Service life prediction and health management system, method and equipment for key device of switching power supply and medium

Abstract

The invention discloses a service life prediction and health management system, method, equipment and medium for key devices of a switching power supply, and relates to the field of electrical element management. The invention aims to solve the problems that a physical degradation mechanism is neglected in a method for predicting the aging of a device, a prediction model lacks the quantification of the uncertainty of a prediction result, and a sufficient risk assessment basis cannot be provided for operation and maintenance personnel; and after abnormity is detected, real-time feedback adjustment of driving parameters of key devices of the switching power supply is lacked. According to the invention, the key parameters of the key device of the switching power supply are collected, the health state and Bayesian uncertainty of the key device of the switching power supply are calculated according to the key parameters, and the adaptive Transform deep learning model is constructed to predict the service life of the key device of the switching power supply. And dynamically adjusting the PWM frequency and the gate driving voltage based on the health state and the Bayesian uncertainty.

Description

Technical Field

[0001] The present invention belongs to the field of electrical component management. Background Art

[0002] With the development of electronic systems towards high power, high efficiency, and high reliability, switching power supplies, as the core components in power management, play a crucial role in ensuring the performance and stability of the overall system. However, despite the continuous improvement of manufacturing processes, in practical applications, key devices (such as power semiconductors like MOSFETs and IGBTs) still face problems such as accelerated aging and performance degradation when subjected to multiple stresses such as high temperature, frequent switching, and external radiation. Traditional monitoring methods mainly rely on the real-time detection of single parameters such as current, voltage, or temperature, and usually take protective measures only after obvious abnormalities or failures occur in the device. This post-alarm method not only makes it difficult to timely reflect the true health status of the device but also unable to predict the remaining life of the device in advance, resulting in unplanned outages or premature failures of the device during actual operation, which has an adverse impact on the long-term reliability and maintenance cost of the system.

[0003] In addition, although some current improvement schemes attempt to introduce data-driven or machine learning methods to predict device aging, these methods often ignore the internal physical degradation mechanisms of key devices, and it is difficult to comprehensively characterize the degradation process of devices under complex working conditions with only data-driven models. At the same time, most prediction models lack the quantification of the uncertainty of prediction results and cannot provide sufficient risk assessment basis for operation and maintenance personnel. After detecting abnormalities, existing systems mostly stay at the alarm level and lack real-time feedback regulation of the driving parameters (such as PWM frequency and gate voltage) of key devices in switching power supplies, thus unable to effectively reduce the device load, improve the heat dissipation condition, and reduce switching losses. Summary of the Invention

[0004] The present invention aims to solve the problems that existing methods for predicting device aging ignore the internal physical degradation mechanisms of key devices, and it is difficult to comprehensively characterize the degradation process of devices under complex working conditions with only data-driven models; most prediction models lack the quantification of the uncertainty of prediction results and cannot provide sufficient risk assessment basis for operation and maintenance personnel; after detecting abnormalities, there is a lack of real-time feedback regulation of the driving parameters of key devices in switching power supplies, thus unable to effectively reduce the device load, improve the heat dissipation condition, and reduce switching losses. Now, a life prediction and health management system and method for key devices in switching power supplies are provided.

[0005] A life prediction and health management system for key devices in switching power supplies includes: an edge computing terminal and a cloud computing terminal;

[0006] The edge computing terminal is used to collect the key parameters of the key devices of the switching power supply and send the key parameters to the cloud computing terminal. The key parameters include: source-drain voltage drop V DS , source-drain current I D , source-drain resistance R DS(on) , threshold voltage V th , transient switching voltage transient switching current and switching loss power P sw ;

[0007] The cloud computing terminal is used to calculate the health status and Bayesian uncertainty of the key devices of the switching power supply according to the received key parameters, and to construct an adaptive Transformer deep learning model to predict the service life of the key devices of the switching power supply. It is also used to dynamically adjust the PWM frequency and gate drive voltage based on the health status and Bayesian uncertainty.

[0008] Further, calculating the health status of the key devices of the switching power supply includes:

[0009] Using the source-drain voltage drop V DS , source-drain current I D , source-drain resistance R DS(on) , threshold voltage V th to establish a conduction state degradation model of the key devices of the switching power supply;

[0010] Using the transient switching voltage transient switching current and switching loss power P sw to establish a switching process degradation model of the key devices of the switching power supply;

[0011] Combining the conduction state degradation model and the switching process degradation model through weight coefficients to establish a total degradation index model, and calculating the total degradation index;

[0012] Using the total degradation index to calculate the health status.

[0013] Further, the expression of the conduction state degradation model f cond (t) is:

[0014]

[0015] where V DS (t), I D (t), R DS(on) (t) and V th (t) are the source-drain voltage drop, source-drain current, source-drain resistance and threshold voltage at time t respectively, and V DS (0), ID (0), R DS(on) (0) and V th (0) are the source-drain voltage drop, source-drain current, source-drain resistance, and threshold voltage at the initial moment respectively. α1, α2, α3, and α4 are the parameter exponential factors corresponding to each low-frequency parameter, β1 is the low-frequency exponential factor, and E(t) is the low-frequency coupling factor;

[0016] The switching process degradation model f sw (t) of the key device of the switching power supply is expressed as:

[0017]

[0018] Among them, and P sw (t) are the transient switching voltage, transient switching current, and switching loss power at time t respectively, and P sw (0) are the transient switching voltage, transient switching current, and switching loss power at the initial moment respectively. γ1, γ2, and γ3 are the parameter exponential factors corresponding to each high-frequency parameter, β2 is the high-frequency exponential factor, and E(t) (sw) is the high-frequency coupling factor;

[0019] The expression of the total degradation index model is:

[0020] AI(t) = w cond ·f cond (t) + w sw ·f sw (t),

[0021] Among them, AI(t) is the total degradation index at time t, w cond is the weight of the degradation of the key device of the switching power supply in the on state, w sw is the weight of the degradation of the switching process of the key device of the switching power supply, and there is w cond + w sw = 1;

[0022] The expression of the health state SOH(t) is:

[0023]

[0024] Among them, AI crit is the set critical aging index. When AI(t) reaches AI crit , the key device of the switching power supply is considered to have reached the failure boundary.

[0025] Furthermore, the expression of the above low-frequency coupling factor E(t) is:

[0026]

[0027] The high-frequency coupling factor E(t) (sw) has the following expression:

[0028]

[0029] where T(t) is the operating temperature at time t, T ref is the reference temperature, T s is the temperature scale factor, Φ(t) represents the dose rate at time t, λ is the low-frequency irradiation sensitivity coefficient, and μ is the high-frequency irradiation sensitivity coefficient.

[0030] Furthermore, calculating the Bayesian uncertainty of the key devices of the switching power supply includes:

[0031] Obtaining the Bayesian uncertainty using the following formula:

[0032]

[0033] where U represents the Bayesian uncertainty value, y pred represents the life prediction value, y true represents the true life value, σ represents the standard deviation of the prediction distribution of the adaptive Transformer deep learning model, and κ and z are both uncertainty adjustment parameters.

[0034] Furthermore, the loss function of the above-mentioned adaptive Transformer deep learning model is the adaptive Bayesian incremental loss function, and its expression is:

[0035]

[0036] where i = 1, 2,..., N, j = 1, 2,..., M, N is the number of output samples of the adaptive Transformer deep learning model, M is the number of historical samples of the adaptive Transformer deep learning model, and M < N, λ1 is the mean square error weight, λ2 is the Bayesian regularization weight, λ3 is the data distillation weight, is the predicted value of the i-th output sample, is the actual value of the i-th output sample, ω i is the dynamic weighting coefficient of the i-th output sample, is the uncertainty estimation value of the i-th output sample, KL(·) represents Bayesian regularization, μ i represents the predicted average value of the i-th output sample, represents the prior distribution, is the output of the hidden layer of the model under the current iteration, is the output of the hidden layer of the model under the previous iteration.

[0037] Further, dynamically adjusting the PWM frequency and the gate drive voltage based on the health state and Bayesian uncertainty includes:

[0038] Coupling the health state error and the uncertainty to obtain a PWM frequency adjustment control law Δf PWM (t):

[0039]

[0040] where SOH ref is the target health state, H scale is the health state change scale factor, k0 is the base gain, k1 is the uncertainty adjustment coefficient, and U(t) is the Bayesian uncertainty value at time t;

[0041] Using the PWM frequency adjustment control law Δf PWM (t) to update the PWM frequency f PWM,new (t) at time t:

[0042] f PWM,new (t) = f PWM,orig ·[1 - Δf PWM (t)],

[0043] where f PWM,orig is the initial set frequency;

[0044] Coupling the health state error and the uncertainty to obtain a gate drive voltage adjustment control law ΔV GS (t):

[0045]

[0046] where δ0 is the base voltage adjustment coefficient and δ1 is the uncertainty sensitivity coefficient;

[0047] Using the gate drive voltage adjustment control law ΔV GS (t) to update the gate drive voltage V GS,new (t) at time t:

[0048] V GS,new (t) = V GS,orig - ΔV GS (t),

[0049] where V GS,orig is the initial drive voltage.

[0050] A method for predicting the life and health management of key devices in a switching power supply, including: collecting key parameters of key devices in a switching power supply, where the key parameters include: source-drain voltage drop V DS and source-drain current ID 、 Source-drain resistance R DS(on) 、 Threshold voltage V th 、 Transient switching voltage Transient switching current and switching loss power P sw ;

[0051] Calculate the health status and Bayesian uncertainty of the key devices of the switching power supply, and construct an adaptive Transformer deep learning model to predict the service life of the key devices of the switching power supply;

[0052] Dynamically adjust the PWM frequency and gate drive voltage based on the health status and Bayesian uncertainty.

[0053] A life prediction and health management device for the key devices of a switching power supply. The life prediction and health management device for the key devices of a switching power supply includes a processor and a memory. At least one instruction is stored in the memory, and the at least one instruction is loaded and executed by the processor to implement the life prediction and health management method for the key devices of the switching power supply as described above.

[0054] A computer storage medium stores at least one instruction, and the at least one instruction is loaded and executed by a processor to implement the life prediction and health management method for the key devices of the switching power supply as described above.

[0055] The beneficial effects of the present invention are as follows:

[0056] 1. The service life of the key devices is extended by 40%-80%.

[0057] 2. By intelligently controlling the switching frequency and adjusting the gate voltage, the switching loss is reduced by about 30%-50%, the heat accumulation is reduced by 20%-40%, the temperature fluctuation of the device is reduced by 15%-30%, and the energy efficiency and stability of the overall system are improved.

[0058] 3. The physics-based adaptive deep learning model in the present invention realizes real-time training and updating of the model, improves the model training efficiency by 25%, the model accuracy is above 90%, and the response time of the feedback control system is improved by 30%-50% compared with the traditional system.

[0059] 4. Through uncertainty analysis, the error rate of the system is reduced by about 15%-30%, avoiding excessive losses or premature failures caused by unstable factors. Brief Description of the Drawings

[0060] Figure 1 It is the schematic diagram of the life prediction and health management system for the key devices of the switching power supply;

[0061] Figure 2Schematic diagram of the operation interface of the life prediction and health management system for key devices of a switching power supply. Specific implementation manners

[0062] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention. It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0063] The present invention aims at the above problems and proposes a life prediction and health management system and method for key devices of a switching power supply. By integrating multi-parameter monitoring technology, this system combines low-frequency parameters (such as source-drain voltage drop, source-drain current, on-resistance, threshold voltage) with high-frequency transient parameters (such as transient switching voltage, transient switching current, and switching loss power), and at the same time introduces external stress factors such as ambient temperature and irradiation to construct a comprehensive device health assessment model. With the help of physical modeling and adaptive deep learning methods, this system can real-time predict the health state (SOH) and remaining useful life (RUL) of key devices, and provide credibility quantification for the prediction results through Bayesian uncertainty analysis. On this basis, this system also designs an intelligent feedback control module to realize real-time adjustment of drive parameters, so as to reduce the switching frequency and optimize the gate drive voltage under the device health state, thereby reducing power loss and heat accumulation and extending the device service life.

[0064] Specific implementation manner 1: Refer to Figure 1 and Figure 2 to specifically illustrate this implementation manner. The life prediction and health management system for key devices of a switching power supply described in this implementation manner includes: an edge computing end and a cloud computing end, which can be used in any switching power supply with a MOSFET or IGBT device as the core. Taking the MOSFET device of a DCDC converter as an example to introduce this implementation manner.

[0065] I. Edge computing end, this part mainly realizes:

[0066] 1. Collect key parameters of MOSFET, such as: source-drain voltage drop V DS , source-drain current I D , source-drain resistance R DS(on) , threshold voltage V th , transient switching voltage transient switching current and switching loss power P sw .

[0067] 2. Perform high-speed ADC sampling and filtering processing.

[0068] 3. Perform data preprocessing through FPGA and send the data to the cloud computing terminal via 5G / LoRa.

[0069] The specific hardware components are as follows:

[0070] Table 1: Main hardware components and functions of the edge computing terminal of the DCDC converter key device health monitoring system

[0071]

[0072]

[0073] 1.1 Data acquisition and processing part

[0074] Use an ADC (digital-to-analog converter) and an ADC (high-speed) to collect the low-frequency and high-frequency signals of the device respectively, and obtain the discrete sampling sequence x[n].

[0075] For the high-speed ADC data noise, a new recursive low-pass filter is proposed:

[0076] y[n] = y[n - 1] + α(x[n] - y[n - 1]) + β(x[n] - 2x[n - 1] + x[n - 2]),

[0077] where y[n] is the output of the recursive low-pass filter, and n represents the sampling point index.

[0078] α controls the basic filtering response, and β introduces a second-order difference correction to suppress high-frequency jitter. By introducing the difference term, the robustness of the filter to sudden interference is improved.

[0079] For the MOSFET low-frequency signal collected by the ADC (digital-to-analog converter): the source-drain voltage drop V DS and the source-drain current I D , use the dynamic adaptive exponential filtering and drift correction method for data processing. This processing can achieve both rapid response to dynamic changes and suppression of noise and drift. The specific formula is as follows:

[0080]

[0081] In the formula, is the estimated value after filtering and drift correction, and α n represents the dynamic adaptive filtering coefficient, and Δ d represents the drift correction term.

[0082] For the dynamic adaptive filtering coefficient α n it can be obtained by the following formula:

[0083]

[0084] In the formula, represents the noise variance calculated based on the most recent n samples; k is an adjustment constant used to control the influence of noise on the filtering weights.

[0085] The drift correction term Δ d is estimated by linear regression within a sliding window:

[0086]

[0087] In the formula, is the slope, obtained by fitting a linear model x = mt + b within the sliding window {x[n - N + 1],..., x[n]}, and Δt is the sampling interval.

[0088] The second derivative and adaptive weights are introduced to more accurately capture the smooth transition region of the curve inflection point for the threshold voltage V th which is calculated as follows:

[0089]

[0090] In the formula, V GS is the gate voltage; represents the second derivative of the I D -V GS curve; is an adaptive weighting function used to enhance the calculation accuracy near the inflection point; V c is the estimated critical voltage; η and are both adjustment parameters.

[0091] 1.2 Data Transmission Part

[0092] In the data transmission module, TLS is used to encrypt data transmission to ensure the security of data under the MQTT or UDP protocol. To prevent data loss during transmission, a dynamic polynomial CRC check is introduced. This method makes it adapt to environmental changes and further improves data integrity and security. The specific formula is:

[0093]

[0094] In the formula, CRC dyn represents the check result; Q represents the number of data bytes to be checked this time, data[q] represents the q-th data byte; h(q) represents a dynamic perturbation function based on the sine function and h(q) = |ζ·sin(ωq + φ)|; P q (x) is a dynamically selected polynomial generator, selected according to the real-time detected channel noise characteristics; MOD(·, P q(x) represents the polynomial modulo operation on the dynamic generator P q (x) polynomial modulo operation, where x represents a formal variable and does not represent a specific data value. It is used as a placeholder. ζ represents the sine perturbation amplitude, which is used to control the mask amplitude. ω represents the angular frequency, and φ represents the phase shift.

[0095] II. Cloud Computing End

[0096] 2.1 Physics-Based Adaptive Deep Learning Model

[0097] First, establish the MOSFET physical model, considering the device's low-frequency and high-frequency parameters: V DS 、V GS 、I D 、R DS(on) 、V th 、 and P sw etc. By normalizing, the dimensional differences are eliminated, enabling the contributions of each parameter to be superimposed within the same framework.

[0098] For the low-frequency parameters: V DS 、I D 、R DS(on) and V th , while considering the operating temperature and radiation effects, the MOSFET conduction state degradation model f cond (t) is established as:

[0099]

[0100] In the formula, V DS (t), I D (t), R DS(on) (t) and V th (t) are the V DS , I D , R DS(on) and V th at time t respectively. V DS (0), I D (0), R DS(on) (0) and V th (0) are the V DS , I D , R DS(on) and V th at the initial moment respectively. Then and respectively represent V DS (t), I D (t), R DS(on) (t) and V th(t) The normalized value relative to the initial state, which is used to reflect the change of device performance over time; α1, α2, α3, and α4 are the parameter exponential factors corresponding to each low-frequency parameter, which are used to characterize the non-linear effect of each parameter on aging; β1 is the low-frequency exponential factor for overall normalization and adjustment; E(t) is the low-frequency coupling factor, which is used to reflect the influence of the environment. The specific formula is as follows:

[0101]

[0102] In the formula, T(t) is the operating temperature at time t; T ref is the reference temperature; T s is the temperature scale factor, which determines the sensitivity of the temperature change to the aging rate; Φ(t) represents the dose rate at time t; λ is the low-frequency irradiation sensitivity coefficient, which reflects the acceleration effect of irradiation on device aging.

[0103] For high-frequency parameters and P sw , the MOSFET switching process degradation model f sw (t) is established as:

[0104]

[0105] In the formula, and P sw (t) are respectively the and P sw at time t, and P sw (0) are respectively the and P sw at the initial moment, and respectively represent the normalized values of the transient parameters and P sw (t) relative to the initial state. γ1, γ2, and γ3 are the parameter exponential factors corresponding to each high-frequency parameter respectively, β2 is the high-frequency exponential factor, and E(t) (sw) represents the high-frequency coupling factor. The specific formula is:

[0106]

[0107] In the formula, μ is the high-frequency irradiation sensitivity coefficient, which describes the irradiation sensitivity during the switching process.

[0108] Subsequently, the MOSFET on-state degradation model and the MOSFET switching process degradation model are integrated through the weight coefficient to establish the total degradation index model:

[0109] AI(t) = w cond ·f cond (t) + wsw ·f sw (t),

[0110] where AI(t) represents the total degradation index at time t; w cond represents the weight of on-state degradation, w sw represents the weight of switching process degradation, w cond +w sw = 1.

[0111] Furthermore, to intuitively reflect the health change of the MOSFET from the initial state to the degradation failure process, the health state SOH(t) of the device is defined as:

[0112]

[0113] where AI crit is the set critical aging index. When AI(t) reaches AI crit , the device is considered to have reached the failure boundary.

[0114] An adaptive Transformer deep learning model (Ad-Transformer) is established, which has the ability of adaptive dynamic adjustment. By adding a dynamically learnable attention adjustment term, a dynamic attention correction matrix A t is introduced:

[0115] A t = W A ·h t ,

[0116] where W A represents the learnable weight matrix for calculating the dynamic adjustment value; h t is the hidden state of the new data input.

[0117] The dynamic attention calculation is as follows:

[0118]

[0119] where Q, K, and V represent the multi-dimensional key parameter matrices of the device. The row index of the matrix corresponds to the time step, and the column index corresponds to the multi-dimensional feature coordinate system; d k represents the dimension of matrix K; softmax represents the attention weight determined after normalization.

[0120] For the loss function of Ad-Transformer, an Adaptive Bayesian Incremental Loss (ABiLoss) is established, specifically as follows: An incremental learning mechanism is established to enable the model to be partially updated when new data arrives, rather than being completely retrained, and to ensure that the model does not forget historical data; Bayesian regularization is established to model probability distributions, and an adaptive weighting term is established to enable the model to be adaptively adjusted. The specific form of the Adaptive Bayesian Incremental Loss function ABiLoss is as follows:

[0121]

[0122] In the formula, N is the number of input samples, M is the number of historical samples, M < N, i = 1, 2,..., N, j = 1, 2,..., M; λ1 is the MSE error weight, λ2 is the Bayesian regularization weight, and λ3 is the data distillation weight; is the predicted value of the i-th sample, is the actual value of the i-th sample; μ i represents the predicted average value of the i-th output sample, represents the prior distribution, i.e., the standard normal distribution; ω i is the dynamic weighting coefficient of the i-th sample, is the uncertainty estimate of the i-th sample, ∈ is a stability parameter to prevent unstable formula solving; KL(·) represents Bayesian regularization; is the output of the hidden layer of the model under the current iteration, is the output of the hidden layer of the model under the previous iteration.

[0123] To address the fast adaptability of the Ad-Transformer model adjustment under different working conditions of MOSFET, during the training of the Ad-Transformer model parameters, an Adaptive Incremental Meta-Learning method (AIML) for the Ad-Transformer model is established to calculate the loss under each working condition and update the hyperparameters of the Ad-Transformer model:

[0124] θ′ i = θ - αe i ▽L Ti (θ),

[0125] In the formula, θ i ′ is the updated parameter of the model under the i-th sample environmental condition, θ is the original parameter of the Ad-Transformer model, α is the inner loop learning rate, e i is the meta-learning adaptive factor, is the loss function under the i-th sample environmental condition, represents the loss gradient.

[0126] The Ad-Transformer model can automatically adjust and update the meta-learning adaptive factor e through the Bayesian incremental loss (ABiLoss). i :

[0127]

[0128] Subsequently, the Ad-Transformer model calculates the updated losses on multiple working conditions and calculates the global optimization direction θ of the parameters * :

[0129]

[0130] where β is the learning rate of the outer loop of the model.

[0131] To dynamically evaluate the credibility of the prediction, a Bayesian uncertainty quantification formula that combines the prediction error and the output variance of the Ad-Transformer model is introduced:[[]]

[0132]

[0133] where y pred represents the predicted life value, y true represents the true life value, σ represents the standard deviation of the prediction distribution of the Ad-Transformer model, κ and z are uncertainty adjustment parameters, and when the prediction error exceeds zσ, the uncertainty value U rises rapidly.

[0134] 2.2 Intelligent feedback control

[0135] The control part dynamically adjusts the drive parameters according to the health state SOH(t) and the Bayesian uncertainty U(t) output by the physics-based adaptive deep learning model, reduces the device working load, improves heat dissipation, and reduces switching losses. A decrease in SOH(t) or a high U(t) indicates that the device is in an aging or unstable working state. At this time, the controller will automatically adjust the numerical value of the PWM frequency or the gate drive voltage V GS to reduce heat accumulation and lower switching losses.

[0136] For the PWM frequency adjustment control law Δf PWM (t), a non-linear adjustment function is introduced to couple the health state error and the uncertainty:[[]]

[0137]

[0138] where SOH ref is the target health state, such as custom safety margins of 100%, 90%, etc.; H scale is the health state change scale factor; k0 is the base gain; k1 is the uncertainty adjustment coefficient.

[0139] When SOH(t) is lower than the target state, the tanh output tends to 1, thereby increasing Δf PWM (t); if U(t) is large, it further increases Δf PWM (t), and the updated real-time PWM frequency f PWM,new (t) is calculated by the formula:

[0140] f PWM,new (t) = f PWM,orig ·[1 - Δf PWM (t)],

[0141] where f PWM,orig is the initial set frequency.

[0142] For the gate drive voltage adjustment control law ΔV GS (t), the gate drive voltage is adjusted through the non-linear control law:

[0143]

[0144] where δ0 is the basic voltage adjustment coefficient and δ1 is the uncertainty sensitivity coefficient.

[0145] After updating, the formula for the real-time gate drive voltage is:

[0146] V GS,new (t) = V GS,orig -ΔV GS (t),

[0147] where V GS,orig is the initial drive voltage.

[0148] 2.3 Intelligent Report Generation

[0149] The "Intelligent Report Generation" module automatically generates a comprehensive report covering key contents such as real-time monitoring data, health status analysis, remaining life prediction, abnormal alarm, and operation and maintenance suggestions by integrating data and analysis results from the edge side and the cloud.

[0150] Obtain real-time parameters from the edge side, such as: source-drain voltage drop V DS , source-drain current I D , source-drain resistance R DS(on) , threshold voltage V th , transient switching voltage transient switching current and switching loss power P swExtract the predicted health state SOH(t), remaining life, Bayesian uncertainty U(t), and their historical parameter data of the deep learning model from the cloud. Analyze the data by selecting a large language model to complete the comprehensive interpretation of the data, and display the report through a text information list.

[0151] Specific implementation method 2: The life prediction and health management method for key devices of the switching power supply described in this implementation method includes:

[0152] Collect the key parameters of the key devices of the switching power supply. The key parameters include: source-drain voltage drop V DS , source-drain current I D , source-drain resistance R DS(on) , threshold voltage V th , transient switching voltage transient switching current and switching loss power P sw .

[0153] Specifically, this implementation method uses an ADC (analog-to-digital converter) and an ADC (high-speed) to collect the low-frequency and high-frequency signals of the device respectively, and obtains a discrete sampling sequence x[n].

[0154] For the high-speed ADC data noise, use a new recursive low-pass filter for filtering:

[0155] y[n] = y[n - 1] + α(x[n] - y[n - 1]) + β(x[n] - 2x[n - 1] + x[n - 2]),

[0156] where y[n] is the output of the recursive low-pass filter, and n represents the sampling point index.

[0157] α controls the basic filtering response, and β introduces a second-order difference correction to suppress high-frequency jitter. By introducing a difference term, the robustness of the filter to sudden interference is improved.

[0158] For the MOSFET low-frequency signals collected by the ADC (analog-to-digital converter): source-drain voltage drop V DS and source-drain current I D , use the dynamic adaptive exponential filtering and drift correction method for data processing. This processing can achieve both rapid response to dynamic changes and suppression of noise and drift. The specific formula is as follows:

[0159]

[0160] In the formula, is the estimated value after filtering and drift correction, α n represents the dynamic adaptive filtering coefficient, and Δ d represents the drift correction term.

[0161] For the dynamic adaptive filtering coefficient α n It can be obtained by the following formula:

[0162]

[0163] In the formula, represents the noise variance calculated based on the most recent n samples; k is an adjustment constant used to control the influence of noise on the filtering weight.

[0164] The drift correction term Δ d Adopts linear regression estimation within a sliding window:

[0165]

[0166] In the formula, is the slope, obtained by fitting the linear model x = mt + b within the sliding window {x[n - N + 1],..., x[n]}, and Δt is the sampling interval.

[0167] Introduce the second derivative and adaptive weight to more accurately capture the smooth transition region of the curve inflection point, and calculate the threshold voltage V th as follows:

[0168]

[0169] In the formula, V GS is the gate voltage; represents the second derivative of the I D -V GS curve; is the adaptive weighting function used to enhance the calculation accuracy near the inflection point; V c is the estimated critical voltage; η and are both adjustment parameters.

[0170] Use TLS to encrypt data transmission to ensure the security of data under the MQTT or UDP protocol. To prevent data loss during transmission, introduce dynamic polynomial CRC check, which makes it adapt to environmental changes and further improves data integrity and security. The specific formula:

[0171]

[0172] In the formula, CRC dyn represents the check result; Q represents the number of data bytes to be checked this time, data[q] represents the qth data byte; h(q) represents the dynamic perturbation function based on the sine function and h(q) = |ζ·sin(ωq + φ)|; P q (x) is a dynamically selected polynomial generator, selected according to the real-time detected channel noise characteristics; MOD(·, Pq (x) represents the polynomial modulo operation, and takes the modulus of the polynomial P q (x). Here, x represents a formal variable, which does not represent a specific data value and is used as a placeholder. ζ represents the sine perturbation amplitude, which is used to control the mask amplitude. ω represents the angular frequency, and φ represents the phase shift.

[0173] Using the source-drain voltage drop V DS and the source-drain current I D and the source-drain resistance R DS(on) and the threshold voltage V th to establish the conduction state degradation model f cond (t) of the key devices in the switching power supply:

[0174]

[0175] where V DS (t), I D (t), R DS(on) (t), and V th (t) are the source-drain voltage drop, source-drain current, source-drain resistance, and threshold voltage at time t respectively. V DS (0), I D (0), R DS(on) (0), and V th (0) are the source-drain voltage drop, source-drain current, source-drain resistance, and threshold voltage at the initial time respectively. α1, α2, α3, and α4 are the parameter exponential factors corresponding to each low-frequency parameter, β1 is the low-frequency exponential factor, and E(t) is the low-frequency coupling factor. The expression of the low-frequency coupling factor E(t) is:

[0176]

[0177] In the formula, T(t) is the working temperature at time t, T ref is the reference temperature, T s is the temperature scale factor, Φ(t) represents the dose rate at time t, and λ is the low-frequency irradiation sensitivity coefficient.

[0178] Using the transient switching voltage the transient switching current and the switching loss power P sw to establish the switching process degradation model f sw (t) of the key devices in the switching power supply:

[0179]

[0180] where and P sw (t) are the transient switching voltage, transient switching current, and switching loss power at time t respectively, and P sw (0) are the transient switching voltage, transient switching current, and switching loss power at the initial moment, respectively. γ1, γ2, and γ3 are the parameter exponential factors corresponding to each high-frequency parameter, β2 is the high-frequency exponential factor, and E(t) (sw) is the high-frequency coupling factor, and the high-frequency coupling factor E(t) (sw) has the following expression:

[0181]

[0182] μ is the high-frequency irradiation sensitivity coefficient.

[0183] The on-state degradation model and the switching process degradation model are integrated through weight coefficients to establish a total degradation index model, and the total degradation index is calculated as follows:

[0184] AI(t) = w cond ·f cond (t) + w sw ·f sw (t),

[0185] where AI(t) is the total degradation index at time t, w cond is the weight of the on-state degradation of the key device of the switching power supply, w sw is the weight of the switching process degradation of the key device of the switching power supply, and w cond + w sw = 1.

[0186] The health state SOH(t) is calculated using the total degradation index:

[0187]

[0188] where AI crit is the set critical aging index. When AI(t) reaches AI crit , the key device of the switching power supply is considered to have reached the failure boundary.

[0189] The Bayesian uncertainty is obtained using the following formula:

[0190]

[0191] where U represents the Bayesian uncertainty value, y pred represents the life prediction value, y true represents the true life value, σ represents the standard deviation of the prediction distribution of the adaptive Transformer deep learning model, and κ and z are both uncertainty adjustment parameters.

[0192] The loss function of the adaptive Transformer deep learning model is the adaptive Bayesian incremental loss function, and its expression is:

[0193]

[0194] where \(i = 1, 2, \cdots, N\), \(j = 1, 2, \cdots, M\), \(N\) is the number of output samples of the adaptive Transformer deep learning model, \(M\) is the number of historical samples of the adaptive Transformer deep learning model, and \(M < N\), \(\lambda_1\) is the mean square error weight, \(\lambda_2\) is the Bayesian regularization weight, \(\lambda_3\) is the data distillation weight. \(\hat{y}_i\) is the predicted value of the \(i\)-th output sample. \(y_i\) is the actual value of the \(i\)-th output sample, \(\omega\) i is the dynamic weighting coefficient of the \(i\)-th output sample. \(u_i\) is the uncertainty estimation value of the \(i\)-th output sample, \(KL(\cdot)\) represents Bayesian regularization, \(\mu\) i represents the predicted average value of the \(i\)-th output sample. \(p(\cdot)\) represents the prior distribution. \(h_t\) is the output of the hidden layer of the model at the current iteration. \(h_{t - 1}\) is the output of the hidden layer of the model at the previous iteration.

[0195] Couple the health state error and uncertainty to obtain the PWM frequency adjustment control law \(\Delta f\) PWM (t):

[0196]

[0197] where \(SOH\) ref is the target health state, \(H\) scale is the health state change scale factor, \(k_0\) is the base gain, \(k_1\) is the uncertainty adjustment coefficient, and \(U(t)\) is the Bayesian uncertainty value at time \(t\).

[0198] Use the PWM frequency adjustment control law \(\Delta f\) PWM (t) to update the PWM frequency \(f\) PWM,new (t) at time \(t\):

[0199] f PWM,new (t)=f PWM,orig \(\cdot[1 - \Delta f\) PWM (t)],

[0200] where \(f\) PWM,orig is the initial set frequency.

[0201] Couple the health state error and uncertainty to obtain the gate drive voltage adjustment control law \(\Delta V\) GS (t):

[0202]

[0203] Among them, δ0 is the basic voltage adjustment coefficient, and δ1 is the uncertainty sensitivity coefficient;

[0204] Using the gate drive voltage adjustment control law ΔV GS (t) to update the gate drive voltage V GS,new (t) at time t:

[0205] V GS,new (t) = V GS,orig - ΔV GS (t),

[0206] where V GS,orig is the initial drive voltage.

[0207] Specific Embodiment 3: The life prediction and health management device for key devices of a switching power supply according to this embodiment is characterized in that the life prediction and health management device for key devices of a switching power supply includes a processor and a memory, and at least one instruction is stored in the memory, and the at least one instruction is loaded and executed by the processor to implement the life prediction and health management method for key devices of a switching power supply as described in Specific Embodiment 2.

[0208] Specific Embodiment 4: A computer storage medium according to this embodiment is characterized in that at least one instruction is stored in the computer storage medium, and the at least one instruction is loaded and executed by the processor to implement the life prediction and health management method for key devices of a switching power supply as described in Specific Embodiment 2.

[0209] Generally speaking, the present invention overcomes the defects of traditional monitoring systems that only rely on a single parameter, lack real-time prediction and feedback control, have insufficient uncertainty analysis, and cannot automatically generate comprehensive maintenance reports. By integrating multi-parameter monitoring, health prediction technologies combining physics and deep learning, and a closed-loop intelligent control mechanism, the present invention not only significantly improves the health management level of key devices of a switching power supply, but also provides a quantitative and intuitive basis for operation and maintenance decisions, promoting the development of switching power supply systems towards more efficient and reliable directions.

[0210] Although the present invention has been described herein with reference to particular embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that numerous modifications may be made to the exemplary embodiments and that other arrangements may be devised, without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the different dependent claims and the features described herein may be combined in a manner different from that described in the original claims. It should also be understood that the features described in connection with a separate embodiment may be used in other described embodiments.

Claims

1. A life prediction and health management system for key components of a switching power supply, characterized in that, Including: An edge computing terminal and a cloud computing terminal; The edge computing terminal is used to collect the key parameters of the key devices of the switching power supply and send the key parameters to the cloud computing terminal. The key parameters include: source-drain voltage drop V DS , source-drain current I D , source-drain resistance R DS(on) , threshold voltage V th , transient switching voltage transient switching current and switching loss power P sw ; The cloud computing terminal is used to calculate the health state and Bayesian uncertainty of the key devices of the switching power supply according to the key parameters received by it, and to construct an adaptive Transformer deep learning model to predict the service life of the key devices of the switching power supply. It is also used to dynamically adjust the PWM frequency and gate drive voltage based on the health state and Bayesian uncertainty.

2. The life prediction and health management system for key devices of a switching power supply according to claim 1, wherein Calculating the health state of the key devices of the switching power supply includes: Using the source-drain voltage drop V DS 、source-drain current I D 、source-drain resistance R DS(on) 、threshold voltage V th to establish a conduction state degradation model for key devices of a switching power supply; Using the transient switching voltage Transient switching current And the switching loss power P sw To establish a degradation model for the switching process of key components in a switching power supply; Integrating the conduction state degradation model and the switching process degradation model through weight coefficients to establish a total degradation index model, and calculating the total degradation index; Calculating the health state using the total degradation index.

3. The life prediction and health management system for the key devices of the switching power supply according to claim 2, characterized in that The conduction state degradation model f cond (t) of the key devices of the switching power supply is expressed as: Among them, V DS (t), I D (t), R DS(on) (t) and V th (t) are the source-drain voltage drop, source-drain current, source-drain resistance, and threshold voltage at time t, respectively. V DS (0), I D (0), R DS(on) (0) and V th (0) are the source-drain voltage drop, source-drain current, source-drain resistance, and threshold voltage at the initial time, respectively. α1, α2, α3, and α4 are the parameter exponent factors corresponding to each low-frequency parameter, β1 is the low-frequency exponent factor, and E(t) is the low-frequency coupling factor; The switching process degradation model f sw (t) of the key devices of the switching power supply is expressed as: Among them, and P sw (t) are the transient switching voltage, transient switching current and switching loss power at time t, respectively, and P sw (0) are the transient switching voltage, transient switching current and switching loss power at the initial time, respectively. γ1, γ2 and γ3 are the parameter exponential factors corresponding to each high-frequency parameter, β2 is the high-frequency exponential factor, and E(t) (sw) is the high-frequency coupling factor; The expression of the total degradation index model is: AI(t) = w cond ·f cond (t) + w sw ·f sw (t), where AI(t) is the total degradation index at time t, w cond is the weight of the conduction state degradation of the key devices in the switching power supply, w sw is the weight of the switching process degradation of the key devices in the switching power supply, and w cond + w sw = 1; The expression of the health state SOH(t) is: Among them, AI crit is the set critical aging index. When AI(t) reaches AI crit , the key devices of the switching power supply are considered to have reached the failure boundary.

4. The life prediction and health management system for the key devices of the switching power supply according to claim 3, characterized in that The expression of the low-frequency coupling factor E(t) is: The high-frequency coupling factor E(t) (sw) has the following expression: Where, T(t) is the working temperature at time t, T ref is the reference temperature, T s is the temperature scale factor, Φ(t) represents the dose rate at time t, λ is the low-frequency irradiation sensitivity coefficient, and μ is the high-frequency irradiation sensitivity coefficient.

5. The life prediction and health management system for key devices of a switching power supply according to claim 1, wherein Calculating the Bayesian uncertainty of the key devices of the switching power supply includes: Obtaining the Bayesian uncertainty using the following formula: where U represents the Bayesian uncertainty value, y pred represents the predicted life value, y true represents the true life value, σ represents the standard deviation of the predicted distribution of the adaptive Transformer deep learning model, and both κ and z are uncertainty adjustment parameters.

6. The life prediction and health management system for key devices of a switching power supply according to claim 1, characterized in that, The loss function of the adaptive Transformer deep learning model is the adaptive Bayesian incremental loss function, and the expression is: where \(i = 1, 2, \cdots, N\), \(j = 1, 2, \cdots, M\), \(N\) is the number of output samples of the adaptive Transformer deep learning model, \(M\) is the number of historical samples of the adaptive Transformer deep learning model, and \(M\lt N\), \(\lambda_1\) is the mean square error weight, \(\lambda_2\) is the Bayesian regularization weight, \(\lambda_3\) is the data distillation weight, is the predicted value of the \(i\)-th output sample, is the actual value of the \(i\)-th output sample, \(\omega\) i is the dynamic weighting coefficient of the \(i\)-th output sample, is the uncertainty estimate value of the \(i\)-th output sample, \(KL(\cdot)\) represents Bayesian regularization, \(\mu\) i represents the predicted average value of the \(i\)-th output sample, represents the prior distribution, is the output of the hidden layer of the model under the current iteration, is the output of the hidden layer of the model under the previous iteration.

7. The life prediction and health management system for key devices of a switching power supply according to claim 1, wherein Dynamically adjusting the PWM frequency and gate drive voltage based on the health state and Bayesian uncertainty includes: Couple the health state error with uncertainty to obtain the PWM frequency adjustment control law Δf PWM (t): Among them, SOH ref is the target health state, H scale is the health state change scale factor, k0 is the base gain, k1 is the uncertainty adjustment coefficient, and U(t) is the Bayesian uncertainty value at time t; Adjust the PWM frequency using the control law Δf PWM (t) to update the PWM frequency f PWM,new (t) at time t: f PWM,new f(t) = f PWM,orig ·[1 - Δf PWM (t)], where f PWM,orig is the initial set frequency; Couple the health state error with uncertainty to obtain the gate drive voltage adjustment control law ΔV GS (t): Where, δ0 is the basic voltage adjustment coefficient, and δ1 is the uncertainty sensitivity coefficient; Adjust the control law of the gate drive voltage ΔV GS (t) to update the gate drive voltage V GS,new (t) at time t: V GS,new V(t)=V GS,orig -ΔV GS (t), Among them, V GS,orig is the initial driving voltage.

8. A method for predicting the lifespan and managing the health of key components in a switching power supply, characterized in that, Including: Collect the key parameters of the key devices of the switching power supply, where the key parameters include: source-drain voltage drop V DS , source-drain current I D , source-drain resistance R DS(on) , threshold voltage V th , transient switching voltage transient switching current and switching loss power P sw ; Calculating the health state and Bayesian uncertainty of the key devices of the switching power supply, and constructing an adaptive Transformer deep learning model to predict the service life of the key devices of the switching power supply; Dynamically adjusting the PWM frequency and gate drive voltage based on the health state and Bayesian uncertainty.

9. Life prediction and health management device for key components of a switching power supply, characterized in that, The life prediction and health management device for the key devices of the switching power supply includes a processor and a memory. At least one instruction is stored in the memory, and the at least one instruction is loaded and executed by the processor to implement the life prediction and health management method for the key devices of the switching power supply as claimed in claim 8.

10. A computer storage medium, characterized in that, At least one instruction is stored in the computer storage medium, and the at least one instruction is loaded and executed by the processor to implement the life prediction and health management method for the key devices of the switching power supply as claimed in claim 8.

Images