Power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion

Abstract

The application relates to the technical field of intelligent diagnosis of power supply systems and optical communication cross technology, and discloses a power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion, which comprises the following steps: short-time coarse sampling of a power supply ripple signal, dynamic adjustment of sampling parameters and window function types according to the coarse sampling result; extraction of frequency domain features, time domain features and time-frequency domain features to form a multi-dimensional feature vector; dynamic updating of a fault threshold value through Bayesian estimation based on historical normal feature samples in a fault feature sample library; adoption of weighted D-S evidence fusion of the fault confidence of the multi-dimensional feature vector based on the fault threshold value, and determination of a system state based on the fault confidence to output a diagnosis result; and hierarchical fault tracing when the system state is an abnormal state. The application realizes adaptive sampling and feature fusion, reduces the fault misjudgment rate, shortens the tracing time consumption, has low computing power consumption, and is suitable for real-time health diagnosis of a data center optical module power supply system.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostics of power supply systems and the cross-disciplinary field of optical communication, and in particular to a method for health diagnosis of power supply systems based on adaptive multi-dimensional ripple feature fusion. Background Technology

[0002] The communication links between switches and optical modules in data centers operate under high load for extended periods. Core components within the optical modules, such as lasers and DSP chips, are highly sensitive to power supply voltage ripple at the μV level. Even minor anomalies in the power supply link (such as increased impedance due to aging power lines, filter failure caused by dried-out filter capacitors, or drift in the switching power supply control chip) can distort the spectral characteristics of the voltage ripple, thereby interfering with signal modulation and data transmission in the optical modules and causing sudden packet loss, bit error rate jumps, and other faults. The root causes of these faults are often concealed, and traditional monitoring methods struggle to capture the dynamic spectral changes in the ripple, resulting in long troubleshooting cycles and high maintenance costs.

[0003] Currently, power supply ripple signal monitoring solutions in the industry mainly suffer from three technical limitations: First, time-domain peak monitoring based on a fixed sampling rate only judges faults by whether the ripple peak exceeds a threshold, but cannot distinguish the type of fault; second, frequency-domain analysis schemes using fixed-parameter FFT have fixed sampling rates and FFT points, making it difficult to adapt to voltage ripple characteristic changes in different power supply environments, and only extracting a single power spectral density feature, resulting in weak anti-interference capabilities; third, centralized monitoring schemes relying on dedicated hardware are costly, have low deployment flexibility, and cannot achieve distributed node-level monitoring.

[0004] Therefore, a power supply system health diagnosis scheme based on adaptive multi-dimensional ripple feature fusion is needed to solve the problem of inaccurate diagnosis caused by inaccurate power supply ripple signal monitoring. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problem that abnormalities caused by power supply failures in optical modules are difficult to diagnose accurately and are easily confused with faults in the optical modules themselves. This invention provides a power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion, which achieves adaptive adaptation to different voltage ripple scenarios, multi-dimensional feature anti-interference, accurate fault tracing, and low computing power consumption.

[0006] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0007] A power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion is characterized by the following steps:

[0008] S1. Perform short-time coarse sampling on the power supply ripple signal, and dynamically adjust the sampling parameters and window function type of the power supply ripple signal based on the coarse sampling results;

[0009] S2. Extract frequency domain features, time domain features, and time-frequency domain features from the power supply ripple signal acquired after adjustment by S1 to form a multi-dimensional feature vector;

[0010] S3. Based on historical normal feature samples in the fault feature sample library, the fault threshold of the multi-dimensional feature vector is dynamically updated by Bayesian estimation.

[0011] S4. Based on the fault threshold, the fault confidence is obtained by weighted DS evidence fusion of multi-dimensional feature vectors, and the system status is determined based on the fault confidence to output the diagnostic result.

[0012] S5. When the system is in an abnormal state, perform graded fault tracing.

[0013] Furthermore, the dynamically adjusted sampling parameters in S1 include:

[0014] Adaptive sampling rate: ;

[0015] Fast Fourier Transform Points: ;

[0016] Where fs is the adaptive sampling rate; N is the number of points in the Fast Fourier Transform; f max,est γ is the estimated highest frequency of the ripple; T is the sampling redundancy coefficient; win For a single batch of sampling time window; This represents the frequency resolution; round indicates rounding to the nearest integer.

[0017] The window function type is based on temporal smoothness. The selection is as follows: where n is the index of the sampling point within the window function, with a value range of [0,1,⋯,N-1]; x(n) is the discrete-time signal sequence; max(x(n)) and min(x(n)) are the maximum and minimum values ​​of the short-time coarse-sampled data; std(x(n)) is the standard deviation of the short-time coarse-sampled data; when S≥8, the Blackman window is selected; when 4 < S < 8, the Hanning window is selected; and when S≤4, the rectangular window is selected.

[0018] Furthermore, in S2, the constructed multi-dimensional feature vector is A`=[PSD,THD,SE,CF,K,IF,WPEE], where PSD is the power spectral density, THD is the total harmonic distortion rate, SE is the spectral entropy, CF is the peak factor, K is the kurtosis, IF is the impulse factor, and WPEE is the wavelet packet energy entropy.

[0019] Furthermore, S3 specifically includes the following steps:

[0020] S31. Based on historical normal feature samples of multi-dimensional feature vectors in the fault feature sample library, the prior parameters of the multi-dimensional feature vectors are obtained by maximum likelihood estimation.

[0021] S32. After collecting every n' multi-dimensional feature vectors diagnosed as normal, calculate the mean of the latest n' multi-dimensional feature vectors. and variance Based on the mean and variance Update posterior mean and posterior variance ;

[0022] S33. Calculate the fault threshold Thrt based on the posterior mean and posterior variance.

[0023] Furthermore, in step S32, the update formula for the posterior mean is:

[0024] ;

[0025] The update formula for the posterior variance is:

[0026] ;

[0027] in, The prior mean; For prior variance; Let be the mean of the latest n' multi-dimensional feature vectors; Let be the variance of the latest n' multi-dimensional feature vectors; τ0 is the confidence weight of the mean; represents the confidence weight of the variance.

[0028] Furthermore, S4 specifically includes the following steps:

[0029] S41. Obtain the weights of multi-dimensional feature vectors on fault types;

[0030] S42. Based on the fault threshold, calculate the fault confidence assignment function of the multi-dimensional feature vector;

[0031] S43. The fault confidence is obtained based on the confidence assignment function of the multi-dimensional feature vector fusion of weighted DS evidence.

[0032] S44. Determine the system status based on the fault confidence level.

[0033] Furthermore, the graded fault tracing in S5 includes: identifying the fault type, quantifying the degree of aging, and pinpointing the fault location.

[0034] Furthermore, the quantified degree of aging is calculated using the aging degree index D:

[0035] ;

[0036] Where m(Fault) is the fault confidence level; 50%≤D<80% indicates mild aging of the power supply system, 80%≤D<95% indicates moderate aging of the power supply system, and 95%≤D indicates severe aging of the power supply system.

[0037] Furthermore, the fault location is locked based on the feature contribution C. i The final contribution obtained by normalization Sure:

[0038] ;

[0039] ;

[0040] Among them, A i Let C be the i-th feature in the multi-dimensional feature vector A', where i = 1, 2, ..., 7; i For feature A i Characteristic contribution; w i For feature A i The weight; Thr t,i For feature A i Fault threshold; C j The feature contribution of feature j; take The feature corresponding to the largest value is used for localization.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] (1) Significantly improved fault diagnosis accuracy: The combination of multi-dimensional feature fusion and Bayesian dynamic threshold reduces the fault misjudgment rate from 20% of the existing technology to below 0.3%, and the fault type identification accuracy is ≥99%;

[0043] (2) Realize graded fault tracing: For the first time, a three-level diagnosis of "fault type-aging degree-fault location" is realized, which solves the limitation of traditional solutions that can only determine "whether there is a fault" and reduces the fault tracing time from 4 hours to 30 seconds;

[0044] (3) Strong adaptive capability: The adaptive sampling rate and window function selection can cover the ripple frequency range of 1kHz-2MHz, adapt to the power supply environment of different brands and models of switches, and improve compatibility by 80%;

[0045] (4) Low computing power consumption: Feature dimensionality reduction and time-sharing scheduling mechanism, the entire process of a single diagnosis takes ≤300μs, only occupies 5% of the MCU's computing power resources, and does not affect the normal communication function of the optical module. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0048] 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0049] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0050] Example 1:

[0051] This invention is achieved through the following technical solutions, such as... Figure 1 As shown, this invention provides a power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion, the method comprising the following steps:

[0052] S1. Perform short-time coarse sampling on the power supply ripple signal, and dynamically adjust the sampling parameters and window function type of the power supply ripple signal based on the coarse sampling results.

[0053] The short-time coarse sampling uses a fixed sampling rate and number of sampling points for initial data acquisition to quickly estimate the peak frequency and waveform pulse characteristics of the power supply ripple signal. Specifically, the sampling rate f of the short-time coarse sampling... s,coarse =500kHz, number of sampling points N coarse =256.

[0054] The coarse sampling results include the coarse sampling time-domain sequence, the peak frequency of the spectrum, the estimated highest frequency, and the time-domain smoothness.

[0055] The sampling parameters include the adaptive sampling rate f.s And the number of points N in the Fast Fourier Transform.

[0056] The system initially determines the characteristics of the power supply ripple signal based on short-time coarse sampling results, and dynamically adjusts the sampling parameters, as shown in the following formula:

[0057] ;

[0058] ;

[0059] Among them, f s For adaptive sampling rate; f max,est This is the estimate of the highest frequency of the ripple. The highest frequency of the ripple can be estimated by short-time coarse sampling (sampling rate f). s,coarse =500kHz, number of sampling points N coarse The peak frequency of the spectrum (=256) is obtained by multiplying it by 1.2; γ is the sampling redundancy coefficient, which ensures coverage of the true highest frequency, γ=1.1; T win For a single batch sampling time window, T win =200μs; To meet the system's required frequency resolution, N represents the number of points in the Fast Fourier Transform; it should be noted that round means rounding to the nearest integer. In this embodiment, the rounding is taken as an integer power of 2, which can ensure efficient operation of the radix-2 Fast Fourier Transform (radix-2 FFT).

[0060] The system dynamically selects a window function based on the time-domain smoothness S of the power supply ripple signal. The formula for calculating the time-domain smoothness S of the power supply ripple signal is as follows:

[0061] ;

[0062] Where n is the index of the sampling point within the window function, and its value range is [0,1,⋯,N-1]; x(n) is the discrete-time signal sequence; max(x(n)) and min(x(n)) are the maximum and minimum values ​​of the short-time coarse-sampled data; std(x(n)) is the standard deviation of the short-time coarse-sampled data.

[0063] The window function type w(n) includes Blackman window, Hanning window, and rectangular window.

[0064] Specifically, when the time domain smoothness S is greater than or equal to 8, the ripple contains sharp pulses, and the Blackman window w(n) = 0.42 - 0.5cos(2πn / (N-1)) + 0.08cos(4πn / (N-1)) is selected; when the time domain smoothness S is greater than 4 and less than 8, the ripple stability is moderate, and the Hanning window w(n) = 0.5[1-cos(2πn / (N-1))] is selected; when the time domain smoothness is less than 4, the ripple is stable, and the rectangular window w(n) = 1 is selected.

[0065] The choice of window function can balance spectrum leakage suppression and computational efficiency, and achieve adaptive sampling scheduling.

[0066] It should be noted that after adjusting the sampling parameters and selecting the window function, a preprocessing process is required. The preprocessing process includes removing the DC component, windowing, and ripple noise suppression filtering, all of which are performed using existing technologies and will not be elaborated here.

[0067] S2. Extract frequency domain features, time domain features, and time-frequency domain features from the power supply ripple signal acquired after adjustment by S1 to form a multi-dimensional feature vector.

[0068] The frequency domain features include power spectral density, total harmonic distortion rate, and spectral entropy.

[0069] The system collects the adjusted power supply ripple signal. Perform an N-point Fast Fourier Transform, where This represents the DC component of the power supply ripple signal. Then, the k-th power spectral density PSD(k) is calculated:

[0070] ;

[0071] Where k is the frequency domain index, k=0,1,...,N / 2, and k only takes positive values ​​when calculating the power spectral density; X(k) is the result of the fast Fourier transform, and we have e is the natural constant, j is the imaginary unit, n = 0, 1, ..., N-1; R eq R is the equivalent resistance of the power supply circuit for the optical module. eq =50Ω.

[0072] Assuming N=8, then k=0,1,2,3,4. Since k only takes positive values ​​(i.e., k only takes 1,2,3,4) when calculating the power spectral density, the characteristic frequency set can be represented as F={f1,f2,f3,f4}. Calculate the corresponding power spectral density to construct the power spectral density vector PSD:

[0073] ;

[0074] Among them, PSD(f k ) represents the k-th characteristic frequency f k The corresponding power spectral density is k=1,2,3,4. For example, f1=100kHz, f2=250kHz, f3=500kHz, f4=1MHz, that is, the characteristic frequency set can be expressed as F={100kHz,250kHz,500kHz,1MHz}.

[0075] It should be noted that the system needs to first complete the power spectral density (PSD) value for all positive frequencies corresponding to k values, and then selectively extract the PSD results for some k values ​​according to different calculation objectives.

[0076] Total harmonic distortion (THD) reflects the modulation stability of a switching power supply. Assuming the system operates at a fundamental frequency f0 = 100kHz for the switching power supply, calculate the THD:

[0077] ;

[0078] Where h is the harmonic order (h=2,3,...,H, H is the total harmonic order, H=5); PSD(f0,h) represents the power spectral density of the h-th harmonic of the fundamental frequency f0 of the switching power supply.

[0079] Spectral entropy quantifies the degree of disorder in a spectral distribution. A higher spectral entropy indicates a more chaotic power supply ripple spectrum (e.g., due to capacitor drying out). Specifically, the system calculates the spectral entropy SE using the following formula:

[0080] ;

[0081] in, The power spectral density ratio at the k-th frequency point satisfies .

[0082] The time-domain features include peak factor, kurtosis, and impulse factor.

[0083] The peak factor reflects the pulse characteristics of the ripple. For example, the pulse characteristics are more pronounced when the power line ages. In this embodiment, the system calculates the peak factor CF using the following formula:

[0084] ;

[0085] Kurtosis is used to quantify the steepness of the ripple distribution; for example, kurtosis increases when the capacitor dries out. In this embodiment, the system calculates the kurtosis K using the following formula:

[0086] ;

[0087] The pulse factor is used to highlight the instantaneous pulse component in the ripple. The system calculates the pulse factor IF using the following formula:

[0088] ;

[0089] The time-frequency domain features include wavelet packet energy entropy.

[0090] The power supply ripple signal after adjustment was obtained using db4 wavelet packets. A 3-level decomposition yields 8 frequency band components. The system calculates the wavelet packet energy entropy (WPEE) of each frequency band component using the following formula:

[0091] ;

[0092] ;

[0093] Where m is the frequency band component index, m=1,2,...,8; E represents the signal of the m-th frequency band component; m This represents the energy of the m-th frequency band component; N represents the energy percentage of the m-th frequency band component; m This represents the signal length of the m-th frequency band component, which is the total number of sampling points within that frequency band.

[0094] Wavelet packet energy entropy (WPEE) comprehensively reflects the uniformity of energy distribution of the power supply ripple signal in the time and frequency domain. During a fault, energy concentrates in a specific frequency band, causing a decrease in the WPEE value.

[0095] Thus, the multi-dimensional feature vector A=[PSD,THD,SE,CF,K,IF,WPEE] is obtained. To eliminate the difference in dimensions, the mean of the features in the fault feature sample library is used. and standard deviation The multi-dimensional feature vector A is Z-score standardized to obtain the standardized multi-dimensional feature vector A'.

[0096] The fault feature sample library is used to store various reference feature data required for power supply system health diagnosis. This library contains multiple sets of labeled multidimensional feature samples, including multidimensional feature vectors.

[0097] S3. Based on historical normal feature samples in the fault feature sample library, the fault threshold of the multi-dimensional feature vector is dynamically updated through Bayesian estimation.

[0098] Furthermore, the specific steps for updating the fault threshold of the multi-dimensional feature vector in the system include:

[0099] S31. Based on historical normal feature samples (i.e., historical normal feature samples of PSD, THD, SE, CF, K, IF, WPEE) of multi-dimensional feature vectors in the fault feature sample library, the prior parameters of the multi-dimensional feature vectors are obtained by maximum likelihood estimation.

[0100] The prior parameters include the prior mean and the prior variance.

[0101] Specifically, the prior mean of the multi-dimensional feature vector The calculation formula is as follows:

[0102] ;

[0103] Where c is the index of the historical normal feature sample, c=1,2,...,M, and M is the number of historical normal feature samples; A a,c This is the c-th historical normal feature sample.

[0104] The prior variance of the multidimensional feature vector The calculation formula is as follows:

[0105] ;

[0106] S32. After collecting every n' multi-dimensional feature vectors diagnosed as normal, calculate the mean of the latest n' multi-dimensional feature vectors. and variance Based on the mean and variance Update posterior mean and posterior variance .

[0107] Specifically, the posterior mean and posterior variance The calculation formula is as follows:

[0108] ;

[0109] ;

[0110] Where t is the number of updates; Let be the mean of the latest n' multi-dimensional feature vectors; τ is the variance of the latest n' multi-dimensional feature vectors; τ0 is the confidence weight of the mean of the latest n' multi-dimensional feature vectors. In this embodiment, τ0=10. Let n' be the confidence weights of the variances of the latest n multidimensional feature vectors. n` represents the number of samples of the latest acquired multi-dimensional feature vector that have been diagnosed as normal. In this embodiment, n`=10.

[0111] S33. Calculate the fault threshold based on the posterior mean and posterior variance.

[0112] Specifically, the system uses the confidence interval method to calculate the fault threshold Thr using the following formula. t :

[0113] ;

[0114] Where kt is the dynamic confidence coefficient, calculated based on the false negative rate of fault samples. α is the allowable false negative rate, α = 0.001. As the inverse function of the standard normal distribution, kt≈3.09 is calculated, which ensures a false positive rate of ≤0.1% under normal operating conditions; is the posterior standard deviation.

[0115] S4. Based on the fault threshold, the fault confidence is obtained by weighted DS evidence fusion of multi-dimensional feature vectors, and the system status is determined based on the fault confidence to output the diagnostic result.

[0116] Fault confidence is a decision variable used to quantify the degree of confidence that a power supply system is in a fault state.

[0117] The specific steps for the system to fuse fault confidence based on multi-dimensional feature vectors and determine the system state include:

[0118] S41. Obtain the weights of the multi-dimensional feature vectors for the fault type.

[0119] Offline calculation of mutual information entropy (MI(A)) between multi-dimensional feature vectors and fault types i The mutual information entropy (Fault) is used to measure the ability of a feature to distinguish faults.

[0120] Furthermore, the formula for calculating the weights is as follows:

[0121] ;

[0122] Where i is the index of a feature in the multi-dimensional feature vector A, i = 1, 2, ..., 7, and A is used to represent the feature. i w represents the i-th feature in the multi-dimensional feature vector A; i Let be the weight of the i-th feature, and MI(A) i Fault) = H(A i )+H(Fault)-H(A i H(.) is the information entropy; A j Let w represent the j-th feature in the multi-dimensional feature vector A. i The higher the value, the greater the contribution of the i-th feature to fault diagnosis (e.g., the weight w1≈0.3 for the power spectral density vector PSD, and the weight w7≈0.2 for the wavelet packet energy entropy WPEE).

[0123] S42. Based on the fault threshold, calculate the fault confidence assignment function of the multi-dimensional feature vector.

[0124] The fault confidence assignment function is a basic confidence function that is independently assigned to each feature, used to quantify the degree of confidence of the corresponding feature in whether the power supply system is in a fault state or a normal state.

[0125] In this embodiment, the system can calculate the fault confidence assignment function using the following formula:

[0126] ;

[0127] ;

[0128] ;

[0129] Where b is the confidence decay coefficient, b=5; m i (Fault) represents the confidence level of classifying the i-th feature as a fault; m i (Normal) represents the confidence level of the i-th feature being judged as normal; It is an uncertain set. The confidence level for determining the i-th feature as uncertain.

[0130] S43. The fault confidence m(Fault) is obtained by using the confidence assignment function based on the weighted DS evidence fusion multi-dimensional feature vector.

[0131] Specifically, the formula for calculating the fault confidence level m (Fault) is as follows:

[0132] ;

[0133] S44. Determine the system status based on the fault confidence level m (Fault).

[0134] The system determines the system state based on the fault confidence m(Fault) calculated by S43, specifically including:

[0135] When m(Fault)≤0.5, the power supply system is determined to be in a normal state; when 0.5<m(Fault)<0.8, the power supply system is determined to be in a sub-healthy state; when 0.8≤m(Fault), the power supply system is determined to be in a fault state.

[0136] S5. When the system is in an abnormal state (including sub-health state and fault state), perform graded fault tracing.

[0137] The graded fault tracing includes identifying fault types, quantifying aging levels, and pinpointing fault locations.

[0138] When a faulty or sub-healthy state is determined, the current characteristic A is calculated. i Sim(A) cosine similarity with various fault types in the fault feature sample library i A k The fault type is identified using the following formula:

[0139] ;

[0140] Where A k Let T be the standard feature vector of the k-th type of fault (switching power supply failure k=1, capacitor drying k=2, power line aging k=3) in the fault feature sample library; T is the vector transpose; and Sim(A) is the cosine similarity. i A k The maximum value of k corresponds to the identified fault type.

[0141] The system calculates the aging index based on the fault confidence level m (Fault), using the following formula:

[0142] ;

[0143] Where D is the aging index of the power supply system, 50%≤D<80% indicates mild aging of the power supply system, 80%≤D<95% indicates moderate aging of the power supply system, and 95%≤D indicates severe aging of the power supply system (components need to be replaced immediately).

[0144] Furthermore, the system pinpoints the fault location based on feature contribution analysis, including:

[0145] The feature contribution C of the i-th feature i Defined as:

[0146] ;

[0147] Among them, w i For feature A i The weight; Thr t,i represents the fault threshold of the i-th feature; |.| represents the absolute value, used to indicate the degree of deviation from normal.

[0148] Furthermore, to ensure that the contributions of all features are comparable within [0,1], the feature contributions need to be normalized, including:

[0149] ;

[0150] in, C represents the final contribution of the i-th feature; j For feature A j The feature contribution. In this embodiment, the final contribution. The larger the value, the greater the contribution of the corresponding i-th feature to the current fault.

[0151] Furthermore, when locating the fault, the final contribution is taken. The feature corresponding to the maximum value is identified, and the fault location is determined based on the location of that feature.

[0152] Specifically, when the final contribution of feature PSD (100kHz) is the highest, the locking position is the switching power supply; when the final contributions of features THD and WPEE rank first and second, the locking position is the filter capacitor; when the contributions of features CF and IF are the highest, the locking position is the power cable.

[0153] In this embodiment, the present invention further includes a computing power optimization step:

[0154] The sampling and preprocessing tasks are set to high priority and are triggered by timer interrupts; the fast Fourier transform, feature extraction, Bayesian update, and weighted DS evidence fusion tasks are set to low priority and are executed during the optical communication frame intervals to achieve time-sharing scheduling.

[0155] Principal component analysis is used to reduce the dimensionality of features, thereby reducing the computational load of fusion.

[0156] The mean, standard deviation, and weight parameters of the multi-dimensional feature vectors are pre-stored in Flash, so that they can be updated once in a preset batch, avoiding repeated calculations.

[0157] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion, characterized in that, Includes the following steps: S1. Perform short-time coarse sampling on the power supply ripple signal, and dynamically adjust the sampling parameters and window function type of the power supply ripple signal based on the coarse sampling results; S2. Extract frequency domain features, time domain features, and time-frequency domain features from the power supply ripple signal acquired after adjustment by S1 to form a multi-dimensional feature vector; S3. Based on historical normal feature samples in the fault feature sample library, the fault threshold of the multi-dimensional feature vector is dynamically updated by Bayesian estimation. S4. Based on the fault threshold, the fault confidence is obtained by weighted DS evidence fusion of multi-dimensional feature vectors, and the system status is determined based on the fault confidence to output the diagnostic result. S5. When the system is in an abnormal state, perform graded fault tracing.

2. The power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion according to claim 1, characterized in that, The dynamically adjusted sampling parameters in S1 include: Adaptive sampling rate: ; Fast Fourier Transform Points: ; Among them, f s For adaptive sampling rate; N is the number of points in the Fast Fourier Transform; f max,est γ is the estimated highest frequency of the ripple; T is the sampling redundancy coefficient; win For a single batch of sampling time window; This represents the frequency resolution; round indicates rounding to the nearest integer. The window function type is based on temporal smoothness. The selection is as follows: where n is the index of the sampling point within the window function, with a value range of [0,1,⋯,N-1]; x(n) is the discrete-time signal sequence; max(x(n)) and min(x(n)) are the maximum and minimum values ​​of the short-time coarse-sampled data; std(x(n)) is the standard deviation of the short-time coarse-sampled data; when S≥8, the Blackman window is selected; when 4 < S < 8, the Hanning window is selected; and when S≤4, the rectangular window is selected.

3. The power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion according to claim 1, characterized in that, In S2, the multi-dimensional feature vector is A`=[PSD,THD,SE,CF,K,IF,WPEE], where PSD is the power spectral density, THD is the total harmonic distortion rate, SE is the spectral entropy, CF is the peak factor, K is the kurtosis, IF is the impulse factor, and WPEE is the wavelet packet energy entropy.

4. The power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion according to claim 1, characterized in that, S3 specifically includes the following steps: S31. Based on historical normal feature samples of multi-dimensional feature vectors in the fault feature sample library, the prior parameters of the multi-dimensional feature vectors are obtained by maximum likelihood estimation. S32. After collecting every n' multi-dimensional feature vectors diagnosed as normal, calculate the mean of the latest n' multi-dimensional feature vectors. and variance Based on the mean and variance Update posterior mean and posterior variance ; S33. Based on the posterior mean and posterior variance, calculate the fault threshold Thr. t .

5. The power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion according to claim 4, characterized in that, In S32, the update formula for the posterior mean is: ； The update formula for the posterior variance is: ； in, The prior mean; For prior variance; Let be the mean of the latest n' multi-dimensional feature vectors; Let be the variance of the latest n' multi-dimensional feature vectors; τ0 is the confidence weight of the mean; represents the confidence weight of the variance.

6. The power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion according to claim 1, characterized in that, S4 specifically includes the following steps: S41. Obtain the weights of multi-dimensional feature vectors on fault types; S42. Based on the fault threshold, calculate the fault confidence assignment function of the multi-dimensional feature vector; S43. The fault confidence is obtained based on the confidence assignment function of the multi-dimensional feature vector fusion of weighted DS evidence. S44. Determine the system status based on the fault confidence level.

7. The power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion according to claim 1, characterized in that, The graded fault tracing in S5 includes: identifying fault types, quantifying aging degree, and locating fault locations.

8. The power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion according to claim 7, characterized in that, The quantitative aging degree is calculated using the aging degree index D: ； Where m(Fault) is the fault confidence level; 50%≤D<80% indicates mild aging of the power supply system, 80%≤D<95% indicates moderate aging of the power supply system, and 95%≤D indicates severe aging of the power supply system.

9. The power supply system health diagnosis method based on adaptive multi-dimensional ripple feature fusion according to claim 7, characterized in that, The fault location is determined based on the feature contribution C. i The final contribution obtained by normalization Sure: ； ； Among them, A i Let C be the i-th feature in the multi-dimensional feature vector A', where i = 1, 2, ..., 7; i For feature A i Characteristic contribution; w i For feature A i The weight; Thr t,i For feature A i Fault threshold; C j The feature contribution of feature j; take The feature corresponding to the largest value is used for localization.

Images