阅读说明：本技术 一种并联直流-直流变换器的寿命延长方法 (Method for prolonging service life of parallel direct current-direct current converter ) 是由 张九思 罗浩 田纪伦 李翔 李明磊 尹珅 于 2021-08-24 设计创作，主要内容包括：一种并联直流-直流变换器的寿命延长方法,本发明涉及并联直流-直流变换器的寿命延长方法。本发明的目的是为了解决现有方法难以精确预测系统的健康状态,导致无法缓解模块化功率转换系统中的老化程度不匹配,从而缩短电力电子系统寿命的问题。过程为：1.获取训练数据；2.搭建神经网络模型；3.训练神经网络,得到训练好的神经网络；4.在线获得测试数据,预测并联式DC-DC变换器系统的健康状态；5.重复执行3和4N次,保留在测试数据上预测效果最好的神经网络用于最终的在线预测,执行6；6.延长系统的剩余使用寿命。本发明用于电力电子系统人工智能健康管理领域。(The invention discloses a method for prolonging the service life of a parallel direct current-direct current converter, and relates to a method for prolonging the service life of a parallel direct current-direct current converter. The invention aims to solve the problem that the existing method is difficult to accurately predict the health state of a system, so that the mismatching of the aging degree in a modular power conversion system cannot be relieved, and the service life of a power electronic system is shortened. The process is as follows: 1. acquiring training data; 2. building a neural network model; 3. training a neural network to obtain a trained neural network; 4. obtaining test data on line, and predicting the health state of the parallel DC-DC converter system; 5. repeating the execution for 3 and 4N times, reserving the neural network with the best prediction effect on the test data for final online prediction, and executing 6; 6. the remaining service life of the system is extended. The invention is used for the field of artificial intelligence health management of power electronic systems.)

1. A method for prolonging the service life of a parallel DC-DC converter is characterized in that: the method comprises the following specific processes:

step 1, acquiring training data;

step 2, building an LM-BPNN neural network model;

the LM-BPNN neural network is a Levenberg-Marquardt-reverse feedforward neural network;

step 3, training the LM-BPNN neural network to obtain a trained LM-BPNN neural network;

step 4, obtaining test data on line, and predicting the health state of the parallel DC-DC converter system;

step 5, repeating the step 3 and the step 4 for N times, reserving the LM-BPNN neural network with the best prediction effect on the test data for final online prediction, and executing the step 6;

n is a positive integer greater than or equal to 2;

and 6, prolonging the residual service life of the system.

2. The method for prolonging the service life of the parallel direct current-direct current converter according to claim 1, wherein: acquiring training data in the step 1; the specific process is as follows:

extracting voltage, current and temperature offline historical data generated by the parallel DC-DC converter system in the working process as characteristics; taking the health state of the parallel DC-DC converter system at corresponding time as a target;

the features and targets constitute training data.

3. The method for prolonging the service life of the parallel direct current-direct current converter according to claim 2, wherein: building an LM-BPNN neural network model in the step 2; the specific process is as follows:

the LM-BPNN neural network model comprises an input layer, a first hidden layer, a second hidden layer and an output layer.

4. The method of claim 3, wherein the method further comprises: training the LM-BPNN neural network in the step 3 to obtain a trained LM-BPNN neural network; the specific process is as follows:

inputting the training data obtained in the step 1 into the neural network established in the step 2, and constructing a mapping relation between voltage, current and temperature characteristic variables generated in the working process of the parallel DC-DC converter system and the health state of the system;

the optimization algorithm of the LM-BPNN neural network training is an LM optimization algorithm, the learning rate is 0.001, and the network training process is carried out under the hardware environment of 1 CPU;

and obtaining the trained neural network.

5. The method for prolonging the service life of the parallel DC-DC converter according to claim 4, wherein: the LM-BPNN neural network training process comprises the following steps:

establishing an objective optimization function, wherein an expression is shown as a formula (1):

where F (θ) is the objective optimization function, m is the number of samples in the training data set, x_iIs the input vector of the i-th training sample, y_iIs the label value of the ith training sample, f_θ(x_i) A predicted value of the network in the ith training sample is obtained, and theta is a mixed matrix formed by the weight of the network and the offset vector;

solving the minimum value of F (theta), wherein the expression is shown as a formula (2):

wherein r (θ) is a residual function between the predicted value and the true value of the network; r is_i(theta) is a residual function between a predicted value and a true value of the network corresponding to the ith training sample, n is the dimension of theta, RⁿIs n-dimensional real number, T is transposition;

the specific process for solving the minimum value of F (theta) is as follows:

suppose J_a(θ) is the Jacobian matrix for r (θ) having:

wherein the content of the first and second substances,gradient matrix being a residual function, r_iFor the residual error, θ, corresponding to the ith training sample_nAn nth dimension vector of a mixed matrix formed by the weight of the network and the offset vector;

the gradient g (θ) of F (θ) is then expressed as:

the Hessian matrix considering F (θ) may be defined as a form shown by equation (5):

h (theta) is a Hessian matrix of F (theta), and P (theta) is a second-order term of a residual error;

wherein P (θ) is shown in equation (6):

on the basis, the quadratic model m of the objective optimization function in the formula (2)_k(θ) is expressed as:

wherein, F (theta)_k) Is F (theta), g (theta) in the k-th iteration process_k) Is g (theta), theta in the k iteration process_kIs theta, H (theta) in the k-th iteration process_k) Is H (theta), r (theta) in the k-th iteration process_k) Is r (theta), J in the k-th iteration_a(θ_k) Is J in the k-th iteration process_a(θ)，P(θ_k) Is P (theta) in the k-th iteration process;

the Newton method is adopted to convert the nonlinear least squares problem of the formula (2) into a form shown in a formula (8):

θ_k+1＝θ_k-[(J_a(θ_k))^TJ_a(θ_k)+P(θ_k)]^-1(J_a(θ_k))^Tr(θ_k) (8)

wherein, theta_k+1In the k +1 th iteration processTheta of (a);

neglecting P (theta)_k) Second order term in (1), amount of change Δ during the kth iteration_kCan be expressed in the LM algorithm as the form shown in equation (10):

Δ_k＝-[(J_a(θ_k))^TJ_a(θ_k)+μ_kI]^-1(J_a(θ_k))^Tr(θ_k) (10)

wherein mu_kFor the penalty factor, I is the unit matrix.

6. The method for prolonging the service life of the parallel DC-DC converter according to claim 5, wherein: in the step 4, test data are obtained on line, and the health state of the parallel DC-DC converter system is predicted; the specific process is as follows:

collecting data generated in the online working process of the parallel DC-DC converter system, wherein the data comprises online voltage, current and temperature data generated in the working process of the parallel DC-DC converter system;

inputting data generated in the online working process of the parallel DC-DC converter system into a trained LM-BPNN neural network model;

estimating the prediction effect of the constructed LM-BPNN neural network model on the system health state in the online working process by adopting two indexes of root mean square error RMSE and average absolute error MAE;

wherein, the definition of the root mean square error is shown as expression (11):

where m 'is the number of samples in the test data set, i' is the serial number of the test samples, SOH_pi′And SOH_ti′Respectively obtaining a predicted value and a true value of the health state of the ith' test sample;

the definition of the average absolute error is shown in expression (12):

7. the method of claim 6, wherein the method further comprises: the remaining service life of the system is prolonged in the step 6; the specific process is as follows:

step 61, collecting data generated in the online working process of the parallel DC-DC converter system, wherein the data comprises online voltage, current and temperature data generated in the working process of the parallel DC-DC converter system;

inputting data generated in the online working process of the parallel DC-DC converter system into an LM-BPNN neural network model with the best prediction effect to obtain the health state of a system component;

taking the health state obtained in the LM-BPNN as input, and calculating the power weight factor of each component in the DC-DC converter system through a certain distribution rule;

and step 62, realizing the distribution strategy of each component in the system based on the step 61.

8. The method of claim 7, wherein the method further comprises: collecting data generated in the online working process of the parallel DC-DC converter system in the step 61, wherein the data comprises online voltage, current and temperature data generated in the working process of the parallel DC-DC converter system;

inputting data generated in the online working process of the parallel DC-DC converter system into an LM-BPNN neural network model with the best prediction effect to obtain the health state of a system component;

taking the health state obtained in the LM-BPNN as input, and calculating the power weight factor of each component in the DC-DC converter system through a certain distribution rule; the specific process is as follows:

let parameter d be assumed_bThe power weight value of the b-th component is expressed as the sum of the power weight values of all the components is 1, which is expressed in formula (13)Form (a):

wherein n is the total number of power components in the parallel DC-DC converter system;

suppose the damage of the b-th component is D_bThe power weight value d of the b-th component_bExpressed in the form shown in equation (15):

wherein D_b′Is the damage of the b' th component, and α is an exponential term.

9. The method of claim 8, wherein the method further comprises: in the step 62, based on the step 61, the allocation strategy of each component in the system is realized; the specific process is as follows:

for each power component, the total power is multiplied by the power weight value of the corresponding component, so as to determine the individual working power P of each power component_b ^*；

Operating power P_b ^*Divided by a reference voltage V_dObtaining a current I through an inductor_b ^*；

Based on current I_b ^*Obtaining the modulation voltage v of the power component by the PI controller_b；

Obtaining pulses (S) for generating each module based on the modulation voltage of the power component by a normalization method_b) Required duty cycle d_bAnd finishing the final control task.

10. The method of claim 9, wherein the method further comprises: the modulation voltage v based on the power component by the normalization method_iTo obtainGenerating pulses (S) for each module_i) Required duty cycle d_iThe specific process is as follows:

Technical Field

The invention relates to the interdisciplinary field of the combination of health management and artificial intelligence of a power electronic system, in particular to a method for prolonging the service life of a parallel direct current-direct current converter.

Background

With the continuous enlargement of the scale of the power system, the structure of the power grid is increasingly complex, and the smart power grid is more and more paid more attention by people. The power converter is an important device for meeting the power energy requirement and ensuring the normal operation of a power grid.

The power conversion system based on the modular connection refers to a method of connecting a plurality of power converters in series or parallel, wherein the working power of each power converter is a part of the total rated power. A parallel direct current-direct current (DC-DC) converter system, as a typical power conversion system based on modular connection, only needs a very low power loss during operation, and at the same time, has the advantages of high switching frequency and small size, and is widely used in the design of power modules. In the actual working process of a power electronic system, repeated working cycles can cause thermal mechanical fatigue of a power device, and faults such as welding spot cracking, welding wire falling, thermal cycle failure and the like of the device are caused, so that very serious consequences are generated on the power electronic system. Therefore, how to manage the health of the power electronic system is crucial.

In conventional parallel DC-DC converter systems, the power distributed by the various components is inherently constant. However, for the actual operation of the system, the related parameters of different devices, such as the on-resistance, the state voltage drop, and other electrical parameters, and the thermal expansion coefficient of the material are different. Moreover, the health of the various components in the system can vary due to temperature mismatch caused by replacement of components in the system and differences in the design of the cooling system. It is well known that damage to a component of a system can cause the entire system to fail. Therefore, how to alleviate the problem of mismatching of aging degrees in the parallel DC-DC converter system through an effective strategy becomes a research hotspot in the field of power electronic technology at present.

It is worth mentioning that at present, there is little work on prolonging the service life of the power electronic system, and there is a certain challenge on how to alleviate the problem of the mismatch of the aging degrees in the modular power conversion system. Moreover, on the premise of lacking an accurate mathematical model of the subject system, the traditional model-based method is difficult to accurately predict the health state of the system, and the application of the method is greatly limited.

In summary, the existing method is difficult to accurately predict the health status of the system, so that the mismatch of aging degree in the modular power conversion system cannot be relieved, thereby shortening the service life of the power electronic system.

Disclosure of Invention

The invention aims to solve the problem that the service life of a power electronic system is shortened because the aging degree mismatching in a modular power conversion system cannot be relieved due to the fact that the health state of the system is difficult to accurately predict by the existing method, and provides a service life prolonging method of a parallel direct current-direct current converter.

The method for prolonging the service life of the parallel direct current-direct current converter comprises the following specific processes:

step 1, acquiring training data;

step 2, building an LM-BPNN neural network model;

the LM-BPNN neural network is a Levenberg-Marquardt-reverse feedforward neural network;

step 3, training the LM-BPNN neural network to obtain a trained LM-BPNN neural network;

step 4, obtaining test data on line, and predicting the health state of the parallel DC-DC converter system;

step 5, repeating the step 3 and the step 4 for N times, reserving the LM-BPNN neural network with the best prediction effect on the test data for final online prediction, and executing the step 6;

n is a positive integer greater than or equal to 2;

and 6, prolonging the residual service life of the system.

Preferably, training data is acquired in the step 1; the specific process is as follows:

extracting voltage, current and temperature offline historical data generated by the parallel DC-DC converter system in the working process as characteristics; taking the health state of the parallel DC-DC converter system at corresponding time as a target;

the features and targets constitute training data.

Preferably, an LM-BPNN neural network model is built in the step 2; the specific process is as follows:

the LM-BPNN neural network model comprises an input layer, a first hidden layer, a second hidden layer and an output layer.

Preferably, the LM-BPNN neural network is trained in the step 3 to obtain a trained LM-BPNN neural network; the specific process is as follows:

inputting the training data obtained in the step 1 into the neural network established in the step 2, and constructing a mapping relation between voltage, current and temperature characteristic variables generated in the working process of the parallel DC-DC converter system and the health state of the system;

the optimization algorithm of the LM-BPNN neural network training is an LM optimization algorithm, the learning rate is 0.001, and the network training process is carried out under the hardware environment of 1 CPU;

and obtaining the trained neural network.

Preferably, the LM-BPNN neural network training process is:

establishing an objective optimization function, wherein an expression is shown as a formula (1):

where F (θ) is the objective optimization function, m is the number of samples in the training data set, x_iIs the input vector of the i-th training sample, y_iIs the label value of the ith training sample, f_θ(x_i) A predicted value of the network in the ith training sample is obtained, and theta is a mixed matrix formed by the weight of the network and the offset vector;

solving the minimum value of F (theta), wherein the expression is shown as a formula (2):

wherein r (θ) is a residual function between the predicted value and the true value of the network; r is_i(theta) is a residual function between a predicted value and a true value of the network corresponding to the ith training sample, n is the dimension of theta, RⁿIs n-dimensional real number, T is transposition;

the specific process for solving the minimum value of F (theta) is as follows:

suppose J_a(θ) is the Jacobian matrix for r (θ) having:

wherein ^ r_i(theta) is the gradient matrix of the residual function, r_iFor the residual error, θ, corresponding to the ith training sample_nAn nth dimension vector of a mixed matrix formed by the weight of the network and the offset vector;

the gradient g (θ) of F (θ) is then expressed as:

the Hessian matrix considering F (θ) may be defined as a form shown by equation (5):

h (theta) is a Hessian matrix of F (theta), and P (theta) is a second-order term of a residual error;

wherein P (θ) is shown in equation (6):

on the basis of the above, the target in the formula (2)Quadratic model m of the optimization function_k(θ) is expressed as:

wherein, F (theta)_k) Is F (theta), g (theta) in the k-th iteration process_k) Is g (theta), theta in the k iteration process_kIs theta, H (theta) in the k-th iteration process_k) Is H (theta), r (theta) in the k-th iteration process_k) Is r (theta), J in the k-th iteration_a(θ_k) Is J in the k-th iteration process_a(θ)，P(θ_k) Is P (theta) in the k-th iteration process;

the Newton method is adopted to convert the nonlinear least squares problem of the formula (2) into a form shown in a formula (8):

θ_k+1＝θ_k-[(J_a(θ_k))^TJ_a(θ_k)+P(θ_k)]^-1(J_a(θ_k))^Tr(θ_k) (8)

wherein, theta_k+1Theta in the k +1 th iteration process;

neglecting P (theta)_k) Second order term in (1), amount of change Δ during the kth iteration_kCan be expressed in the LM algorithm as the form shown in equation (10):

Δ_k＝-[(J_a(θ_k))^TJ_a(θ_k)+μ_kI]^-1(J_a(θ_k))^Tr(θ_k) (10)

wherein mu_kFor the penalty factor, I is the unit matrix.

Preferably, the test data is obtained online in the step 4, and the health state of the parallel DC-DC converter system is predicted; the specific process is as follows:

collecting data generated in the online working process of the parallel DC-DC converter system, wherein the data comprises online voltage, current and temperature data generated in the working process of the parallel DC-DC converter system;

inputting data generated in the online working process of the parallel DC-DC converter system into a trained LM-BPNN neural network model;

estimating the prediction effect of the constructed LM-BPNN neural network model on the system health state in the online working process by adopting two indexes of root mean square error RMSE and average absolute error MAE;

wherein, the definition of the root mean square error is shown as expression (11):

where m 'is the number of samples in the test data set, i' is the serial number of the test samples, SOH_pi′And SOH_ti′Respectively obtaining a predicted value and a true value of the health state of the ith' test sample;

the definition of the average absolute error is shown in expression (12):

preferably, the remaining service life of the system is prolonged in the step 6; the specific process is as follows:

step 61, collecting data generated in the online working process of the parallel DC-DC converter system, wherein the data comprises online voltage, current and temperature data generated in the working process of the parallel DC-DC converter system;

inputting data generated in the online working process of the parallel DC-DC converter system into an LM-BPNN neural network model with the best prediction effect to obtain the health state of a system component;

taking the health state obtained in the LM-BPNN as input, and calculating the power weight factor of each component in the DC-DC converter system through a certain distribution rule;

and step 62, realizing the distribution strategy of each component in the system based on the step 61.

Preferably, the data generated in the online working process of the parallel DC-DC converter system in step 61 is collected, including the online data of the voltage, the current and the temperature generated in the working process of the parallel DC-DC converter system;

inputting data generated in the online working process of the parallel DC-DC converter system into an LM-BPNN neural network model with the best prediction effect to obtain the health state of a system component;

taking the health state obtained in the LM-BPNN as input, and calculating the power weight factor of each component in the DC-DC converter system through a certain distribution rule; the specific process is as follows:

let parameter d be assumed_bThe power weight value of the b-th component is the sum of the power weight values of all the components is 1, and is specifically expressed in a form shown in formula (13):

wherein n is the total number of power components in the parallel DC-DC converter system;

suppose the damage of the b-th component is D_bThe power weight value d of the b-th component_bExpressed in the form shown in equation (15):

wherein D_b′Is the damage of the b' th component, and α is an exponential term.

Preferably, in the step 62, based on the step 61, the allocation policy of each component in the system is implemented; the specific process is as follows:

for each power component, the total power is multiplied by the power weight value of the corresponding component, so as to determine the individual working power P of each power component_b ^*；

Operating power P_b ^*Divided by a reference voltage V_dObtaining a current I through an inductor_b ^*；

Based on current I_b ^*Obtaining the modulation voltage v of the power component by the PI controller_b；

Obtaining pulses (S) for generating each module based on the modulation voltage of the power component by a normalization method_b) Required duty cycle d_bAnd finishing the final control task.

Preferably, the modulation voltage v based on the power component by the normalization method_iObtaining pulses (S) for generating each module_i) Required duty cycle d_iThe specific process is as follows:

the invention has the beneficial effects that:

the invention provides a service life prolonging method of a DC-DC converter system based on Levenberg-Marquardt-reverse feedforward Neural Network (LM-BPNN) and a power routing strategy. Under the condition of no need of prior knowledge of the system, the health state of the system can be predicted by constructing an algorithm model based on LM-BPNN of data, the problem of mismatching of the aging degree in the system is relieved by utilizing a power routing strategy according to the health state, the service life of the system is prolonged, and the method has important significance for predictive maintenance of key equipment of an industrial system.

1. In consideration of the fact that few work on the aspect of prolonging the service life of a power electronic system is carried out at present, the invention provides a power routing strategy which can allocate different powers to different components in a DC-DC converter system according to the health states of the components, optimize the power allocation of the system, and relieve the problem of inconsistent health states of the components in the system, thereby effectively prolonging the service life of the power electronic system.

2. The invention provides a data-based health state prediction method, which can train a data-based health state prediction model according to data generated in the working process of a system under the condition that an accurate mathematical model of a DC-DC converter system does not need to be known in advance, so that the health state of the system can be effectively predicted in real time.

3. The LM-BPNN-based health state prediction method provided by the invention adopts an LM algorithm to train a neural network. The LM algorithm combines the advantages of the gradient descent algorithm and the Gaussian-Newton algorithm, can enable the neural network to have excellent global optimization characteristics while rapidly converging, and therefore effective training is achieved.

Drawings

FIG. 1 is a work flow diagram;

FIG. 2 is a block diagram of a DC-DC converter system;

FIG. 3 is a diagram of a power routing policy architecture, P^*To distribute the total power for operation of the parallel DC-DC converter system,is the power weight value of the b-th component in the parallel DC-DC converter system,operating power, V, allocated to the b-th component for a parallel DC-DC converter system_dIn order to input the voltage, the voltage is,for setting the current of the b-th component in a parallel DC-DC converter system, I_bIs the actual value of the current, z, of the b-th component in a parallel DC-DC converter system_bFor the deviation between the set value of the current and the actual value of the current of the b-th component, k_pIs the proportionality coefficient, k, of a proportional-integral controller_iIs the integral coefficient of a proportional-integral controller, s is the Laplace operator, v_bModulated voltage, V, for the b-th component of a parallel DC-DC converter system_oTo output a voltage, d_bThe voltage duty ratio of the b component in the parallel DC-DC converter system is obtained;

FIG. 4a is a schematic diagram of accumulated damage of component 1 and component 2 of a system over time without a power routing strategy;

FIG. 4b is a schematic diagram of the cumulative difference in damage for component 1 and component 2 of the system over time without a power routing strategy;

FIG. 4c is a schematic diagram of power weights of component 1 and component 2 of the system over time without a power routing policy;

FIG. 4d is a schematic diagram of the cumulative damage of component 1 and component 2 of the system over time under a power routing policy;

FIG. 4e is a schematic diagram of the accumulated damage difference of the system component 1 and the component 2 with time under the power routing policy;

FIG. 4f is a schematic diagram of power weight of system component 1 and component 2 over time under a power routing policy;

FIG. 5 is a diagram of the LM-BPNN neural network architecture.

Detailed Description

The first embodiment is as follows: the method for prolonging the service life of the parallel direct current-direct current converter in the embodiment comprises the following specific processes:

the invention aims to solve the problem of prolonging the service life of a parallel DC-DC converter system.

On the premise of not knowing the priori knowledge of the converter system in advance, the health state of the LM-BPNN algorithm model prediction system is constructed through offline historical data such as voltage, current, temperature and the like generated in the working process of the parallel DC-DC converter system. And designing a power routing strategy, and distributing different powers for the health states of different components in the system to relieve the problem of mismatching of aging degrees, thereby prolonging the service life of the system. The work flow diagram of the present invention is shown in fig. 1.

The technical scheme of the invention is as follows: and the residual service life of the parallel DC-DC converter system is prolonged by utilizing a data-based LM-BPNN algorithm model and a power routing strategy. The method comprises the following main steps in sequence:

step 1, acquiring training data; as shown in FIG. 2;

step 2, building an LM-BPNN neural network model;

the LM-BPNN neural network is a Levenberg-Marquardt-reverse feedforward neural network;

step 3, training the LM-BPNN neural network to obtain a trained LM-BPNN neural network;

step 4, obtaining test data on line, and predicting the health state (accumulated damage) of the parallel DC-DC converter system;

step 5, repeating the step 3 and the step 4 for N times, reserving the LM-BPNN neural network with the best prediction effect on the test data for final online prediction, and executing the step 6;

n is a positive integer greater than or equal to 2;

considering that the neural network has randomness in the training process, the method performs 10 times of training, and reserves the network with the best prediction effect on the test data set for final online prediction;

and 6, prolonging the residual service life of the system.

The second embodiment is as follows: the difference between this embodiment and the first embodiment is that training data is obtained in step 1; the specific process is as follows:

extracting offline historical data such as voltage, current, temperature and the like generated in the working process of the parallel DC-DC converter system as characteristics; taking the health state of the parallel DC-DC converter system at corresponding time as a target;

the features and the targets form training data which are used as the training data of the LM-BPNN;

other steps and parameters are the same as those in the first embodiment.

The third concrete implementation mode: the first or second embodiment is different from the first or second embodiment in that an LM-BPNN neural network model is established in the step 2; the specific process is as follows:

the neural network in the invention is a fully-connected neural network, and is improved on the basis of the traditional reverse feedforward neural network.

The LM-BPNN neural network model comprises an input layer, a first hidden layer, a second hidden layer and an output layer; the structure of the network is shown in fig. 5.

Assuming that the parallel DC-DC converter system has 2 components, the input is the current, the voltage and the temperature of the 2 components in the working process, and the total number of the 6 characteristics is obtained, so the number of the neurons of the input layer is set to be 6;

the number of neurons of the first hidden layer is set to 16;

the number of neurons of the second hidden layer is set to 32;

the output is the health value of the component, so the number of neurons in the output layer is set to 1.

Other steps and parameters are the same as those in the first or second embodiment.

The fourth concrete implementation mode: the difference between the present embodiment and one of the first to third embodiments is that the LM-BPNN neural network is trained in the step 3 to obtain a trained LM-BPNN neural network; the specific process is as follows:

inputting the training data obtained in the step 1 into the neural network established in the step 2, and constructing a mapping relation between voltage, current and temperature characteristic variables generated in the working process of the parallel DC-DC converter system and the health state of the system;

the optimization algorithm of LM-BPNN neural network training is an LM optimization algorithm, the learning rate is 0.001, and the network training process is carried out under the hardware environment of 1 CPU (Intel i 7-9750H);

obtaining a trained neural network;

the neural network in the invention is improved on the basis of the traditional reverse feedforward neural network, and the LM algorithm is adopted to train the neural network. The LM algorithm can keep the gradient descent algorithm to be rapidly converged, and simultaneously has the global optimization characteristic of the Gaussian-Newton algorithm, so that the neural network is effectively trained. The health state prediction problem researched by the invention adopts a mean square error function as a loss function to optimize a neural network;

other steps and parameters are the same as those in one of the first to third embodiments.

The fifth concrete implementation mode: the difference between this embodiment and one of the first to the fourth embodiments is that the LM-BPNN neural network training process is:

establishing an objective optimization function, wherein an expression is shown as a formula (1):

where F (θ) is the objective optimization function, m is the number of samples in the training data set, x_iIs the input vector of the i-th training sample, y_iIs the label value of the ith training sample, f_θ(x_i) A predicted value of the network in the ith training sample is obtained, and theta is a mixed matrix formed by the weight of the network and the offset vector;

the training process of the neural network can be essentially regarded as solving the minimum value of the nonlinear least square problem F (theta), namely minimizing the content in the formula (2);

solving the minimum value of F (theta), wherein the expression is shown as a formula (2):

wherein r (θ) is a residual function between the predicted value and the true value of the network; r is_i(theta) is a residual function between a predicted value and a true value of the network corresponding to the ith training sample, n is the dimension of theta, RⁿIs n-dimensional real number, T is transposition;

the LM optimization algorithm is as follows:

the specific process for solving the minimum value of F (theta) is as follows:

suppose J_a(θ) is the Jacobian matrix for r (θ) having:

wherein ^ r_i(theta) is the gradient matrix of the residual function, r_iFor the residual error, θ, corresponding to the ith training sample_nAn nth dimension vector of a mixed matrix formed by the weight of the network and the offset vector;

the gradient g (θ) of F (θ) is then expressed as:

the Hessian matrix considering F (θ) may be defined as a form shown by equation (5):

h (theta) is a Hessian matrix of F (theta), and P (theta) is a second-order term of a residual error;

wherein P (θ) is shown in equation (6):

on the basis, the quadratic model m of the objective optimization function in the formula (2)_k(θ) is expressed as:

wherein, F (theta)_k) F (theta), g (theta) in the k-th iteration (for each of the N loops, it takes many iterations to iterate the result)_k) Is g (theta), theta in the k iteration process_kIs theta, H (theta) in the k-th iteration process_k) Is H (theta), r (theta) in the k-th iteration process_k) Is r (theta), J in the k-th iteration_a(θ_k) Is J in the k-th iteration process_a(θ)，P(θ_k) Is P (theta) in the k-th iteration process;

the Newton method is adopted to convert the nonlinear least squares problem of the formula (2) into a form shown in a formula (8):

θ_k+1＝θ_k-[(J_a(θ_k))^TJ_a(θ_k)+P(θ_k)]^-1(J_a(θ_k))^Tr(θ_k) (8)

wherein, theta_k+1Theta in the k +1 th iteration process;

to reduce complexity in the calculation process, P (θ) is ignored_k) A second order term in (1) can be obtainedVariation delta of nonlinear least squares problem Gaussian-Newton (Gauss-Newton) algorithm in k-th iteration process_kAs shown in formula (9):

Δ_k＝-[(J_a(θ_k))^TJ_a(θ_k)]^-1(J_a(θ_k))^Tr(θ_k) (9)

the LM algorithm is improved based on Gauss-Newton method, and the iteration step length, i.e. the variation delta in the k-th iteration process, is changed_kCan be expressed in the LM algorithm as the form shown in equation (10):

Δ_k＝-[(J_a(θ_k))^TJ_a(θ_k)+μ_kI]^-1(J_a(θ_k))^Tr(θ_k) (10)

wherein mu_kFor penalty factor, I is unit matrix, when mu_kThe value of (a) is large, for example, when the value is far more than 1, elements on the main diagonal line dominate, and the LM algorithm is equivalent to a gradient descent algorithm; when mu is_kThe LM method is equivalent to Gauss-Newton's method when the value of (a) is small, for example, close to 0. Repeating the iteration formulas (3), (4), (5), (6), (7), (8) and (10) until the iteration number reaches the upper limit;

in general, a larger μ is used in the initial stage of the neural network training process_kIt is equivalent to adopting a gradient descent algorithm to accelerate the convergence of the network. Using smaller mu in later stage of neural network training_kAnd equivalently, oscillation generated in the training process of the gradient descent algorithm is avoided by adopting a Gauss-Newton method, so that the network converges to the global optimal solution as much as possible. The LM-BPNN adopted by the invention realizes the mapping of the characteristics to the target health state by taking the 2 layers of full connection layers as hidden layers.

Other steps and parameters are the same as in one of the first to fourth embodiments.

The sixth specific implementation mode: the difference between the present embodiment and one of the first to fifth embodiments is that, in the step 4, the test data is obtained online to predict the health status of the parallel DC-DC converter system; the specific process is as follows:

collecting data generated in the online working process of the parallel DC-DC converter system, wherein the data comprises online voltage, current and temperature data generated in the working process of the parallel DC-DC converter system;

data generated in the online working process of the parallel DC-DC converter system is input into a trained LM-BPNN neural network model, so that the real-time prediction of the system health state can be realized;

evaluating the prediction effect of the built LM-BPNN neural network model on the system health state in the online working process by adopting two indexes, namely Root Mean Square Error (RMSE) and Mean Absolute Error (MAE);

the root mean square error and the average absolute error describe the difference between a predicted value and a true value of the health state, and the smaller the value is, the better the health state prediction effect of the algorithm model is; the larger the value is, the worse the health state prediction effect of the algorithm model is;

wherein, the Root Mean Square Error (RMSE) is defined as shown in expression (11):

where m 'is the number of samples in the test data set, i' is the serial number of the test samples, SOH_pi′And SOH_ti′Respectively obtaining a predicted value and a true value of the health state of the ith' test sample;

the definition of Mean Absolute Error (MAE) is shown in expression (12):

according to the training data and the test data as 5: 1 to construct training and testing data sets for LM-BPNN training and testing.

Other steps and parameters are the same as those in one of the first to fifth embodiments.

The seventh embodiment: the difference between this embodiment and one of the first to sixth embodiments is that, in the step 6, the remaining service life of the system is prolonged; the specific process is as follows:

step 61, collecting data generated in the online working process of the parallel DC-DC converter system, wherein the data comprises online voltage, current and temperature data generated in the working process of the parallel DC-DC converter system;

inputting data generated in the online working process of the parallel DC-DC converter system into an LM-BPNN neural network model with the best prediction effect to obtain the health state of a system component;

taking the health state obtained in the LM-BPNN as input, and calculating the power weight factor of each component in the DC-DC converter system through a certain distribution rule;

and step 62, realizing the distribution strategy of each component in the system based on the step 61.

Other steps and parameters are the same as those in one of the first to sixth embodiments.

The specific implementation mode is eight: the difference between the present embodiment and the first to seventh embodiments is that, in the step 61, data generated in the online working process of the parallel DC-DC converter system is collected, including online data of voltage, current and temperature generated in the working process of the parallel DC-DC converter system;

inputting data generated in the online working process of the parallel DC-DC converter system into an LM-BPNN neural network model with the best prediction effect to obtain the health state of a system component;

taking the health state obtained in the LM-BPNN as input, and calculating the power weight factor of each component in the DC-DC converter system through a certain distribution rule; the specific process is as follows:

the invention adopts a power routing strategy to prolong the residual service life of the system, and is essentially characterized in that different power is distributed to different assemblies by adjusting the voltage duty ratio of each assembly according to the actual condition of the health state, so as to implement the health management of the system. For a component with a poor health state, the power routing policy will allocate a lower power to it; for the component with good health state, the power routing strategy allocates higher power to the component, so as to achieve the purpose of delaying the system degradation. Structurally, the power routing strategy is composed of a power routing control loop and a power module control loop, and the structure diagram is shown in fig. 3. The power routing control loop takes the health state obtained in the LM-BPNN as input and outputs a power weight factor of each component in the DC-DC converter through a certain distribution rule; since the power component distributes more power, the power loss is larger, the corresponding temperature is higher, and the loss is more severe. The purpose of the power routing control loop is therefore to allocate lower operating power to the components with higher damage and higher operating power to the components with lower damage. According to the power weight factor, the voltage duty ratio of each component is changed through a closed-loop control method, so that a power distribution task is completed, the problem that the aging degree of the components in the system is not matched is solved, and the service life of the system is prolonged.

Let parameter d be assumed_bThe power weight value for the b-th component determines the actual power allocated to each component. It should be noted that, regardless of the power distribution condition inside the system, in order to ensure the normal operation of the system, the sum of the power weight values of all the components is 1, which is specifically expressed in the form shown in formula (13):

wherein n is the total number of power components in the parallel DC-DC converter system;

for a conventional parallel DC-DC converter system, the power weighting factor value of each component is equal, as shown in expression (14):

it is worth mentioning that the present invention depends on the health of the components in the DC-DC converter systemAnd the state value realizes the management of the aging degree of the system component. Specifically, the accumulated damage D of each component of the system is calculated by a rain flow counting method according to the working cycle condition of the system, and is used as a parameter for measuring the fault occurrence probability and a basis for the health state of the system components. When the cumulative damage D to the system components accumulates to 1, the system will fail. To allocate different powers for different accumulated damage components, let the damage of the b-th component be D_bThe power weight value d of the b-th component_bExpressed in the form shown in equation (15):

wherein D_b′For the damage of the b' th component, α is an exponential term, aiming to increase the difference between the damages of different components, so as to better realize the distribution of power. The power module control loop is essentially a closed loop control loop that varies the voltage duty cycle of each component according to the power weighting factor in the power routing control loop to accomplish the task of power distribution.

Other steps and parameters are the same as those in one of the first to seventh embodiments.

The specific implementation method nine: the difference between this embodiment and the first to eighth embodiment is that, in the step 62, the allocation policy of each component in the system is implemented based on the step 61; the specific process is as follows:

in the power module control loop, for each power component, the total power is multiplied by the power weight value of the corresponding component, so as to determine the individual working power P of each power component_b ^*；

Operating power P_b ^*Divided by a reference voltage V_dObtaining a current I through an inductor_b ^*；

Based on current I_b ^*Obtaining the modulation voltage v of the power component by the PI controller_b；

Power component based modulation by normalizationThe voltage acquisition generates a pulse (S) for each module_b) Required duty cycle d_bAnd finishing the final control task.

Other steps and parameters are the same as those in one to eight of the embodiments.

The detailed implementation mode is ten: this embodiment differs from the first to eighth embodiments in that the modulation voltage v based on the power component is normalized by the normalization method_iObtaining pulses (S) for generating each module_i) Required duty cycle d_iThe specific process is as follows:

other steps and parameters are the same as those in one of the first to ninth embodiments.

The following examples were used to demonstrate the beneficial effects of the present invention:

the first embodiment is as follows:

the invention adopts a parallel DC-DC converter system to verify the provided service life extension method. The power device used by the DC-DC converter system is IRFP4227 Mosfet, and the whole converter system comprises 2 converter units. The method simulates the health states of different working processes by randomly setting the initial accumulated damage, and particularly takes the data of 30 working processes as training data and the data of 6 working processes as test data which are respectively used for training an LM-BPNN model and testing the prediction effect of the LM-BPNN model on the accumulated damage. On the basis, the length of the simulation period in the working process is set to be 1 day. When the DC-DC converter system is subjected to 1 simulation cycle, the power weight factors are distributed to each component again according to the current accumulated damage of each component at the end of the simulation cycle, so that new working power is distributed, and the degradation of the converter system is delayed. The method comprises the following specific steps:

step 1, acquiring training data: and (3) performing characteristic extraction on offline historical data such as voltage, current, temperature and the like generated in the working process of the parallel DC-DC converter system. And the accumulated damage of the parallel DC-DC converter system in the corresponding time is used as the characteristic and the target of the training data set of the LM-BPNN together.

Step 2, building an LM-BPNN neural network model: and (3) building an LM-BPNN network, setting the number of the neurons of the 2-layer full connection layer as the hidden layer to be 16 and 32 respectively, and optimizing the neural network by adopting an LM algorithm.

Step 3, training an LM-BPNN neural network model: inputting the training data subjected to the feature extraction in the step 1 into the LM-BPNN established in the step 2, and establishing a mapping relation between the system accumulated damage and characteristic variables such as voltage, current and temperature generated in the working process of the parallel DC-DC converter system. The data of 30 working processes are used as training data for training the network model.

And 4, acquiring test data on line: and inputting data generated in 6 online working processes of the parallel DC-DC converter system as test data into the trained LM-BPNN to realize the real-time prediction of the accumulated damage of the system. And (3) evaluating the prediction effect of the proposed LM-BPNN model on the accumulated damage of the system in the online working process by using two indexes, namely the root mean square error and the average absolute error shown in the formulas (11) and (12).

And step 5, repeating the step 3 and the step 4 for 10 times, and reserving the LM-BPNN neural network with the best prediction effect on the test data for final online prediction, wherein the evaluation index results of 6 processes are shown in a table 1. It can be seen that the LM-BPNN provided by the invention can effectively predict the accumulated damage of the parallel DC-DC converter system.

TABLE 1 LM-BPNN prediction of cumulative lesion growth

And 6, prolonging the residual service life of the system: the health state obtained in the LM-BPNN is used as an input, and the power weight factor of each component in the system in the DC-DC converter is calculated through the power routing strategy structure diagram shown in fig. 3. According to the power weight factor, the voltage duty ratio of each component is changed through a closed-loop control method, so that a power distribution task is completed, the problem that the aging degree of the components in the system is not matched is solved, and the service life of the system is prolonged. Fig. 4a, 4b, 4c, 4d, 4e, 4f show a schematic of the life extension results of the present invention in a specific application. Specifically, the initial damage of 2 converter assemblies is set to 0.80 and 0.49 respectively, and fig. 4a, 4b, 4c, 4d, 4e, and 4f plot the accumulated damage value, the accumulated damage difference value, and the power weight value in the parallel DC-DC converter system, respectively. It can be seen from the figure that the time of system failure was 5831 days and 7470 days, respectively, before and after the life extension method of the present invention was introduced. It can be seen that the method provided by the invention can effectively alleviate the aging degree mismatch problem of the components in the system, and prolong the service life of the system, which is of great significance to the predictive maintenance of the critical equipment of the industrial system.

The present invention is capable of other embodiments and its several details are capable of modifications in various obvious respects, all without departing from the spirit and scope of the present invention.