IGBT aging prediction method and system based on BP neural network

Abstract

The invention discloses an IGBT aging prediction method and system based on a BP neural network. The method comprises the following steps: S1, outputting a plurality of different electrical data by an IGBT in operation; s2, the signal acquisition and processing module acquires and processes different electrical data to obtain processed electrical data; s3, training a BP neural network model by using the processed electrical data; inputting the trained BP neural network model into an embedded intelligent processing module; s4, performing aging prediction on the IGBT by using the intelligent processing module embedded with the BP neural network to obtain a prediction error, judging whether the prediction error is smaller than a preset error, and if so, outputting a prediction result; and otherwise, circulating the steps S3 to S4 until the prediction error is smaller than the preset error, and outputting a prediction result. The method is used for evaluating the aging of the IGBT, and guarantees are provided for prolonging the service life of the system and enhancing the reliability of the system.

Description

Technical Field

[0001] The present invention relates to the field of transistor technology, and in particular to an IGBT aging prediction method and system based on a BP neural network. Background Art

[0002] IGBT devices play a key role in many industrial and energy fields due to their excellent performance. According to different application requirements, IGBTs are designed with different operating voltage levels. With the growing global demand for new energy, especially in wind and solar power generation, DC transmission, electric vehicles, and power management of large data exchange stations, grid-connected converters are indispensable. IGBT modules are the core components of power converters. In order to achieve optimal energy conversion, the working state changes randomly with the external environment, making them the most susceptible to aging and damage. According to statistics, the failure rate caused by IGBT failure in converter systems is as high as 31%. In order to extend the service life of power modules and improve their reliability, the reliability research of power devices in power electronic systems has become a hot topic.

[0003] As the application scenarios of IGBT modules become more popular and are used in a wider range of fields, the market has put forward higher standards for the accuracy of predicting their remaining useful life (RUL). Traditional methods that rely on physical and analytical models have shown their limitations when faced with IGBT life prediction tasks with large data sets and complex parameters.

[0004] Researchers at home and abroad generally conduct evaluation research by establishing physical models, probabilistic statistical models or machine learning models.

[0005] The physical model used for IGBT health status assessment has high simulation accuracy, but the physical model is relatively complex to establish and takes a long time to run; although the probabilistic statistical model is simple, its evaluation accuracy is low; the machine learning model takes into account the influence of multiple factors, so its evaluation accuracy is higher, but the operation of the machine learning model requires data support. Summary of the invention

[0006] The present invention provides an IGBT aging prediction method and system based on a BP neural network to solve the technical problems mentioned in the background technology.

[0007] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0008] The present invention provides an IGBT aging prediction method based on BP neural network, comprising the following steps:

[0009] S1, after the voltage and current flow through the control module of the converter, a modulation signal is generated, and the modulation signal is used to control the on and off of the IGBT, so that the IGBT outputs multiple different electrical data;

[0010] S2, the signal acquisition and processing module acquires and processes different electrical data to obtain processed electrical data;

[0011] S3, using the processed electrical data to train the BP neural network model to obtain a trained BP neural network model; inputting the trained BP neural network model into the embedded intelligent processing module;

[0012] S4. Use the embedded intelligent processing module to predict the aging of the IGBT, obtain the prediction error, and determine whether the prediction error is less than the preset error. If so, it is determined that the embedded intelligent processing module can predict the aging of the IGBT and output the prediction result; otherwise, loop S3 to S4 until the prediction error is less than the preset error, and output the prediction result.

[0013] Further, the control module in S1 includes a reactive power compensation and pitch angle outer loop control subunit and a voltage and current inner loop control subunit;

[0014] The control principle of the voltage and current inner loop control subunit is:

[0015]

[0016] Among them, s is the Laplace operator, θ is the control synchronization angle, and w g is the grid angular frequency, T is the time parameter of DC voltage control; J is the inertia parameter of DC voltage control; D is the damping parameter of DC voltage control; is the DC voltage control given value; V dc Indicates the actual value of DC voltage;

[0017] The voltage and current inner loop control subunit realizes DC voltage control by simulating the rotor motion equation; the rotor motion equation is as follows:

[0018]

[0019] Where ω is the output angular frequency, w ref is the rated angular frequency; P is the active power and reactive power output from the grid side, P ref is the power of the DC input to the machine side, C dc is the DC side capacitor.

[0020] Further, the electrical data in S1 include the peak value of the IGBT collector-emitter turn-off voltage spike, the IGBT collector-emitter saturation voltage drop, the collector current, and the experiment duration;

[0021] The IGBT collector-emitter saturation voltage drop calculation formula is:

[0022]

[0023] Where V ce(sat) Indicates the IGBT collector-emitter saturation voltage drop; is the conduction voltage drop of the PNP transistor EB junction; R on is the P region on-channel resistance; I C is the on-state current, i.e., the collector current; I D represents the drain current, R x represents the drift region resistance, and the collector current can be calculated using the following formula:

[0024] I c =I Rs +I D =(1+β 1 )I D

[0025] Among them, I Rs Represents the current on the equivalent on-resistance of the P body region, β 1 Indicates the DC current amplification factor of the common emitter of the upper tube.

[0026] Furthermore, it can be seen from the internal equivalent diagram of the module that the calculation formula of the IGBT collector-emitter saturation voltage drop can be expressed by the following formula: V ce(sat) =V MOSFET +V B +V PN

[0027] Where V MOSFET is the conduction voltage drop of the MOSFET tube, V B is the voltage drop in the lightly doped region, V PN is the conduction voltage drop of the PN junction; and the conduction voltage drop V MOSFET , Lightly doped region voltage drop V B And the conduction voltage drop V PN The calculation formulas are as follows:

[0028]

[0029] Among them, α PNP is the common base gain of the PNP structure, L and Z are the length and width of the conductive channel respectively, μ ni is the inversion layer electron mobility, C oxc is the characteristic capacitance of the oxide layer, V g Represents the gate voltage, V this the threshold voltage, q is the charge, A is the silicon area, p(x) is the residual carrier concentration, μ n and μ p are the mobilities of electrons and holes, respectively, and N B is the doping concentration of the drift region, V G is the MOSFET gate threshold voltage, p 0 YesN - The boundary concentration value of the residual minority carriers, n i is the intrinsic carrier concentration.

[0030] Furthermore, the processing in S2 includes removing outliers, primary exponential smoothing, secondary exponential smoothing, and normalization to reduce data noise and standardize the electrical data range; the electrical data range is the range of the input data of the BP neural network model;

[0031] The calculation formula for a single exponential smoothing process is:

[0032] S t =μp t +(1-μ)S t-1

[0033] Among them, S c represents the exponential smoothing value of the tth period; μ represents the smoothing parameter, and its value range is [0, 1]; x t Indicates the current actual value; S t61 represents the exponential smoothing value of the t-1th period;

[0034] The calculation formula for the quadratic exponential smoothing is:

[0035]

[0036] Among them, S t (2) represents the quadratic exponential smoothing value of the tth period; S t (1) represents the exponential smoothing value of the tth period; represents the quadratic exponential smoothing value of the t-1th period;

[0037] The calculation formula for normalization is:

[0038]

[0039] Among them, y represents the normalized value, x min Represents the minimum value of the sample data, x max represents the maximum value of the sample data; x represents the current value of the sample data.

[0040] This normalization method is more suitable for situations where the values ​​are relatively concentrated. However, if x min and x max Unstable, it is easy to make the normalization result unstable, making the subsequent use effect unstable. In actual use, the empirical constant value can be used to replace x min and x max .

[0041] Furthermore, the S3 specifically includes the following steps:

[0042] S3-1, initialize the BP neural network model: determine whether there is a prediction error at present, if yes, go to S3-2; otherwise go directly to S3-3;

[0043] S3-2, optimizing and modulating the initialized BP neural network model, and updating the maximum number of weights of the current BP neural network model according to the comparison result between the prediction error and the preset error, and optimizing the approximation function of the BP neural network model; then optimizing the structure of the BP neural network model through the optimized approximation function;

[0044] S3-3. Use the processed electrical data obtained in S2 to train the BP neural network model to obtain a trained BP neural network model; input the trained BP neural network model into the embedded intelligent processing module.

[0045] Furthermore, the BP neural network model includes an input layer, a hidden layer and an output layer connected in sequence;

[0046] The input layer is used to receive input signals; the output layer is used to generate prediction errors; and the hidden layer is used to convert input data into higher-level feature representations and generate approximate functions between input and output.

[0047] The calculation formulas for the output of the hidden layer and the output of the output layer are respectively:

[0048]

[0049] Among them, x i is the input of the ith node in the input layer, v ki is the weight from the i-th node in the input layer to the k-th neuron node in the hidden layer, w jk is the weight from the kth neuron node in the hidden layer to the jth node in the output layer, a k is the threshold of the kth neuron node in the hidden layer, b j is the threshold of the jth node in the output layer, m k is the output of the kth neuron node in the hidden layer, n j is the actual output of the jth node in the output layer; as well as are different or the same activation functions respectively; α represents the total number of input layer nodes; β represents the total number of hidden layer neuron nodes; the calculation formula of β is:

[0050]

[0051] Among them, δ is the number of nodes in the output layer, and c is a constant between [0, 1].

[0052] Furthermore, the selection rule of the activation function is as follows:

[0053] a. In the initialization stage of the BP neural network model, the activation function is the Sigmoid function, and the specific formula is as follows:

[0054]

[0055] b. When the prediction error is greater than 1, the activation function is the tanh function, and the specific calculation formula is as follows:

[0056]

[0057] On the other hand, the present invention also provides an IGBT aging prediction system based on a BP neural network. The IGBT aging prediction system uses an IGBT aging prediction method to perform aging prediction on the IGBT, including:

[0058] A control module, whose input terminal is electrically isolated and connected to the IGBT, is used to generate a modulation signal to control the on and off of the IGBT;

[0059] The signal acquisition and processing module has an input end connected to the IGBT main circuit via electrical isolation on one side and connected to the control module on the other side;

[0060] A neural network training module, whose input end is electrically isolated and connected to the signal acquisition and processing module, is used to train the BP neural network model;

[0061] The embedded intelligent processing module has input ends which are electrically isolated and connected to the signal acquisition and processing module and the neural network training module, and is used to predict the aging of the IGBT based on the internally stored BP neural network model and obtain the prediction error.

[0062] Furthermore, the embedded intelligent processing module includes a sensor array, a collector-emitter saturation voltage drop peak value acquisition circuit, a filter circuit, an A / D conversion circuit, an STM32 working circuit, and a host computer communication circuit connected in sequence;

[0063] Wherein, the sensor array includes a plurality of sensors, and the plurality of sensors are respectively electrically connected to the signal acquisition and processing module to receive electrical data from the signal acquisition and processing module;

[0064] The collector-emitter saturation voltage drop peak acquisition circuit is used to collect the IGBT collector-emitter saturation voltage drop output by the sensor array:

[0065] The filter circuit is used for smoothing the IGBT collector-emitter saturation voltage drop output by the collector-emitter saturation voltage drop peak value acquisition circuit;

[0066] The A / D conversion circuit is used to sample the smoothed data and normalize the sampled data to improve the data analysis efficiency;

[0067] The STM32 working circuit is electrically connected to the host computer platform through the host computer communication circuit.

[0068] Beneficial effects of the present invention:

[0069] Compared with the traditional aging prediction method, the present invention overcomes the problem that the traditional aging prediction method relies on physical and analytical models. The BP neural network model established by the present invention only needs to readjust the weight parameters of the BP neural network model according to the changes in the samples, and the optimal BP neural network model is obtained by training with the host computer, and the aging prediction can be realized through the embedded intelligent processing module. The development cost is greatly reduced, and the aging of IGBT in a complex working environment can be scientifically and accurately predicted, thereby providing a strong guarantee for extending the system life of IGBT and enhancing system reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is the overall flow chart of the present invention;

[0071] Figure 2 Schematic diagram of reactive power compensation and pitch angle outer loop control subunit in an embodiment of the present invention, wherein (a) is a DC voltage control block diagram, and (b) is a reactive power control block diagram;

[0072] Figure 3 Schematic diagram of the voltage and current inner loop control subunit in an embodiment of the present invention;

[0073] Figure 4 This is a curve diagram of the collector-emitter turn-off transient peak voltage after smoothing in an embodiment of the present invention;

[0074] Figure 5 It is a structural diagram of the embedded intelligent processing in the present invention;

[0075] Figure 6 is a processing flow chart of embedded intelligent processing in the present invention;

[0076] Figure 7 It is a structural diagram of the BP neural network model in an embodiment of the present invention. DETAILED DESCRIPTION

[0077] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Preferred embodiments of the present invention are provided in the drawings. However, the present invention can be implemented in many other different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0078] It should also be noted that, in the embodiments of the present application, the same figure mark is used to represent the same component or the same part. For the same parts in the embodiments of the present application, the figure may only mark one of the parts or components as an example. It should be understood that the figure mark also applies to other identical parts or components.

[0079] The mainstream IGBT in the market is a modular semiconductor product composed of an insulated gate bipolar transistor chip and a FWD (freewheeling diode chip) packaged through a specific circuit bridge. During long-term operation and use, IGBT is affected by electrical stress, thermal cycles, environmental factors, etc., and its performance gradually deteriorates and its reliability decreases.

[0080] IGBTs usually have multiple devices working at the same time. The long-term effects of high voltage and high current will lead to charge accumulation inside the IGBT and electric field concentration, making the loss data such as the collector-emitter turn-off voltage spike peak and voltage complicated. The previous linear calculation formula is no longer applicable. Therefore, it is possible to consider using a nonlinear BP neural network model to solve the problem of calculating the aging prediction of IGBTs.

[0081] Neural network mainly simulates certain characteristics of brain neurons and is a mathematical method for simulating nonlinear functions. Neural network has self-learning ability and can summarize rules from a large amount of data. Even if the operation mechanism inside the network is not clear, as long as there is enough data for the network to learn and train by itself, a good input-output model can be established. At the same time, the data can be classified through the memory and analysis of the network itself, and the accuracy of the model prediction results is guaranteed through association processing. The BP neural network model (backward propagation neural network) can use nonlinear differentiable functions for weight training, successfully realizes the problem of adjusting weights during multi-layer feedforward neural network training, has powerful nonlinear mapping ability, simple structure, strong applicability, and the mathematical meaning of the BP neural network model is clear, and the learning algorithm is concise, so the BP neural network model is used to calculate the aging data. Therefore, the present invention uses the collector-emitter turn-off voltage peak value and saturation voltage drop of the IGBT as the input of the BP neural network model, and the aging prediction data of the IGBT can be obtained according to the calculation of the BP neural network model.

[0082] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0083] Reference Figure 1 The embodiment of the present application provides an IGBT aging prediction method based on a BP neural network, comprising the following steps:

[0084] S1, after the voltage and current flow through the control module of the converter, a modulation signal is generated, and the modulation signal is used to control the on and off of the IGBT (insulated gate bipolar transistor), so that the IGBT outputs different electrical data;

[0085] S2, the signal acquisition and processing module acquires and processes different electrical data to obtain processed electrical data;

[0086] S3, using the processed electrical data to train a BP (Back Propagation Neural Network, BPNN) neural network model (i.e., back propagation neural network BPNN) to obtain a trained BP neural network model; inputting the trained BP neural network model into an embedded intelligent processing module;

[0087] S4. Use the embedded intelligent processing module to predict the aging of the IGBT, obtain the prediction error, and determine whether the prediction error is less than the preset error. In this embodiment, the preset error is preferably 0.05. If so, it is determined that the embedded intelligent processing module can predict the aging of the IGBT and output the prediction result; otherwise, loop S3 to S4, feed the data back to step S3, adjust the weights of the BP neural network model, retrain the neural network until the false detection error is less than 0.05, and output the prediction result.

[0088] The IGBT chip and the anti-parallel diode (hereinafter referred to as the diode) constitute the IGBT. The present invention fully considers the actual layout of the IGBT. Instead of considering the voltage problem of the power components of the upper or lower bridge arm of the converter as a whole, each IGBT chip and diode that affects the module voltage change is considered separately, avoiding the voltage measurement error from being concentrated on one IGBT chip or diode, and further improving the accuracy of the prediction result.

[0089] In this embodiment, the control module in S1 includes a reactive power compensation and pitch angle outer loop control subunit and a voltage and current inner loop control subunit;

[0090] The control principle of the voltage and current inner loop control subunit is:

[0091]

[0092] Among them, s is the Laplace operator, θ is the control synchronization angle, and w g is the grid angular frequency, T is the time parameter of DC voltage control; J is the inertia parameter of DC voltage control; D is the damping parameter of DC voltage control; is the DC voltage control given value; V dc Indicates the actual value of DC voltage;

[0093] The voltage and current inner loop control subunit realizes DC voltage control by simulating the rotor motion equation; the rotor motion equation is as follows:

[0094]

[0095] Where ω is the output angular frequency, w ref is the rated angular frequency; P is the active power and reactive power output from the grid side, P ref is the power of the DC input to the machine side, C dc is the DC side capacitor.

[0096] The voltage and current inner loop control subunit is a cascade structure of AC voltage loop and current loop. This structure is widely used in network control and has a certain universality. Figure 3 After the DC voltage control loop and reactive power control loop in the reactive power compensation and pitch angle outer loop control subunit obtain the synchronization angle θ and output voltage U respectively, the synchronization angle θ is used as u abc and i abcThe coordinate transformation angle is converted from the stationary abc coordinate system to the rotating dq coordinate system through space vector pulse width modulation. The output voltage U is used as the d-axis given signal of the AC voltage loop. The grid voltage d-axis orientation is adopted. The inner loop cascade structure can realize the tracking of the outer loop command by the grid voltage. Among them, the voltage and current double closed loops are both controlled by PI, u is the grid voltage, i is the filter inductor current, and U is the grid voltage. t is the modulation voltage, the subscript abc represents the component in the three-phase stationary coordinate system, and the subscripts d and q represent the d-axis and q-axis components in the synchronous rotating coordinate system. The voltage loop output is the given value i of the current loop ref The modulation voltage is modulated by space vector pulse width modulation to generate the modulation signal S of the converter abc Since there is coupling between the d and q axes in the dq coordinate system, this method uses the traditional cross-decoupling method to eliminate the coupling effect of the d and q axis control, and then transforms the rotating dq coordinate system into the stationary abc coordinate system to generate a PWM signal for the converter.

[0097] Multiple aging state parameters of IGBT power modules are related to module aging, and multiple state parameters affect each other. ce It continues to rise as the module ages, and there is a definite relationship trend between the two.

[0098] In this embodiment, the electrical data in S1 includes the peak value of the IGBT collector-emitter turn-off voltage V CE-p , IGBT collector-emitter saturation voltage drop, collector current, experimental duration and other data, the IGBT collector-emitter saturation voltage drop calculation formula is:

[0099]

[0100] Where V ce(sat) Indicates the IGBT collector-emitter saturation voltage drop; is the conduction voltage drop of the PNP transistor EB junction; R on is the P region on-channel resistance; I C is the on-state current, i.e., the collector current; I D represents the drain current, R x represents the drift region resistance, and the collector current can be calculated using the following formula:

[0101] I c =I Rs +I D =(1+β 1 )I D (4)

[0102] Among them, I Rs Represents the current on the equivalent on-resistance of the P body region, β 1Indicates the DC current amplification factor of the common emitter of the upper tube.

[0103] Furthermore, it can be seen from the module internal equivalent diagram that the calculation formula of the IGBT collector-emitter saturation voltage drop can be expressed by the following formula:

[0104] V ce(sat) =V MOSFET +V B +V PN (5)

[0105] Where V MOSFET is the conduction voltage drop of the MOSFET tube, V B is the voltage drop in the lightly doped region, V PN is the conduction voltage drop of the PN junction, and the conduction voltage drop V MOSFET , Lightly doped region voltage drop V B And the conduction voltage drop V PN The calculation formulas are as follows:

[0106]

[0107] Among them, α PNP is the common base gain of the PNP structure, L and Z are the length and width of the conductive channel respectively, μ ni is the inversion layer electron mobility, C oxc is the characteristic capacitance of the oxide layer, V g Represents the gate voltage, V th is the threshold voltage, q is the charge, A is the silicon area, p(x) is the residual carrier concentration, μ n and μ p are the mobilities of electrons and holes, respectively, and N B is the doping concentration of the drift region, V G is the MOSFET gate threshold voltage, p 0 YesN - The boundary concentration value of the residual minority carriers, n i is the intrinsic carrier concentration;

[0108] Among them, the common base gain α PNP The calculation formula is as follows:

[0109]

[0110] Where W 1 is the drift region width, D A is the ambipolar diffusion coefficient, T is the residual carrier lifetime; the ambipolar diffusion coefficient D A The calculation formula is as follows:

[0111]

[0112] D m and D p They are the electron diffusion coefficient and the hole diffusion coefficient, and their calculation formulas are as follows:

[0113]

[0114] N - The boundary concentration value p of excess carriers in the region 0 It can be expressed as

[0115]

[0116] Where W 2 is the effective neutral base width, L T is the bipolar diffusion length, I snc is the reverse saturation current of the PN junction, b is the electron mobility μ n and hole mobility μ p ratio.

[0117] The intrinsic carrier concentration n i , inversion layer electron mobility μ ni The calculation formula is as follows:

[0118]

[0119] Where N c is the effective state density of conduction band electrons, N V is the valence band effective state density, and the values ​​of both are generally 2.8×10 19 cm -3 and 1.04×10 19 cm -3 ; E g is the bandgap width of silicon material.

[0120] The IGBT collector-emitter saturation voltage drop V can be accurately calculated by using the above equations (3) to (15). ce(sat) .

[0121] In this embodiment, the processing in S2 includes removing outliers, primary exponential smoothing, secondary exponential smoothing, and normalization to reduce data noise and standardize the electrical data range; the electrical data range is the range of the BP neural network model input data;

[0122] The calculation formula for a single exponential smoothing process is:

[0123] S t =μp t +(1-μ)S t-1 (16)

[0124] Among them, S t represents the exponential smoothing value of the tth period; μ represents the smoothing parameter, and its value range is [0, 1]; x t Indicates the current actual value; S t-1 represents the exponential smoothing value of the t-1th period;

[0125] The calculation formula for the quadratic exponential smoothing is:

[0126]

[0127] Among them, S t (2) represents the quadratic exponential smoothing value of the tth period; S t (1) represents the exponential smoothing value of the tth period; represents the quadratic exponential smoothing value of the t-1th period;

[0128] In this embodiment, the smoothing parameter μ is selected as 0.25, and the collected data is processed to obtain the curve of the collector-emitter turn-off transient voltage spike peak after smoothing in Figure (4).

[0129] The goal of normalization is to limit the preprocessed data to a specific interval (such as [v, 1] or [-1, 1]) to reduce the adverse effects of outliers. The benefits include: improving the efficiency of the gradient descent algorithm in finding the optimal solution, thereby accelerating the convergence of the BP neural network model training; in addition, it can also enhance the prediction accuracy of the BP neural network model.

[0130] The calculation formula for normalization is:

[0131]

[0132] Among them, y represents the normalized value, x min Represents the minimum value of the sample data, x max represents the maximum value of the sample data; x represents the current value of the sample data.

[0133] This normalization method is more suitable for situations where the values ​​are relatively concentrated. However, if x min and x max Unstable, it is easy to make the normalization result unstable, making the subsequent use effect unstable. In actual use, the empirical constant value can be used to replace x min and x max .

[0134] In this embodiment, S3 specifically includes the following steps:

[0135] S3-1, initialize the BP neural network model; determine whether there is a prediction error at present, if yes, go to S3-2; otherwise go directly to S3-3;

[0136] S3-2, optimizing and modulating the initialized BP neural network model, and updating the maximum number of weights of the current BP neural network model according to the comparison result between the prediction error and the preset error, and optimizing the approximation function of the BP neural network model; then optimizing the structure of the BP neural network model through the optimized approximation function;

[0137] S3-3. Use the processed electrical data obtained in S2 to train the BP neural network model to obtain a trained BP neural network model; input the trained BP neural network model into the embedded intelligent processing module.

[0138] In this embodiment, a data set consisting of 418 transient data points is used, wherein each data point contains 8 measurement values, the first 7 are used as input features, and the eighth is used as the target output, thereby forming 410 groups of data samples. Among these samples, the first 360 groups are used in the training phase of the BP neural network model, and the last 50 groups are used to test the prediction ability of the BP neural network model to evaluate the accuracy of the BP neural network model.

[0139] In this embodiment, refer to Figure 7 , the BP neural network model includes an input layer, a hidden layer and an output layer connected in sequence;

[0140] The input layer is used to receive input signals, and the number of its neurons is consistent with the eigenvector dimension of the input signal, which is 7. At this stage, the neurons in the input layer do not perform any signal transformation, but transmit the signal intact to the hidden layer.

[0141] Since the BP neural network model is a feedforward learning, a large number of input-output pattern mapping relationships are stored in the hidden layer without revealing the mathematical equations describing this mapping relationship in advance. Its learning rule is to use the steepest descent method to continuously adjust the network weights and thresholds through back propagation to minimize the sum of square errors of the network. By arbitrarily selecting a set of weights, the given target output is directly used as the algebraic sum of linear equations to establish a linear equation system and solve the required weights. For the determination of the number of hidden layer neuron nodes β, it is necessary to carefully consider the amount of data in the sample set, the calculation requirements and the number of sample sets, because the determination of the number of hidden layer neuron nodes β has a great impact on network training. Too many numbers will lead to too long network training time and overfitting; too few will lead to underfitting. Therefore, the calculation of the number of hidden layer neuron nodes β can refer to the empirical formula, and then continuously test within the range, and finally determine the optimal number of hidden layer neuron nodes β according to the neural network training error E and accuracy; the calculation formula of the number of hidden layer neuron nodes β is:

[0142]

[0143] Among them, δ is the number of nodes in the output layer, and c is a constant between [0,1].

[0144] The output layer is used to generate prediction errors; the number of its neurons corresponds to the dimension of the output value. The output layer aggregates the outputs of the hidden layers and determines the final output value of the network through a specific division operation.

[0145] The hidden layer is used to convert the input data into a higher-level feature representation and generate an approximate function between the input and the output;

[0146] The calculation formulas for the output of the hidden layer and the output of the output layer are respectively:

[0147]

[0148] Among them, x i is the input of the ith node in the input layer, v ki is the weight from the i-th node in the input layer to the k-th neuron node in the hidden layer, w jk is the weight from the kth neuron node in the hidden layer to the jth node in the output layer, a k is the threshold of the kth neuron node in the hidden layer, b j is the threshold of the jth node in the output layer, m k is the output of the kth neuron node in the hidden layer, n j is the actual output of the jth node in the output layer; as well as are different or the same activation functions respectively; α represents the total number of input layer nodes; β represents the total number of hidden layer neuron nodes;

[0149] In this embodiment, the calculation formulas for adjusting the output layer weight and the calculation formulas for adjusting the hidden layer weight are respectively:

[0150]

[0151] Wherein, η is the learning rate, which is set to 0.1 in this embodiment; t ji represents the actual value from the i-th neuron node to the j-th neuron node; y ji Represents the output value from the i-th neuron node to the j-th neuron node; Represents the derivative of the first activation function with respect to the normalized value y; Represents the derivative of the second activation function with respect to the normalized value y;

[0152] In this embodiment, the selection rule of the activation function is as follows:

[0153] a. In the initialization stage of the BP neural network model, the activation function is the Sigmoid function, and the specific formula is as follows:

[0154]

[0155] b. When the prediction error is greater than 1, the activation function is the tanh function, and the specific calculation formula is as follows:

[0156]

[0157] If the activation function tanh still cannot reduce the error, the weight parameters are no longer updated using formulas (22) and (23), but the PSO algorithm is used to obtain the weights.

[0158] Compared with the traditional aging prediction method, the present invention overcomes the problem that the traditional aging prediction method relies on physical and analytical models, and breaks through the limitations of the traditional aging prediction method when facing the IGBT life prediction task with large-scale data sets and complex parameters. The BP neural network model relationship between loss and aging is directly established. It only needs to readjust the parameters of the BP neural network model according to the changes in the samples. The parameters here include weight parameters, excitation functions, and initialization (exponential smoothing and normalization) data of the BP neural network model; the optimal BP neural network model is obtained by training with the host computer, and aging prediction can be achieved through the embedded intelligent processing module. The development cost is greatly reduced, and the aging of IGBT in a complex working environment can be scientifically and accurately predicted, thereby providing a strong guarantee for extending the system life of the IGBT and enhancing the system reliability.

[0159] Reference Figure 5 and Figure 6 On the other hand, the present invention also provides an IGBT aging prediction system based on a BP neural network. The IGBT aging prediction system uses an IGBT aging prediction method to perform aging prediction on the IGBT, including:

[0160] A control module, whose input terminal is electrically isolated and connected to the IGBT, is used to generate a modulation signal to control the on and off of the IGBT;

[0161] The signal acquisition and processing module has an input end connected to the IGBT main circuit via electrical isolation on one side and connected to the control module on the other side;

[0162] A neural network training module, whose input end is electrically isolated and connected to the signal acquisition and processing module, is used to train the BP neural network model;

[0163] The embedded intelligent processing module has input ends which are electrically isolated and connected to the signal acquisition and processing module and the neural network training module, and is used to predict the aging of the IGBT based on the internally stored BP neural network model and obtain the prediction error.

[0164] Reference Figure 5 In this embodiment, the embedded intelligent processing module includes a sensor array, a collector-emitter saturation voltage drop peak value acquisition circuit, a filter circuit, an A / D conversion circuit, an STM32 working circuit, and a host computer communication circuit connected in sequence;

[0165] The sensor array includes a plurality of sensors, and the plurality of sensors are electrically connected to the signal acquisition and processing module respectively to receive the electrical data of the signal acquisition and processing module; the electrical data of the signal acquisition and processing module is the peak value of the IGBT collector-emitter turn-off voltage spike V CE-p , IGBT collector-emitter saturation voltage drop, collector current, experimental duration and other data;

[0166] The collector-emitter saturation voltage drop peak value acquisition circuit is used to collect the IGBT collector-emitter saturation voltage drop output by the sensor array;

[0167] The filter circuit is used to smooth the IGBT collector-emitter saturation voltage drop output by the collector-emitter saturation voltage drop peak value acquisition circuit, so as to enhance the accuracy of the aging prediction of the embedded intelligent processing module;

[0168] The A / D conversion circuit is used to sample the smoothed data and normalize the sampled data to improve the data analysis efficiency;

[0169] The STM32 working circuit is electrically connected to the host computer platform through the host computer communication circuit.

[0170] Reference Figure 6 ,There are four threads at the software level of the embedded intelligent ,processing system, including initialization thread, data input thread, data ,processing thread, and refresh thread;

[0171] Initialization thread: allocates space and interfaces required for neural network calculations;

[0172] Data input thread: monitors the input and output of data. If the output prediction error is less than the preset error, it means that this BP neural network model can be used for IGBT aging prediction. If the output prediction error is greater than the preset error, it returns to S3-S4, updates the weights and modulation parameters, and reinitializes.

[0173] Data processing thread: including primary exponential smoothing, secondary exponential smoothing, normalization, BP neural network model initialization, modulation and weight update;

[0174] Refresh thread: When the error is greater than the preset error, the BP neural network model is updated and a prompt is sent to the host computer to retrain the BP neural network model.

[0175] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered within the protection scope of the present invention. In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. An IGBT aging prediction method based on BP neural network, characterized in that: The steps include: S1, after the voltage and current flow through the control module of the converter, a modulation signal is generated, and the modulation signal is used to control the on and off of the IGBT, so that the IGBT generates different electrical data; S2, the signal acquisition and processing module acquires and processes different electrical data to obtain processed electrical data; S3, using the processed electrical data to train the BP neural network model to obtain a trained BP neural network model; inputting the trained BP neural network model into the embedded intelligent processing module; S4. Use the embedded intelligent processing module to predict the aging of the IGBT, obtain the prediction error, and determine whether the prediction error is less than the preset error. If so, it is determined that the embedded intelligent processing module can predict the aging of the IGBT and output the prediction result; otherwise, loop S3 to S4 until the prediction error is less than the preset error, and output the prediction result.

2. The IGBT aging prediction method according to claim 1, characterized in that: The control module in S1 includes a reactive power compensation and pitch angle outer loop control subunit and a voltage and current inner loop control subunit; The control principle of the voltage and current inner loop control subunit is: Among them, s is the Laplace operator, θ is the control synchronization angle, and w g is the grid angular frequency, T is the time parameter of DC voltage control; J is the inertia parameter of DC voltage control; D is the damping parameter of DC voltage control; is the DC voltage control given value; V dc Indicates the actual value of DC voltage; The voltage and current inner loop control subunit realizes DC voltage control by simulating the rotor motion equation; the rotor motion equation is as follows: Where ω is the output angular frequency, w ref is the rated angular frequency; P represents the active power and reactive power output from the grid side; P ref Indicates the power of the DC input side of the machine; C dc Indicates the DC link capacitance.

3. The IGBT aging prediction method according to claim 1, characterized in that: The electrical data in S1 include the peak value of the IGBT collector-emitter turn-off voltage spike, the IGBT collector-emitter saturation voltage drop, the collector current, and the experiment duration; The IGBT collector-emitter saturation voltage drop calculation formula is: Where V ce(sat) Indicates the IGBT collector-emitter saturation voltage drop; is the conduction voltage drop of the PNP transistor EB junction; R on is the P region on-channel resistance; I C is the on-state current, i.e., the collector current; I D represents the drain current, R x Represents the drift region resistance, and the collector current calculation formula is as follows: I c =I Rs +I D =(1+β1)I D Among them, I Rs It represents the current on the equivalent on-resistance of the P body region, and β1 represents the DC current amplification factor of the common emitter of the upper tube.

4. The IGBT aging prediction method according to claim 3, characterized in that: The calculation formula of the IGBT collector-emitter saturation voltage drop can be expressed by the following formula: V ce(sat) =V MOSFET +V B +V PN Where V MOSFET is the conduction voltage drop of the MOSFET tube, V B is the voltage drop in the lightly doped region, V PN is the conduction voltage drop of the PN junction; and the conduction voltage drop V MOSFET , Lightly doped region voltage drop V B And the conduction voltage drop V PN The calculation formulas are as follows: Among them, α PNP is the common base gain of the PNP structure, L and Z are the length and width of the conductive channel respectively, μ ni is the inversion layer electron mobility, C oxc is the characteristic capacitance of the oxide layer, V g Represents the gate voltage, V th is the threshold voltage, q is the charge, A is the silicon area, p(x) is the residual carrier concentration, μ n and μ p are the mobilities of electrons and holes, respectively, and N B is the doping concentration of the drift region, V G is the gate threshold voltage of the MOSFET tube, p0 is the boundary concentration value of the residual carrier minority of N-, and n i is the intrinsic carrier concentration.

5. The IGBT aging prediction method according to claim 1, characterized in that: The processing in S2 includes removing outliers, primary exponential smoothing, secondary exponential smoothing, and normalization to reduce data noise and standardize the electrical data range, which is the range of the input data of the BP neural network model; The calculation formula for a single exponential smoothing process is: S t =μx t +(1-μ)S t-1 Among them, S c represents the exponential smoothing value of the tth period; μ represents the smoothing parameter, and its value range is [0, 1]; x i Indicates the current actual value; S t-1 represents the exponential smoothing value of the t-1th period; The calculation formula for the quadratic exponential smoothing is: Among them, S t (2) represents the quadratic exponential smoothing value of the tth period; S t (1) represents the exponential smoothing value of the tth period; represents the quadratic exponential smoothing value of the t-1th period; The calculation formula for normalization is: Among them, y represents the normalized value, x min Represents the minimum value of the sample data, x max represents the maximum value of the sample data; x represents the current value of the sample data.

6. The IGBT aging prediction method according to claim 4, characterized in that: The S3 specifically includes the following steps: S3-1, initialize the BP neural network model; determine whether there is a prediction error at present, if yes, go to S3-2; otherwise go directly to S3-3; S3-2, optimizing and modulating the initialized BP neural network model, and updating the weight of the current BP neural network model according to the comparison result between the prediction error and the preset error, and optimizing the approximation function of the BP neural network model; then optimizing the structure of the BP neural network model through the optimized approximation function; S3-3. Use the processed electrical data obtained in S2 to train the BP neural network model to obtain a trained BP neural network model; input the trained BP neural network model into the embedded intelligent processing module.

7. The IGBT aging prediction method according to claim 6, characterized in that: The BP neural network model includes an input layer, a hidden layer and an output layer connected in sequence; The input layer is used to receive input signals; the output layer is used to generate prediction errors; the hidden layer is used to convert input data into higher-level feature representations and generate approximate functions between input and output; The calculation formulas for the output of the hidden layer and the output of the output layer are respectively: Among them, x i is the input of the ith node in the input layer, v ki is the weight from the i-th node in the input layer to the k-th neuron node in the hidden layer, w jk is the weight from the kth neuron node in the hidden layer to the jth node in the output layer, a k is the threshold of the kth neuron node in the hidden layer, b j is the threshold of the jth node in the output layer, m k is the output of the kth neuron node in the hidden layer, n j is the actual output of the jth node in the output layer; as well as are different or the same activation functions respectively; a represents the total number of input layer nodes; β represents the total number of hidden layer neuron nodes; the calculation formula of β is: Among them, δ is the number of nodes in the output layer, and c is a constant between [0, 1].

8. The IGBT aging prediction method according to claim 7, characterized in that: The selection rules of the activation function are as follows: a. In the initialization stage of the BP neural network model, the activation function is the Sigmoid function, and the specific formula is as follows: b. When the prediction error is greater than 1, the activation function is the tanh function, and the specific calculation formula is as follows:

9. An IGBT aging pre-system based on BP neural network, characterized in that: Predicting IGBT aging using the IGBT aging prediction method according to any one of claims 1 to 8 comprises: A control module, whose input terminal is electrically isolated and connected to the IGBT, and is used to generate a modulation signal to control the on and off of the IGBT; The signal acquisition and processing module has an input end connected to the IGBT main circuit via electrical isolation on one side and connected to the control module on the other side; A neural network training module, whose input end is electrically isolated and connected to the signal acquisition and processing module, is used to train the BP neural network model; The embedded intelligent processing module has input ends which are electrically isolated and connected to the signal acquisition and processing module and the neural network training module, and is used to predict the aging of the IGBT based on the internally stored BP neural network model and obtain the prediction error.

10. The IGBT aging prediction system according to claim 9, characterized in that: The embedded intelligent processing module includes a sensor array, a collector-emitter saturation voltage drop peak value acquisition circuit, a filter circuit, an A / D conversion circuit, an STM32 working circuit, and a host computer communication circuit connected in sequence; Wherein, the sensor array includes a plurality of sensors, and the plurality of sensors are respectively electrically connected to the signal acquisition and processing module to receive electrical data from the signal acquisition and processing module; The collector-emitter saturation voltage drop peak value acquisition circuit is used to collect the IGBT collector-emitter saturation voltage drop output by the sensor array; The filter circuit is used to smooth the IGBT collector-emitter saturation voltage drop output by the collector-emitter saturation voltage drop peak value acquisition circuit; The A / D conversion circuit is used to sample the smoothed data and normalize the sampled data to improve the data analysis efficiency; The STM32 working circuit is electrically connected to the host computer platform through the host computer communication circuit.

Images