on the loss value and iteratively trained until the preset iteration termination condition is met, thus obtaining the aging detection model.

8. The power supply equipment aging test method according to claim 7, characterized in that, The training sample set also includes data augmentation samples obtained by augmenting the measured data, the controlled aging test data and the simulation data. The data augmentation includes at least one of time shift, sensor noise and disturbances of input voltage and load, but does not include vertical flipping, frequency axis flipping and color transformation.

9. A training method for an aging detection model, characterized in that, The aging detection model is used for aging detection of power supply equipment, which includes a three-phase six-pulse rectifier and a DC bus capacitor bank connected to the DC side of the three-phase six-pulse rectifier. The training method includes: Obtain a training sample set, wherein the training samples in the training sample set include samples when the power supply device is in a healthy state and samples when it is in an aging state; The bus voltage ripple is obtained based on the DC bus voltage of the training sample, and the bus voltage ripple of the training sample is transformed by time and frequency to obtain the ripple time and frequency diagram corresponding to the training sample. The ripple time-frequency map corresponding to the training sample is input into the initial aging detection model to extract the ripple morphology features of the ripple time-frequency map corresponding to the training sample. Based on the frequency statistical features corresponding to multiple frequency positions extracted from the ripple time-frequency map corresponding to the training sample, frequency attention values ​​corresponding to the multiple frequency positions are determined. Physical frequency band prior values ​​are established centered on the third, sixth, and ninth harmonics of the current power grid base frequency. The frequency attention values ​​are fused with the physical frequency band prior values, and the ripple morphology features are reweighted based on the fusion result to obtain the reweighted ripple morphology features corresponding to the training sample. The aging detection result corresponding to the training sample is determined based on the reweighted ripple morphology features corresponding to the training sample. The loss value is determined based on the difference between the aging detection results corresponding to the training samples and the labeled data of the training samples; The parameters of the initial aging detection model are adjusted based on the loss value and iteratively trained until the preset iteration termination condition is met, thus obtaining the aging detection model.

10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the power supply device aging detection method as described in any one of claims 1 to 8 or the aging detection model training method as described in claim 9.

Images