A multi-channel intelligent control method for electronic loads based on a multi-layer hierarchical architecture

Abstract

The present application proposes a multi-channel intelligent control method for electronic loads based on a multi-layer hierarchical architecture. The multi-channel sampling voltage of the parallel loading circuit of the electronic DC load is collected at high frequency, and the high-frequency collected signal is filtered by the recursive least squares RLS algorithm and then slidingly sliced ​​as a training sample set; a preset control module is established, and a fuzzy control algorithm is used to preset the MOS tube gate voltages of multiple parallel circuits to ensure that the load current of multiple parallel electronic DC circuits is evenly distributed; a dense convolutional network is used to predict abnormal circuit conditions based on sliced ​​samples; intelligent compensation is performed on the preset MOS tube gate voltage based on abnormal conditions; a multi-channel PID closed-loop feedback control module is established to achieve tracking control of the multi-channel MOS tube gate voltage after intelligent compensation. The present application proposes a multi-layer hierarchical control method to achieve multi-channel constant current equalization of electronic load circuits and improve the performance indicators of electronic DC load circuits.

Description

Technical Field

[0001] The present application belongs to the field of intelligent control based on big data, and specifically relates to a multi-channel intelligent control method for electronic loads based on a multi-layer hierarchical architecture. Background Art

[0002] The basic working principle of electronic DC load is to control the MOS tube working current flowing from the drain to the source by controlling the MOS tube gate voltage. For high-power test environments, multiple current-pulling circuits need to be connected in parallel to achieve current balancing to increase the working current. When the MOS tubes in the parallel path cannot evenly share the current in the parallel branch due to process differences or abnormal operation, the MOS tubes in the parallel branch with a larger current will be burned out, causing unnecessary losses.

[0003] At present, electronic DC loads do not have the function of current equalization regulation for multi-branch parallel circuits. It is a feasible method to collect the voltage of sampling resistors in each load-pulling circuit and select a multi-channel intelligent control algorithm to adjust the gate voltage of the MOS tube through MCU program control. For some abnormal working conditions of MOS tubes, the gate voltage compensation correction can be given in time to ensure the current equalization of the parallel current-pulling circuit. Summary of the invention

[0004] The purpose of the present application is to provide a current-sharing circuit for a current-pulling electronic DC load in parallel under a high-power environment, thereby ensuring the safe and stable operation of the electronic load.

[0005] In order to solve the above technical problems, the present application proposes a multi-channel intelligent control method of an electronic load based on a multi-layer hierarchical architecture, comprising:

[0006] The multi-channel sampling voltage of the parallel loading circuit of the electronic DC load is collected at high frequency, and the collected signal is filtered by the recursive least squares RLS algorithm. After sliding slicing, a training sample set is formed. The RLS algorithm is a fast algorithm of the least squares algorithm. For the real-time incremental data of this application, it is not necessary to calculate the inverse matrix every time a prediction is made. RLS recursively solves the optimal parameter combination of the multivariable linear regression model based on the existing inverse matrix and the new incremental data. The multivariable linear regression model is a linear model in which the output result is jointly determined by multiple variables. Each variable in the model has a variable coefficient. In this application, the multiple variable coefficients of the model are obtained by recursive solution of the RLS algorithm. The optimized model can predict the data value to be sampled based on the existing multiple sampling points. When the actual sampled data value is too different from the predicted result, the sampled value is replaced by the predicted value;

[0007] Establish a preset control module and use fuzzy control algorithm to preset the MOS tube gate programmable voltage of multiple parallel circuits to ensure that the load current of multiple parallel electronic DC circuits is evenly distributed;

[0008] Establish a dense convolutional network to predict abnormal circuit conditions based on slice samples. The dense convolutional network is a deep neural network. The slice sample data with abnormal conditions as labels is used to train the optimal parameters of the network. The trained network can predict abnormal conditions of the electronic load circuit based on input features.

[0009] Intelligent compensation of preset MOS tube gate voltage based on abnormal working conditions;

[0010] A multi-channel PID closed-loop feedback control module is established to achieve tracking control of the gate voltage of multi-channel MOS tubes after intelligent compensation.

[0011] Optionally, the high frequency acquisition of multi-channel sampling voltages of the parallel loading circuit of the electronic DC load, filtering the acquired signals using a recursive least squares RLS algorithm, and forming a training sample set after sliding slicing, includes:

[0012] The recursive least squares (RLS) algorithm is used to update the parameters of the multivariate linear regression model based on the least squares objective function. The regression model predicts the data value to be sampled according to the historical sampling points, and gives a prediction correction to the sampling value to achieve filtering. The high-frequency sampling data after filtering is slidingly sliced, and each slice is taken as a sample. The samples are divided into training set, validation set and test set.

[0013] Optionally, a preset control module is established, and a fuzzy control algorithm is used to preset the gate program-controlled voltage of the MOS tubes of multiple parallel circuits to ensure that the load current of multiple parallel electronic DC circuits is evenly distributed, including:

[0014] Based on big data driving technology, a fuzzy relationship rule between the DC load circuit load current and the MOS tube gate programming voltage is established. The fuzzy control algorithm is used to calculate the MOS tube gate programming voltage of multiple parallel load circuits to ensure that the load current of multiple parallel electronic loads is evenly distributed.

[0015] Optionally, establishing a dense convolutional network to predict abnormal circuit conditions based on slice samples includes:

[0016] A dense convolutional network Dense-Net with 4 dense blocks and 3 transition layers is built. The network improves the feature extraction ability of the model by improving the information exchange between multiple layers. A meta-learning strategy is adopted to establish a small sample circuit fault training dataset based on easily available circuit faults to pre-train the model. A generative adversarial network is selected to expand the training samples to obtain a dense convolutional network that can predict abnormal circuit conditions.

[0017] Optionally, intelligent compensation is performed on the preset MOS tube gate voltage based on abnormal operating conditions, including:

[0018] A fuzzy reasoning model of circuit abnormal conditions and electronic load current is established, and intelligent compensation is performed on multiple preset MOS tube gate voltages according to the circuit abnormal conditions. The fuzzy reasoning model takes the abnormal conditions in the form of fuzzy variables as input and the MOS tube gate voltage in the form of fuzzy variables as output. The fuzzy logic relationship between fuzzy input and fuzzy output is mined from historical data, and the MOS tube gate voltage value is given based on fuzzy reasoning according to the type of abnormal working condition of the electronic load circuit.

[0019] Optionally, a multi-channel PID closed-loop feedback control module is established to achieve tracking control of the gate voltage of multi-channel MOS tubes after intelligent compensation, including:

[0020] The MOS tube gate voltage after intelligent compensation is used as the set value, and a fuzzy PID control algorithm is designed. The MOS tube gate voltage is collected as the feedback quantity, and the program-controlled DA value is calculated by MCU based on the PID algorithm to control the MOS tube gate voltage and realize multi-channel tracking feedback control of the electronic load current. The three key control parameters of the PID algorithm, "proportional, integral, and differential", are obtained through mining using a big data driven method.

[0021] The present application proposes a multi-channel intelligent control method for electronic loads based on a multi-layer hierarchical architecture. The multi-channel sampling voltage of the parallel loading circuit of the electronic DC load is collected at high frequency, and the high-frequency collected signal is filtered by the recursive least squares RLS algorithm and then slidingly sliced ​​into a training sample set, and divided into a training set, a test set, and a verification set; a preset control module is established, and a fuzzy control algorithm is used to preset the MOS tube gate voltage of multiple parallel circuits to ensure that the load current of multiple parallel electronic DC circuits is evenly distributed; a dense convolutional network is used to predict abnormal circuit conditions based on sliced ​​samples; intelligent compensation is performed on the preset MOS tube gate voltage based on abnormal conditions; a multi-channel PID closed-loop feedback control module is established to achieve tracking control of the multi-channel MOS tube gate voltage after intelligent compensation. The present application proposes a multi-layer hierarchical control method to achieve multi-channel constant current equalization of electronic load circuits and improve the performance indicators of electronic DC load circuits. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] FIG1 is a schematic diagram of a flow chart of a multi-layer hierarchical control method for implementing multi-channel current sharing in an electronic load circuit according to an embodiment of the present application;

[0024] FIG2 is a single-branch circuit for pulling current from an electronic load provided by an embodiment of the present invention;

[0025] FIG3 is a flow chart of an abnormal operating condition identification algorithm provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0027] As shown in FIG. 1 , FIG. 1 is a flow chart of a multi-channel intelligent control method for an electronic load based on a multi-layer hierarchical architecture provided in an embodiment of the present application, which specifically includes five contents.

[0028] S11: High frequency acquisition of multi-channel sampling voltages of the electronic DC load parallel loading circuit, establishment of a recursive least squares RLS algorithm to filter the acquired signals, and a training sample set is formed after sliding slicing.

[0029] It should be noted that the least squares RLS algorithm can directly calculate the model parameters at one time through the formula, but this application continuously samples new samples in the form of a data stream. In this case, the workload of obtaining the model parameters using the inverse formula is very large and cannot be used in practice. This application chooses the recursive least squares RLS algorithm to reduce the amount of calculation, and slides and slices the sampled data after filtering, which mainly includes 4 steps.

[0030] Step 11: First sample a piece of data and use the self-supervised learning method to build a data set, using the first 10 points sampled as input attributes x 1 , with the 11th point as supervision y 1 , and so on rolling to generate n sampling voltage input samples X = [ x 1, x 2, … , x n ] T , n sampled voltage output samples y = [ y1, y 2, … , y n ] T According to formula (1), the predicted value corresponding to each output sample can be calculated = [ , , … , ] T , the weights can be optimized using the gradient descent method w The calculation formula w = ( X T X ) -1 X T y ,in, X T Represents the sample voltage input sample X The transpose of .

[0031] Step 12: When there is a new sampling voltage point, the input sample, output sample and predicted value are expressed as, X = [ x 1, x 2, … , x n, x n+1 ] T , y = [ y 1, y 2, … , y n, y n+1 ] T and = [ , , … , , ] T, where n+1 represents the newly added sampling point. To avoid repeated matrix inversion, the step 11 X T X is defined as R, X T y Defined as z , then the weight of formula (1) w The calculation formula is w = R -1 z ， For the data with new samples, the weight w* The calculation formula can be rewritten as w* = R* -1 z * ， in R* Represents the value after adding new samples X T X, z * Represents the value after adding new samples X T y。

[0032] Step 13: Calculate the new inverse matrix R* -1 Just need the old inverse matrix R -1 and new sample data x n+1 , this rolling recursive method is much less computationally intensive than directly solving the inverse matrix. According to mathematical derivation, R* The recursion formula is λR + x n+1 x n+1 T , z * The recursive calculation formula is λz + x n+1 y n+1 , λ is the forgetting factor, let p = R* -1 x n+1 , e = wx T n+1 - y n+1 , then after derivation, the weight recursive formula is w* = w – pe .

[0033] Step 14: Use the weights calculated in step 13 w As the parameters of the multivariate linear regression model, the regression model is used to predict the new sampling point. If the difference between the predicted value and the true sampling value is less than 3 times the mean of the variance, the sampling data is retained. Otherwise, the sampling point is replaced by the mean of the previous 10 sampling values ​​to achieve real-time smoothing filtering of the adopted data.

[0034] The high-frequency sampled data after filtering is intercepted by sliding with the set window size and sliding step size, and the data intercepted each time is used as a slice sample. All slice samples are randomly divided into training set, validation set and test set in proportion as the model training set.

[0035] Based on the above discussion, in an optional embodiment of the present application, the sampling frequency is 10 kHz, 10 point length is used as the number of model variables, the least squares objective function is selected, the window size m = 800 is set, the sliding step n = 6, and the division ratio of the training set, validation set and test set is selected as 2:3:5.

[0036] S12: Establish a preset control module, and use a fuzzy control algorithm to preset the MOS tube gate programmable voltage of multiple parallel circuits to ensure that the load current of multiple parallel electronic DC circuits is evenly distributed.

[0037] It should be noted that the current size of the current-pulling circuit of the electronic DC load is controlled by the size of the MOS tube gate voltage. Due to the differences in the performance parameters of each MOS tube, differentiated MOS tube gate voltages need to be set for different parallel branches to ensure that the parallel branch currents are evenly distributed. This application adopts fuzzy feedforward control based on data rules to calculate the MOS tube gate voltages of all parallel branches to ensure current equalization. The fuzzy feedforward control mainly includes 7 key steps.

[0038] Step 21: Fuzzy rules: Based on the feedback mean value of the sampling resistor voltage and the quantitative relationship between the gate voltage of the branch MOS tube, fuzzy rules for the gate voltage and the branch load current are established.

[0039] Step 22: Domain transformation, if the actual input of the feedback voltage sampled by the resistor is , The actual range of variation is , and the domain of the algorithm design in this application is , then the feedback voltage can be transformed into the domain range using equation (2), where is the scale factor, ; This application chooses to discretize the domain, so the values ​​transformed into the domain must be divided into intervals according to the range.

[0040] Step 23: fuzzy segmentation. Fuzzy segmentation is performed on the sampled voltage after the domain transformation. This application selects 7 fuzzy variables [NBNM NS ZE PS PM PB]. In order to determine the membership relationship between the domain discrete value and the fuzzy variable, a membership table in tabular form is established to realize fuzzy segmentation.

[0041] Step 24: Fuzzy control rules. The fuzzy control rule antecedents of this application select two input variables: sampling feedback voltage and sampling feedback voltage and program-controlled DA error. The rule form selects state evaluation fuzzy control rules. The quantitative relationship of fuzzy rules needs to be continuously optimized through data-driven statistical methods.

[0042] Step 25: Fuzzy reasoning, the fuzzy rule base can be expressed as formula (3). It can be expressed as , the fuzzy meaning of all fuzzy control rules is Assume that the fuzzy variable of the fuzzy controller input is known to be: , y is , then according to the fuzzy control rules, the output The fuzzy variables (consisting of fuzzy sets indicates) . (3)

[0043] Step 26: Clarification: Based on the area centroid method of formula (4), the fuzzy value of the BG3 gate voltage is clarified, where: is the membership value of the fuzzy variable in the membership table, is the domain discrete value of the corresponding variable, and then the fuzzy value is inversely transformed into the actual control quantity, that is, the real control value of the gate voltage of the MOS tube. (4)

[0044] Step 27: The preset module inputs the gate voltage of the MOS tube of each parallel circuit calculated in step 26 to the MCU program control module. The program control module outputs a digital quantity for each parallel branch DA module according to the relationship between the digital quantity and the analog quantity. The digital quantity controls the gate voltage of the MOS tube through the analog quantity generated by the DA, thereby controlling the current of the parallel branch.

[0045] Based on the above discussion, in an optional embodiment of the present application, 7 fuzzy variables are selected, fuzzy rule antecedents select 2 inputs of feedback quantity and error value, three parallel branches are selected, and 3 MOS tube gate voltage values ​​are set at one time.

[0046] S13: Use a dense convolutional network to predict circuit abnormal conditions based on slice samples.

[0047] It should be noted that when an abnormality occurs in the electronic load current circuit, the MOS gate voltage preset based on the normal state fuzzy rule is no longer applicable, and the preset parameters need to be corrected and compensated according to the abnormality type. The types of abnormalities that affect the load current mainly include "current mutation, MOS tube start and stop asynchronous, and current amplitude limit". This application is based on high-frequency collected resistor sampling voltage data and uses a dense convolutional network to predict the abnormal type of the electronic load load circuit. The model training process is shown in Figure 2, which specifically includes 3 steps.

[0048] Step 31: Build a dense convolution network Dense-Net with 4 dense blocks and 3 transition layers. The network establishes dense connections for all layers of convolution features, directly connects the input layer to the output layer, and improves the feature extraction ability of the model by improving the information exchange between multiple layers. Establish a training data set with three types of anomalies as labels: "current mutation, MOS tube start-stop asynchrony, and current amplitude limit". ① To identify the "current mutation" anomaly, extract the FFT spectrum time series and original voltage waveform slices of the sampled slice data as identification features. ② To identify the "MOS tube start-stop asynchrony" anomaly, extract the envelope morphology of the data collected when the electronic load is powered on and off as the identification feature. ③ To identify the "current amplitude limit" anomaly, establish the maximum load current threshold of each parallel branch according to the MOS tube performance index parameters, and use whether the current value converted from the sampled voltage is greater than the threshold as the identification feature.

[0049] Step 32: The training samples established in step 31 are small samples. In order to increase the amount of training samples, the generative adversarial network is selected to expand the training samples.

[0050] Step 33: In order to further improve the model training effect, this application first selects a dense convolutional network that has been pre-trained with a large amount of time series data and image data, and then adopts a meta-learning strategy to establish a small sample circuit fault training data set based on easily available circuit faults, and fine-tune the pre-trained model to finally obtain a dense convolutional network to identify abnormal operating conditions of electronic load circuits.

[0051] Based on the above discussion, in an optional embodiment of the present application, the Dense-Net network architecture of the present application includes 4 dense blocks and 2 transition layers, and each basic block of the dense block has 4 layers, including 2 normalization layers, 2 activation layers, 1 1×1 convolution layer and 1 3×3 convolution layer. The transition layer includes a normalization layer, an activation layer, a 1×1 convolution layer and 5 filters, and a 2×2 average pooling layer.

[0052] S14: Intelligently compensate the preset MOS tube gate voltage based on abnormal working conditions.

[0053] It should be noted that when the three abnormalities of "current mutation, MOS tube start and stop asynchronous, and current amplitude limit" occur, the preset parameters cannot ensure equal current distribution of the parallel load branches. This application adopts a fuzzy reasoning method to establish a fuzzy logic relationship between various abnormal types and the MOS tube gate voltage, and intelligently compensates the preset parameters based on the predicted load circuit abnormalities, which specifically includes two steps.

[0054] Step 41: The sudden change of current is generally related to the damage of MOS tube. When the current of a parallel branch MOS tube suddenly increases, the gate voltage of the branch MOS tube needs to be reduced to suppress the increase of current. On the contrary, when the branch current suddenly decreases, the gate voltage of the MOS tube needs to be increased. When the load circuit is powered on, if a branch MOS tube has a delayed opening phenomenon, the current of the branch will be much lower than the load current of other branches. At this time, a momentary pulse gate voltage is needed to increase the current of the branch in time to protect other branch MOS tubes from damage. When the load current has reached the critical maximum working current of the MOS tube, the gate voltage should be reduced in time to prevent the tube from being burned by the large current. Based on the above logical relationship, a fuzzy reasoning model of the branch load circuit is established.

[0055] Step 42: When the dense convolution network running in real time predicts that the load circuit is abnormal, an intelligent compensation method for the gate voltage of each parallel branch MOS tube is given based on the fuzzy reasoning model to compensate and correct the gate voltage given by the preset module.

[0056] Based on the above discussion, in an optional embodiment of the present application, the "current mutation" anomaly is designed with 7 fuzzy variables of current "too high, very high, relatively high, normal, relatively low, very low, too low", corresponding to 7 MOS tube gate voltage regulation strategies of "large reduction, medium reduction, small reduction, unchanged, small increase, medium increase, large increase". Similarly, the two anomaly types of "MOS tube start and stop are not synchronized, current amplitude limit" also use 7 fuzzy variables to establish a fuzzy reasoning model.

[0057] S15: Establish a multi-channel PID closed-loop feedback control module to achieve tracking control of the gate voltage of the multi-channel MOS tube after intelligent compensation.

[0058] It should be noted that the MOS tube gate voltage finally given by the intelligent compensation module is the voltage value actually required by the electronic load pulling circuit, which can ensure the current equalization of multiple parallel branches. The voltage value is generated by the DA chip conversion after the digital quantity given by the MCU program control module. In order to overcome the unstable gate voltage caused by the environmental differences of the hardware circuit, this application continuously corrects the MCU program control output according to the gate voltage sampling value through the feedback adjustment method, and finally realizes the stable tracking of the MOS tube gate multi-channel voltage setting value given by the intelligent compensation module, which mainly includes 3 steps.

[0059] Step 51: Figure 3 is a single-branch circuit of an electronic DC load. The sampling resistor voltage collected by this application is proportional to the load current. Taking a single-branch circuit as an example, the intelligent compensation module gives the gate voltage of the MOS tube BG3, which determines the internal resistance of the MOS tube, and the internal resistance determines the load current. The load current converts the current into voltage through the sampling resistor. In order to stably control the BG3 gate voltage at the given value of the intelligent compensation module, it is necessary to collect the BG3 gate voltage through AD, compare it with the given value of the intelligent compensation inside the MCU and calculate the error. The fuzzy PID control algorithm is used to calculate the program-controlled output digital quantity based on the error, and the gate control voltage is output through the DA converter. The gate voltage of BG3 is stably controlled at the set value of the compensation module through dynamic adjustment based on feedback.

[0060] Step 52: The PID control algorithm has three key parameters, namely, proportion, integration and differentiation. The proportion P multiplied by the feedback error gives the current adjustment value. The integral I accumulates the past adjustment values ​​to stabilize the output within the set value range. The differential D is the difference between the current feedback and the previous feedback. According to the direction of change, a predictive extraction adjustment is given. According to the parameter calculation principle of the PID algorithm, a quantitative PID fuzzy adjustment rule is designed to give a programmable digital quantity based on the feedback error and the difference between the two previous and subsequent samplings.

[0061] Step 53: The three key parameters of the PID algorithm need to be calibrated according to real data. This application selects a big data statistical method and a step response method to learn the optimal parameter combination from a large amount of feedback adjustment history records through continuous adjustment.

[0062] Based on the above discussion, in an optional embodiment of the present application, the electronic DC load circuit selects three parallel branches, each branch selects three MOS tubes in series to improve the circuit voltage tolerance range, and the MCU selects 89C52 single-chip microcomputer.

[0063] The present application uses specific examples to illustrate the principles and implementation methods of the present invention, and the description of the above embodiments is only used to help understand the method and core ideas of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A multi-channel intelligent control method for electronic loads based on a multi-layer hierarchical architecture, characterized in that: include: The multi-channel sampling voltage of the parallel loading circuit of the electronic DC load is collected at high frequency, and the high-frequency collected signal is filtered by the recursive least squares RLS algorithm and then the sliding slice is used as the training sample set; Establish a preset control module and use fuzzy control algorithm to preset the MOS tube gate program-controlled voltage of multiple parallel circuits to ensure that the load current of multiple parallel electronic DC circuits is evenly distributed; Using dense convolutional networks, abnormal circuit conditions are predicted based on slice samples; Intelligent compensation of preset MOS tube gate voltage based on abnormal working conditions; A multi-channel PID closed-loop feedback control module is established to achieve stable tracking control of the gate voltage of multi-channel MOS tubes after intelligent compensation.

2. The electronic load multi-channel intelligent control method based on a multi-layer hierarchical architecture as claimed in claim 1, characterized in that: The high-frequency acquisition signal is filtered by the recursive least squares RLS algorithm and then the sliding slice is used as the training sample set, including: The recursive least squares (RLS) algorithm is used to update the parameters of the multivariate linear regression model based on the least squares objective function. The regression model predicts the data value to be sampled according to the historical sampling points, gives a prediction correction to the sampling value to achieve filtering, and slides the filtered high-frequency sampling data to form a training sample set.

3. The electronic load multi-channel intelligent control method based on a multi-layer hierarchical architecture as claimed in claim 1, characterized in that: The fuzzy control algorithm is used to preset the MOS tube gate programmable voltage of multiple parallel circuits, including: Based on big data driving technology, a fuzzy relationship rule between the DC load circuit loading current and the MOS tube gate programming voltage is established. The fuzzy control algorithm is used to calculate the MOS tube gate programming voltage of multiple parallel loading circuits to ensure that the current of multiple parallel electronic load loading circuits is evenly distributed.

4. The electronic load multi-channel intelligent control method based on a multi-layer hierarchical architecture as claimed in claim 1, characterized in that: Using dense convolutional networks, circuit abnormal conditions are predicted based on slice samples, including: A dense convolutional network Dense-Net with 4 dense blocks and 3 transition layers is built. The network improves the feature extraction ability of the model by improving the information exchange between multiple layers. A meta-learning strategy is adopted to establish a small sample circuit fault training dataset based on easily available circuit faults to pre-train the model. A generative adversarial network is selected to expand the training samples to obtain a dense convolutional network that can predict abnormal circuit conditions.

5. The electronic load multi-channel intelligent control method based on a multi-layer hierarchical architecture as claimed in claim 1, characterized in that: Intelligent compensation of preset MOS tube gate voltage based on abnormal working conditions, including: A fuzzy reasoning model of circuit abnormal conditions and electronic load current is established, and intelligent compensation is performed on the preset gate voltages of multiple MOS tubes according to the circuit abnormal conditions.

6. The electronic load multi-channel intelligent control method based on a multi-layer hierarchical architecture as claimed in claim 1, characterized in that: Establish a multi-channel PID closed-loop feedback control module to achieve tracking control of multi-channel MOS tube gate voltage after intelligent compensation, including: The gate voltage of the MOS tube after intelligent compensation is used as the set value, and a fuzzy PID control algorithm is designed. The gate voltage of the MOS tube is collected as the feedback quantity. The program-controlled DA value is calculated by MCU based on the PID algorithm to control the gate voltage of the MOS tube, thereby realizing multi-channel tracking feedback control of the electronic load current. The three key control parameters of the PID algorithm, namely "proportional, integral, and differential", are obtained through mining using a big data-driven method.

Images