A statistical system for air heat dissipation losses of DC power electronic switches

Abstract

The present invention relates to the technical field, specifically a statistical system for air heat dissipation loss of DC power electronic switches. In the present invention, dynamic difference merging is used to establish a deviation matrix, which helps to continuously track heat loss and form high-resolution heat difference information, and the product sequence of the deviation term and the electrical signal within the cycle is input into a feedforward neural network. By using multiple rounds of weight iteration and convergence judgment of the residual sum of squares, an estimation sequence that can reflect the dynamic loss characteristics is obtained. By dividing the equipment groups through support vector machines and screening high-similarity samples through silhouette coefficients, it is possible to significantly reduce abnormal sample interference, extract representative central loss values, and construct a loss benchmark with high consistency. Combining the deviation feature set to locate the change peak, and dynamically correcting the loss results through weight reduction and reorganization estimation, the time series continuity of the estimation results and the accuracy of the thermal management closed-loop control can be significantly improved, thereby achieving accurate monitoring of the thermal load of high-power density equipment and energy efficiency optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of power loss calculation, and in particular to a system for calculating air heat dissipation losses of a DC power electronic switch. Background Art

[0002] The field of power loss calculation technology mainly focuses on the conduction loss, switching loss and static power consumption caused by power electronic devices during the conduction and shutdown processes. By collecting operating parameters such as current, voltage and temperature in real time, combined with electrothermal coupling models or empirical algorithms to estimate losses, it is widely used in thermal management and energy efficiency evaluation of high power density equipment such as converters, power modules, and DC distribution systems.

[0003] A DC power electronic switch air heat loss statistics system aims to achieve dynamic monitoring and quantitative evaluation of device heat consumption, construct device thermal load curves, optimize heat dissipation structure configurations, evaluate device energy efficiency performance under different load cycles, improve overall power density and thermal stability, and avoid performance degradation or failure caused by overheating. It aims to provide accurate thermal management basic data for power electronic systems and support closed-loop calibration of heat dissipation design and intelligent control strategies.

[0004] Existing technologies rely on electrothermal coupling models for loss estimation, which lacks specificity when processing dynamic parameter changes within a cycle. Heat input and heat dissipation are inferred only through a single data point or average value, ignoring the transient characteristics of temperature rise during equipment operation and the periodic fluctuations of air channel heat dissipation power, causing the thermal balance results to deviate from the actual operating conditions. When processing abnormal data, there is a lack of a multi-round iterative correction mechanism, and sudden deviations cannot be corrected in a timely manner. The stability and accuracy of loss estimation are insufficient. Group sample analysis uses simple mean or extreme value statistics, and does not fully distinguish between equipment structure differences and sample similarities. It is easy to include atypical samples in the benchmark value calculation, affecting the reliability of group analysis results. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a system for calculating air heat dissipation losses of a DC power electronic switch.

[0006] In order to achieve the above-mentioned object, the present invention adopts the following technical solution: A DC power electronic switch air heat dissipation loss statistics system includes:

[0007] The conduction parameter module: Based on the case temperature, voltage, and current information during the conduction cycle of the DC power device, it samples the voltage and current and marks the conduction time, calculates the RMS current and the mean voltage, processes the temperature rise, and generates a conduction thermal parameter set.

[0008] Thermal difference calculation module: based on the conduction thermal state parameter set, calculates the heat input value and the air channel heat dissipation power, performs a comparison between the input and heat dissipation heat difference, and establishes a heat loss deviation matrix;

[0009] Loss estimation module: Based on the heat loss deviation matrix, the product sequence of the deviation term and the electrical signal corresponding to each cycle is extracted. A feedforward neural network is used to perform weighted operations according to the initial weight combination and simultaneously calculate the residual sum of squares. The residual change rate is compared to determine the convergence condition and the weight path is updated round by round to obtain the periodic loss estimation sequence.

[0010] Group calibration module: Based on the periodic loss estimation sequence, a support vector machine is used to divide the equipment into groups according to the air duct structure. The mean loss and maximum deviation of each group are processed separately, and a difference matrix is ​​constructed to perform silhouette coefficient discrimination within and between groups. Samples in clusters with silhouette coefficient values ​​above the threshold are extracted and the central loss value is calculated to obtain a group consistent loss benchmark;

[0011] Error correction module: Based on the group consistent loss benchmark, it collects error sequences and identifies the maximum change point, decreases the corresponding estimation weights and reorganizes the estimation to obtain a dynamic correction loss result.

[0012] As a further solution of the present invention, the conduction parameter module includes:

[0013] The sampling and annotation submodule collects voltage signals and records the timestamp of each sampling based on the case temperature, voltage, and current information during the conduction cycle of the DC power device. It also intercepts the current signal and calibrates the initial baseline. It identifies the changing trend of the start and end points of the conduction voltage drop, matches the current rising edge with the conduction voltage segment, and establishes a corresponding relationship to create the conduction raw data.

[0014] Parameter calculation submodule: Based on the original conduction data, the current sampling sequence of each conduction interval is extracted and the square root is calculated after point-by-point accumulation. The corresponding voltage sampling value is extracted and the average value is obtained after point-by-point summation. The temperature change value before and after conduction is read and the difference is measured and processed to form a conduction thermal parameter set.

[0015] As a further solution of the present invention, the heat difference calculation module includes:

[0016] Heat calculation submodule: Based on the conduction thermal state parameter set, extract the current root mean square value and voltage mean value in each conduction cycle and perform numerical product to obtain heat input, read the internal flow velocity information and cross-sectional dimension data of the air channel to calculate the heat dissipation power, and obtain thermal balance data;

[0017] Difference matrix submodule: Based on the thermal balance data, the heat input and heat dissipation power in the corresponding period are numerically differentiated, the differences are merged in chronological order to construct a matrix index structure, and a heat loss deviation matrix is ​​generated.

[0018] As a further solution of the present invention, the loss estimation module includes:

[0019] Deviation product submodule: Based on the heat loss deviation matrix, extract the deviation items of each cycle, align them with the voltage and current sampling power array, and match them point by point. Read the deviation value and power value of the same period and perform multiplication operation according to the index. Reorganize the time series data according to the cycle number to generate the product deviation signal.

[0020] Residual convergence submodule: Based on the product deviation signal, a feedforward neural network is used to perform weighted accumulation on the signal data through the initial weight vector and synchronously collect the current residual square sum value, read the previous round of residual values ​​and perform difference comparison. If the weights have not converged, they are updated according to the difference and the weighted accumulation operation is repeated to obtain multiple rounds of residual sequences;

[0021] Estimation generation submodule: Based on the multi-round residual sequence, read the weight configuration corresponding to the last convergence round, traverse the product deviation signal and weight array of each cycle and perform fusion accumulation, verify the sequence coherence and cycle integrity, and obtain the cycle loss estimation sequence.

[0022] As a further solution of the present invention, the feedforward neural network first initializes the weight vector and applies it to the product deviation signal, performs weighted accumulation operation point by point, and simultaneously calculates the residual sum of squares between the current output and the target value, then reads the residual value obtained by the previous round of calculation, compares the current residual with the previous residual, and if the difference does not reach the preset convergence threshold, updates the weight parameter in descending order according to the difference, re-executes the weighted accumulation operation, generates a new round of output, calculates the new residual sum of squares, and iterates in a loop, updates the weight and calculates the residual in each round until the residual difference change rate is less than the threshold, determines that the weight converges, and outputs multiple rounds of residual sequences and converged weight configurations.

[0023] As a further solution of the present invention, the feedforward neural network is according to the formula:

[0024] ;

[0025] in: Indicates the The weight vector used in the power loss prediction model in the residual convergence module after rounds of iterations, Indicates the The current weight vector at the round iteration, Represents the basic learning rate, which controls the step size of weight update in each iteration. Represents the learning rate attenuation factor, combined with the current round Adjust the step size decreasing speed to improve the residual convergence stability, Indicates the current iteration round, Indicates the number of batch signal samples participating in this round of training, It represents the partial derivative of the residual sum of squares to the current weight, reflecting the sensitivity of the loss fitting model to the parameters. Indicates the The sum of squares of the residuals between the predicted value and the actual air heat loss value in the round iteration, Indicates the The sample in The actual heat dissipation loss value in the round iteration is measured by the thermistor and wind speed sensor. Indicates that the model is effective for the The predicted value of heat dissipation loss of samples, represents the momentum factor coefficient, Indicates the change in weight between the current and previous rounds, represents the weight decay coefficient, Indicates the total number of batch training samples used for model training, represents the weight decay term.

[0026] As a further solution of the present invention, the group calibration module includes:

[0027] Equipment grouping submodule: Based on the periodic loss estimation sequence, a support vector machine is used to read the air duct structure label of each device and assign it to the corresponding subgroup according to the label value. The sequence within the group is traversed and the periodic loss data dimension is extracted to obtain a structure grouping sample set;

[0028] Difference construction submodule: based on the structure grouping sample set, respectively calculate the loss mean and range of each group and calculate the difference between groups pair by group index, merge all difference results into a matrix structure and perform normalization processing to generate a group difference matrix;

[0029] Benchmark extraction submodule: Based on the group difference matrix, extract the distance vector of each group of samples and calculate the silhouette coefficient value, perform center value statistics on the sample set with coefficients higher than the threshold and summarize the center loss data of each cluster to obtain the group consistent loss benchmark.

[0030] As a further solution of the present invention, the support vector machine first reads the air duct structure label corresponding to each device, matches the periodic loss estimation sequence with the label data, establishes a sample feature matrix, and each sample contains the periodic loss data dimension and the structure label. Then, a hyperplane is constructed, the distance from each sample point to the hyperplane is calculated, and the support vector point is determined to form a classification decision function. Based on the decision function, the samples are labeled and classified and assigned to corresponding subgroups. After the classification is completed, each subgroup sample sequence is traversed, the periodic loss data dimension is extracted and sorted, and a structure grouping sample set is output.

[0031] As a further solution of the present invention, the support vector machine is according to the formula: ;

[0032] in: Indicates the The structural label category value corresponding to the support vector, Indicates that the Gaussian kernel function is used to calculate the similarity between the vector to be classified and the support vector. Represents the current cycle loss feature vector to be classified, Indicates the The eigenvectors corresponding to the support vectors are represents the Lagrange multiplier coefficient used to construct the support vector machine decision function, represents the width parameter of the Gaussian kernel function, Indicates the The support vector corresponds to the complexity coefficient of the air duct structure of the equipment represents the bias term, represents the subgroup category imbalance weight coefficient, represents the current subgroup entropy value, represents the regularization penalty coefficient, represents the total number of selected support vectors, Represents the final calculated value of the structure grouping decision function.

[0033] As a further solution of the present invention, the error correction module includes:

[0034] Error collection submodule: Based on the group consistent loss benchmark, it collects the loss deviation value sequence of each cycle, removes outliers through filtering and smoothing, traverses adjacent data points to calculate the differential amplitude, locates the timestamp corresponding to the differential value peak and marks it, extracts deviation samples before and after the time period and organizes the data dimensions to establish a deviation feature set;

[0035] Weight reorganization submodule: Based on the deviation feature set, read the estimated weight corresponding to each sample, reduce and normalize the weight of the change point sample in sequence, splice the weighted estimated value sequence in descending order, summarize the weighted results at each moment and update the estimation matrix to obtain the dynamic correction loss result.

[0036] Compared with the prior art, the advantages and positive effects of the present invention are:

[0037] In the present invention, by collecting and synchronously annotating temperature, voltage, and current signals during the conduction cycle, combined with precise calculations of the current root mean square value, voltage mean, and temperature rise amplitude, a data set reflecting the thermal state characteristics can be generated, providing a basis for subsequent comparison of heat input and heat dissipation power.

[0038] In the present invention, dynamic difference merging is used to establish a deviation matrix, which helps to continuously track heat loss and form high-resolution heat difference information. The product sequence of the deviation term and the electrical signal within the cycle is input into the feedforward neural network. Through multiple rounds of weight iteration and convergence judgment of the residual sum of squares, an estimation sequence that can reflect the dynamic loss characteristics is obtained, which makes the loss calculation process have adaptive adjustment capabilities and strong anti-interference capabilities.

[0039] In this paper, by using a support vector machine to divide equipment groups and screening highly similar samples using the silhouette coefficient, we can significantly reduce the interference of abnormal samples, extract representative central loss values, and establish a highly consistent loss benchmark. By combining a deviation feature set to locate peak values ​​and dynamically correcting loss results through weight reduction and recombination estimation, we can significantly improve the temporal continuity of the estimation results and the accuracy of the thermal management closed-loop control, thereby achieving precise monitoring of the thermal load and energy efficiency optimization of high-power density equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a system flow chart of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0042] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined. Example

[0043] See also Figure 1 The present invention provides a technical solution: a DC power electronic switch air heat dissipation loss statistics system comprising:

[0044] The conduction parameter module: Based on the case temperature, voltage, and current information during the conduction cycle of the DC power device, it samples the voltage and current and marks the conduction time, calculates the RMS current and the mean voltage, processes the temperature rise, and generates a conduction thermal parameter set.

[0045] Thermal difference calculation module: Based on the conduction thermal state parameter set, it calculates the heat input value and the heat dissipation power of the air channel, performs a comparison between the input and heat dissipation heat difference, and establishes a heat loss deviation matrix;

[0046] Loss estimation module: Based on the heat loss deviation matrix, the product sequence of the deviation term and the electrical signal corresponding to each cycle is extracted. A feedforward neural network is used to perform weighted operations based on the initial weight combination and simultaneously calculate the sum of squared residuals. The residual change rate is compared to determine the convergence condition and the weight path is updated round by round to obtain the periodic loss estimation sequence.

[0047] Group calibration module: Based on the periodic loss estimation sequence, a support vector machine is used to divide equipment into groups according to the air duct structure. The mean loss and maximum deviation of each group are processed separately. A difference matrix is ​​constructed and the silhouette coefficient is judged within and between groups. Samples in clusters with silhouette coefficient values ​​above the threshold are extracted and the central loss value is calculated to obtain a group consistent loss benchmark.

[0048] Error correction module: Based on the group consensus loss benchmark, it collects error sequences and identifies the maximum change points, decreases the corresponding estimation weights and reorganizes the estimates to obtain dynamic correction loss results.

[0049] The conduction parameter module includes:

[0050] The sampling and annotation submodule collects voltage signals and records the timestamp of each sampling based on the case temperature, voltage, and current information during the conduction cycle of the DC power device. It also intercepts the current signal and calibrates the initial baseline. It identifies the changing trend of the start and end points of the conduction voltage drop, matches the current rising edge with the conduction voltage segment, and establishes a corresponding relationship to create the conduction raw data.

[0051] Parameter calculation submodule: Based on the original conduction data, the current sampling sequence of each conduction interval is extracted and the square root is calculated after point-by-point accumulation. The corresponding voltage sampling value is extracted and averaged after point-by-point summation. The temperature change value before and after conduction is read and the difference is calculated to form the conduction thermal parameter set.

[0052] Sampling and marking submodule: Based on the shell temperature, voltage and current information during the conduction period of the DC power device, a fixed-frequency clock-triggered sampling mechanism is used to sample the voltage signal. The sampling adopts the differential amplifier front channel input and combines the sampling and holding module to stabilize the input signal amplitude. The analog-to-digital converter is driven by the sampling clock signal to obtain the voltage value point by point, and the timestamp recording function is set in the sampling control register to synchronize the sampling data with the absolute time. The static current signal window baseline calibration scheme is used to intercept the current signal of a fixed period before the conduction to establish a baseline, set the sampling window length and calculate the interval signal average value as the zero-bias reference line, and then the current signal is calibrated during the sampling process. Baseline correction is performed, and the trend of the start and end points of the on-state voltage drop is identified using the differential voltage drop trend identification method. The full-cycle voltage sampling sequence is subjected to sliding differentiation, and the critical points of the falling and rising segments are identified based on the set voltage change threshold. The current rising edge and the on-state voltage segment are matched using a dual-threshold edge interval matching method. The current signal is differentially processed to obtain the starting point of the rising edge. Pairing is performed based on time proximity and overlap of change trends, and finally a one-to-one mapping relationship between the current rising edge and the on-state voltage segment is established. The on-state raw data including the voltage sampling point sequence, current reference calibration value, on-state voltage start and end point index, current rising edge index, and pairing index is generated.

[0053] Parameter calculation submodule: Based on the original conduction data, the conduction interval segmented mean square current calculation scheme is used to square and accumulate each current sampling sequence point by point, and then normalize it with the number of points to obtain the current root mean square value. The voltage sampling value corresponding to the above current sequence is extracted by the voltage sampling segment average extraction method. Based on the equal interval sampling condition, the interval sampling points are summed and the average value is calculated according to the number of samples as the average voltage. The conduction temperature difference estimation scheme is used to read the temperature measurement values ​​before and after the start of the conduction interval. The numerical difference operation is performed and combined with the measurement time difference to calculate the temperature rise per unit time. Finally, the current root mean square value, voltage average value, and temperature rise estimation value are assembled into a conduction thermal parameter set.

[0054] The thermal difference calculation module includes:

[0055] Heat calculation submodule: Based on the conduction thermal state parameter set, it extracts the RMS current and mean voltage in each conduction cycle and multiplies them to obtain the heat input. It also reads the internal flow velocity information and cross-sectional dimension data of the air channel to calculate the heat dissipation power and obtain the thermal balance data.

[0056] Difference matrix submodule: Based on the thermal balance data, the heat input and heat dissipation power in the corresponding period are numerically differentiated, the differences are merged in chronological order to construct a matrix index structure, and a heat loss deviation matrix is ​​generated;

[0057] Heat calculation submodule: Based on the conduction thermal state parameter set, the heat input cumulative product processing method is used to match the current root mean square value and the voltage mean extracted in each conduction cycle one by one. The current and voltage values ​​are read in sequence using the data buffer. The current value is used as the left factor and the voltage value is used as the right factor to input into the multiplication operation unit. The product operation is completed according to the corresponding conduction cycle index. The product result is stored in the heat input temporary register group, and then written into the heat input data area to complete the heat input construction. The heat dissipation power numerical estimation method is used to read the flow rate sensor sampling result set inside the air channel. Sequentially extract channel cross-sectional width and height dimension data, establish a cross-sectional area calculation factor based on the rectangular channel assumption, calculate the volume flow rate value by matching the area and flow rate values, read the fixed specific heat capacity of the heat exchange medium and perform the multiplication and accumulation, combine the temperature difference value measured per unit time as the heat numerator, complete the heat dissipation power processing according to the ratio calculation method and write the result to the heat dissipation power recording area, read the heat input value and heat dissipation power value corresponding to the conduction cycle index, perform the difference processing operation, complete the same-cycle comparison and difference calculation of the two sets of data, and write them to the heat balance data storage area to generate heat balance data;

[0058] Difference matrix submodule: Based on the thermal balance data, the difference is arranged in time sequence and merged to construct a matrix method. The heat input value and heat dissipation power value in each cycle are read in the order of the conduction cycle time index. The difference processing is performed on the two data and the results are written to the difference data buffer. Then all difference samples are read in sequence and the initial matrix structure arranged in rows is constructed. A unique index identifier is set for each difference data entry, and the time series list is called to match the corresponding row position in index order. The complete arrangement operation of the difference matrix is ​​performed. After the matrix mapping relationship of all conduction cycle data is established, it is written into the matrix storage area to finally generate the heat loss deviation matrix

[0059] The loss estimation module includes:

[0060] Deviation product submodule: Based on the heat loss deviation matrix, it extracts the deviation items of each cycle, aligns them with the voltage and current sampling power array, and matches them point by point. It reads the deviation value and power value of the same period and performs multiplication operation according to the index. It reorganizes the time series data according to the cycle number to generate the product deviation signal.

[0061] Residual convergence submodule: Based on the product deviation signal, a feedforward neural network is used to perform weighted accumulation on the signal data through the initial weight vector and synchronously collect the current residual sum of squares. The residual value of the previous round is read and the difference is compared. If the weight has not converged, it is updated according to the difference and the weighted accumulation operation is repeated to obtain multiple rounds of residual sequences.

[0062] Estimation generation submodule: Based on multiple rounds of residual sequences, read the weight configuration corresponding to the last convergence round, traverse the product deviation signal and weight array of each cycle and perform fusion accumulation, verify the sequence coherence and cycle integrity, and obtain the cycle loss estimation sequence;

[0063] Deviation product submodule: Based on the heat loss deviation matrix, the data index alignment product operation method is used to extract the deviation value corresponding to each conduction cycle. The data with the same cycle number in the voltage and current sampling power array are read in time series. The deviation value and the power value are matched point by point. A one-time multiplication operation is performed on each set of matching results. The calculated value is stored in the product buffer area. The product data is arranged in the order of cycle number, and a linear time index structure is established. The data content is written to the product signal storage area with a fixed address, and finally a product deviation signal is generated.

[0064] Residual convergence submodule: Based on the product deviation signal, a static structure feedforward neural network weighted processing method is adopted to initialize the fixed layer network structure and load the first set of weight vector values. The product deviation signal samples corresponding to all cycles are read in sequence as input node data. Each signal value is multiplied by the corresponding weight one by one in a predetermined order, and then weighted accumulation is performed in sequence. The weighted accumulation result is temporarily stored in the output area, and the square operation is performed to obtain the residual square result and stored in the residual record area. The total residual square value of the previous weight corresponding round is read and compared with the current round residual value by direct numerical difference. If the difference is higher than the convergence threshold, the original weight array is corrected by the proportional factor according to the defined difference adjustment ratio, and the weighted accumulation and residual acquisition processing are re-executed with the new weight. The same round of neural network feedforward operation and residual convergence processing process are iterated until the residual change value is lower than the threshold, and finally a multi-round residual sequence is generated;

[0065] Estimation generation submodule: Based on multiple rounds of residual sequences, a weight-guided fusion calculation method is adopted to read the weight configuration data corresponding to the last round of residual sequence that has completed convergence, establish a one-to-one mapping relationship between the cycle number and the weight index, read the product deviation signal of each cycle and the weight value at the corresponding position in chronological order, perform multiplication operation on the two values ​​and write the result into the accumulation register area, then superimpose and summarize the corresponding results of all cycles in sequence to complete a round of fusion operation, and then check the continuity of the cycle number in sequence and confirm that the cycle sampling data is complete, and finally generate a cycle loss estimation sequence.

[0066] Feedforward neural network, first, initialize the weight vector and apply it to the product deviation signal, perform weighted accumulation operation point by point, and synchronously calculate the residual sum of squares between the current output and the target value. Then read the residual value calculated in the previous round, compare the current residual with the previous residual, and if the difference does not reach the preset convergence threshold, update the weight parameter in descending order according to the difference, re-execute the weighted accumulation operation, generate a new round of output and calculate the new residual sum of squares, and iterate in a loop. In each round, update the weight and calculate the residual until the residual difference change rate is less than the threshold, determine that the weight converges, and output multiple rounds of residual sequences and converged weight configurations.

[0067] Feedforward neural network, according to the formula: ;

[0068] in: Indicates the The weight vector used in the power loss prediction model in the residual convergence module after rounds of iterations, Indicates the The current weight vector at the round iteration, Represents the basic learning rate, which controls the step size of weight update in each iteration. Represents the learning rate attenuation factor, combined with the current round Adjust the step size decreasing speed to improve the residual convergence stability, Indicates the current iteration round, Indicates the number of batch signal samples participating in this round of training, It represents the partial derivative of the residual sum of squares to the current weight, reflecting the sensitivity of the loss fitting model to the parameters. Indicates the The sum of squares of the residuals between the predicted value and the actual air heat loss value in the round iteration, Indicates the The sample in The actual heat dissipation loss value in the round iteration is measured by the thermistor and wind speed sensor. Indicates that the model is effective for the The predicted value of heat dissipation loss of samples, represents the momentum factor coefficient, Indicates the change in weight between the current and previous rounds, represents the weight decay coefficient, Indicates the total number of batch training samples used for model training, represents the weight decay term;

[0069] Execution process: First, collect the heat dissipation related signal data of the power electronic switch under different load conditions and form a batch input vector. The batch size is set to m to improve the statistical stability of each round of model update. Then, the weight vector corresponding to the current iteration round t is Perform weighted operations on the input signal to generate a heat loss prediction value for each sample and compared with the sensor measured value Calculate the residual sum of squares to construct the loss function, and then use the partial derivative of the weight vector with respect to the residual function Perform gradient calculations while setting the base learning rate Combined learning rate decay factor Calculate the normalization factor with round t , used to suppress high-round fluctuations, and then introduced the momentum mechanism through the momentum coefficient Calculate the difference between the current weight change direction and the previous round weight to form a direction compensation term , enhance the stability of the convergence path, and introduce the weight decay term To prevent overfitting and constrain model complexity, the three update terms are linearly synthesized and subtracted from the current weight vector to obtain the next round of update weights. The whole process is executed cyclically until the loss prediction model converges to the minimum residual point, thereby completing the statistical modeling of the heat dissipation loss of DC power electronic switches under natural air cooling conditions.

[0070] The Population Calibration Module includes:

[0071] Equipment grouping submodule: Based on the periodic loss estimation sequence, a support vector machine is used to read the air duct structure label of each device and assign it to the corresponding subgroup according to the label value. The sequence within the group is traversed and the periodic loss data dimension is extracted to obtain a structural grouping sample set;

[0072] Difference construction submodule: Based on the structure grouping sample set, the loss mean and range of each group are calculated respectively, and the difference between groups is calculated pair by pair according to the group index. All difference results are merged into a matrix structure and normalized to generate a group difference matrix;

[0073] Benchmark extraction submodule: Based on the group difference matrix, extract the distance vector of each sample group and calculate the silhouette coefficient value. Perform center value statistics on the sample set with the coefficient higher than the threshold and summarize the center loss data of each cluster to obtain the group consistent loss benchmark;

[0074] Equipment grouping submodule: Based on the periodic loss estimation sequence, a support vector machine grouping scheme is adopted. The linear kernel function is called with the parameter items set to the kernel function type linear, the tolerance parameter C to 1.0, and the classification interval maximization parameter epsilon to 0.01. The air duct structure label is bound to each device, and the index mapping is extracted from the structure label table based on the label value. The equipment samples corresponding to different label values ​​are mapped to the corresponding subgroup memory area. After the mapping is completed, all periodic loss data in each group of samples are traversed, and the loss value field is extracted according to the equipment number and period order. After removing the non-numeric field, the sample input vector is generated in the sample group unit. All samples in the group are summarized in a unified format and stored in the structured sample pool, and finally a structured grouping sample set is generated;

[0075] Difference construction submodule: Based on the structured grouped sample set, the group mean and range calculation method is adopted to perform a single-channel mean calculation operation on all sample period loss values ​​in each group. The overall mean of the period loss value sequence is calculated using a sliding window method and stored in the group mean buffer. Subsequently, the difference between the maximum and minimum values ​​is extracted in the same group to obtain the range result. Index pairs are constructed in pairwise order according to the group index sequence. For each index pair, the mean and range of the corresponding group are extracted and the difference between the two groups is calculated. All inter-group difference results are written into a matrix structure buffer with consistent row and column numbers. Finally, all elements in the matrix are normalized and each element is scaled to between zero and one using a linear minimum-maximum mapping method to finally generate a group difference matrix.

[0076] Benchmark extraction submodule: Based on the group difference matrix, the sample silhouette coefficient evaluation and central statistics method are used to extract the average distance vector of each group sample in the matrix, construct local compactness and external dispersion indicators, and use the general formula of the silhouette coefficient to calculate the ratio of the difference between the average distance within the group to which each sample belongs and the minimum average distance of the nearest group to the maximum value to form a sample silhouette coefficient value vector. The sample set with a silhouette coefficient greater than the threshold value of 0.5 is selected and its position in the original periodic loss estimation sequence is extracted. Each group of samples is grouped according to the period number, and the median extraction operation is performed on the periodic loss value of each group of samples and summarized into the central statistical area. The representative value of the center of each cluster is stored in the order of the group label, and finally a group consistent loss benchmark is generated.

[0077] The support vector machine first reads the air duct structure label corresponding to each device, matches the cycle loss estimation sequence with the label data, and establishes a sample feature matrix. Each sample contains the cycle loss data dimension and structure label. Then, a hyperplane is constructed, the distance from each sample point to the hyperplane is calculated, and the support vector points are determined to form a classification decision function. Based on the decision function, the samples are labeled and assigned to the corresponding subgroups. After the classification is completed, the sample sequence of each subgroup is traversed, the cycle loss data dimension is extracted and sorted, and the structure grouping sample set is output.

[0078] Support vector machine, according to the formula: ;

[0079] in: Indicates the The structural label category value corresponding to the support vector, Indicates that the Gaussian kernel function is used to calculate the similarity between the vector to be classified and the support vector. Represents the current cycle loss feature vector to be classified, Indicates the The eigenvectors corresponding to the support vectors are represents the Lagrange multiplier coefficient used to construct the support vector machine decision function, represents the width parameter of the Gaussian kernel function, Indicates the The support vector corresponds to the complexity coefficient of the air duct structure of the equipment represents the bias term, represents the subgroup category imbalance weight coefficient, represents the current subgroup entropy value, represents the regularization penalty coefficient, represents the total number of selected support vectors, Represents the final calculated value of the structure grouping decision function;

[0080] Execution process: First, extract the heat loss characteristics in multiple working cycles from the cycle loss estimation module and construct the feature vector , including statistical indicators of duct thermal resistance, current density, conduction loss change rate, and surface wind speed, and then load the trained support vector set And call in support vector labels , Lagrange multipliers and duct structure complexity coefficient ,Then, the Gaussian kernel function is used to calculate the similarity response between each support vector and the current device vector, and the similarity value is combined with Together they constitute the weighted cumulative term used to characterize the attribution tendency of different structural groups, and then introduce the bias term Numerical correction of model output and introduction of class imbalance calibration term To adapt to the changes in the number of equipment samples and distribution entropy in different air duct subgroups, the regularization term Perform constraint correction to prevent model overfitting and combine all the above items to generate structure grouping judgment values , establish a structural grouping sample set for subsequent heat dissipation loss statistical analysis and grouping modeling.

[0081] The error correction module includes:

[0082] Error collection submodule: Based on the group consensus loss benchmark, it collects the loss deviation value sequence for each cycle, removes outliers through filtering and smoothing, traverses adjacent data points to calculate the differential amplitude, locates the corresponding timestamp of the differential value peak and marks it, extracts deviation samples before and after the time period, organizes the data dimensions, and establishes a deviation feature set;

[0083] Weight Recombination Submodule: Based on the deviation feature set, read the estimated weight corresponding to each sample, decrease and normalize the weight of the change point samples in sequence, splice the weighted estimated value sequence in descending order, summarize the weighted results at each moment and update the estimation matrix to obtain the dynamic correction loss result;

[0084] Error acquisition submodule: Based on the group consistent loss benchmark, a sliding window median filtering scheme is adopted to read the loss deviation value sequence of each period in time series. The window size is set to 5 samples. After sorting the deviation values ​​in each window, the middle value is selected to replace the center point to smooth the sequence. After traversing the smoothed sequence, the difference calculation is performed on two adjacent points. The difference calculation method is the current sample minus the previous sample. The peak recognition threshold is set to 0.05. The time point corresponding to the difference peak exceeding the threshold is recorded as the change mark. 3 samples before and after each change point are extracted to construct the feature interception interval. The 7 samples before and after each interval are sorted and stored in a four-dimensional format according to the deviation amplitude, time interval, peak position, and sample sequence number to generate a deviation feature set.

[0085] Weight reorganization submodule: Based on the deviation feature set, a decreasing weight allocation scheme is adopted. The initial estimated weight vector corresponding to each sample is read, and the decreasing factor is set to 0.9. For each change point and subsequent sample in the sequence, the original weight is multiplied by the decreasing factor to update the weight value. After the update is completed, the weights of all samples are summed and divided by the sum item by item to perform normalization processing. The corresponding product deviation signal and weight value are spliced ​​in the order of the normalized weights. The product operation is performed in each cycle and all product values ​​are accumulated and written into the corresponding position of the estimation matrix in sequence to finally form a dynamic correction loss result.

[0086] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A system for calculating air heat dissipation loss of a DC power electronic switch, characterized in that: The system comprises: The conduction parameter module: Based on the case temperature, voltage, and current information during the conduction cycle of the DC power device, it samples the voltage and current and marks the conduction time, calculates the RMS current and the mean voltage, processes the temperature rise, and generates a conduction thermal parameter set. Thermal difference calculation module: based on the conduction thermal state parameter set, calculates the heat input value and the air channel heat dissipation power, performs a comparison between the input and heat dissipation heat difference, and establishes a heat loss deviation matrix; Loss estimation module: Based on the heat loss deviation matrix, the product sequence of the deviation term and the electrical signal corresponding to each cycle is extracted. A feedforward neural network is used to perform weighted operations according to the initial weight combination and simultaneously calculate the residual sum of squares. The residual change rate is compared to determine the convergence condition and the weight path is updated round by round to obtain the periodic loss estimation sequence. Group calibration module: Based on the periodic loss estimation sequence, a support vector machine is used to divide the equipment into groups according to the air duct structure. The mean loss and maximum deviation of each group are processed separately, and a difference matrix is ​​constructed to perform silhouette coefficient discrimination within and between groups. Samples in clusters with silhouette coefficient values ​​above the threshold are extracted and the central loss value is calculated to obtain a group consistent loss benchmark; Error correction module: Based on the group consistent loss benchmark, it collects error sequences and identifies the maximum change point, decreases the corresponding estimation weights and reorganizes the estimation to obtain a dynamic correction loss result.

2. The DC power electronic switch air heat dissipation loss statistics system according to claim 1, characterized in that: The conduction parameter module includes: The sampling and annotation submodule collects voltage signals and records the timestamp of each sampling based on the case temperature, voltage, and current information during the conduction cycle of the DC power device. It also intercepts the current signal and calibrates the initial baseline. It identifies the changing trend of the start and end points of the conduction voltage drop, matches the current rising edge with the conduction voltage segment, and establishes a corresponding relationship to create the conduction raw data. Parameter calculation submodule: Based on the original conduction data, the current sampling sequence of each conduction interval is extracted and the square root is calculated after point-by-point accumulation. The corresponding voltage sampling value is extracted and the average value is obtained after point-by-point summation. The temperature change value before and after conduction is read and the difference is measured and processed to form a conduction thermal parameter set.

3. The DC power electronic switch air heat dissipation loss statistics system according to claim 1, characterized in that: The heat difference calculation module includes: Heat calculation submodule: Based on the conduction thermal state parameter set, extract the current root mean square value and voltage mean value in each conduction cycle and perform numerical product to obtain heat input, read the internal flow velocity information and cross-sectional dimension data of the air channel to calculate the heat dissipation power, and obtain thermal balance data; Difference matrix submodule: Based on the thermal balance data, the heat input and heat dissipation power in the corresponding period are numerically differentiated, the differences are merged in chronological order to construct a matrix index structure, and a heat loss deviation matrix is ​​generated.

4. The DC power electronic switch air heat dissipation loss statistics system according to claim 1, characterized in that: The loss estimation module includes: Deviation product submodule: Based on the heat loss deviation matrix, extract the deviation items of each cycle, align them with the voltage and current sampling power array, and match them point by point. Read the deviation value and power value of the same period and perform multiplication operation according to the index. Reorganize the time series data according to the cycle number to generate the product deviation signal. Residual convergence submodule: Based on the product deviation signal, a feedforward neural network is used to perform weighted accumulation on the signal data through the initial weight vector and synchronously collect the current residual square sum value, read the previous round of residual values ​​and perform difference comparison. If the weights have not converged, they are updated according to the difference and the weighted accumulation operation is repeated to obtain multiple rounds of residual sequences; Estimation generation submodule: Based on the multi-round residual sequence, read the weight configuration corresponding to the last convergence round, traverse the product deviation signal and weight array of each cycle and perform fusion accumulation, verify the sequence coherence and cycle integrity, and obtain the cycle loss estimation sequence.

5. The DC power electronic switch air heat dissipation loss statistics system according to claim 4, characterized in that: The feedforward neural network first initializes the weight vector and applies it to the product deviation signal, performs weighted accumulation operation point by point, and simultaneously calculates the residual sum of squares between the current output and the target value. Then, the residual value obtained by the previous round of calculation is read, and the current residual is compared with the previous residual. If the difference does not reach the preset convergence threshold, the weight parameter is updated in descending order according to the difference, and the weighted accumulation operation is re-executed to generate a new round of output and calculate a new residual sum of squares and iterate in a loop. The weight is updated and the residual is calculated in each round until the residual difference change rate is less than the threshold. The weight convergence is determined, and a multi-round residual sequence and converged weight configuration are output.

6. The DC power electronic switch air heat loss statistics system according to claim 4, characterized in that: The feedforward neural network is based on the formula: ； in: Indicates the The weight vector used in the power loss prediction model in the residual convergence module after rounds of iterations, Indicates the The current weight vector at the round iteration, Represents the basic learning rate, which controls the step size of weight update in each iteration. Represents the learning rate attenuation factor, combined with the current round Adjust the step size decreasing speed to improve the residual convergence stability, Indicates the current iteration round, Indicates the number of batch signal samples participating in this round of training, Represents the partial derivative of the residual sum of squares to the current weight, reflecting the sensitivity of the loss fitting model to the parameters. Indicates the The sum of squares of the residuals between the predicted value and the actual air heat loss value in the round iteration, Indicates the The sample in The actual heat dissipation loss value in the round iteration is measured by the thermistor and wind speed sensor. Indicates that the model is effective for the The predicted value of heat dissipation loss of samples, represents the momentum factor coefficient, Indicates the change in weight between the current and previous rounds, represents the weight decay coefficient, Indicates the total number of batch training samples used for model training, represents the weight decay term.

7. The DC power electronic switch air heat dissipation loss statistics system according to claim 1, characterized in that: The population calibration module includes: Equipment grouping submodule: Based on the periodic loss estimation sequence, a support vector machine is used to read the air duct structure label of each device and assign it to the corresponding subgroup according to the label value. The sequence within the group is traversed and the periodic loss data dimension is extracted to obtain a structure grouping sample set; Difference construction submodule: based on the structure grouping sample set, respectively calculate the loss mean and range of each group and calculate the difference between groups pair by group index, merge all difference results into a matrix structure and perform normalization processing to generate a group difference matrix; Benchmark extraction submodule: Based on the group difference matrix, extract the distance vector of each group of samples and calculate the silhouette coefficient value, perform center value statistics on the sample set with coefficients higher than the threshold and summarize the center loss data of each cluster to obtain the group consistent loss benchmark.

8. The system for calculating air heat dissipation loss of a DC power electronic switch according to claim 7, characterized in that: The support vector machine first reads the air duct structure label corresponding to each device, matches the periodic loss estimation sequence with the label data, and establishes a sample feature matrix. Each sample contains the periodic loss data dimension and the structure label. Then, a hyperplane is constructed, the distance from each sample point to the hyperplane is calculated, and the support vector point is determined to form a classification decision function. Based on the decision function, the samples are labeled and classified and assigned to corresponding subgroups. After the classification is completed, each subgroup sample sequence is traversed, the periodic loss data dimension is extracted and sorted, and a structure grouping sample set is output.

9. The DC power electronic switch air heat dissipation loss statistics system according to claim 7, characterized in that: The support vector machine is based on the formula: ； in: Indicates the The structural label category value corresponding to the support vector, Indicates that the Gaussian kernel function is used to calculate the similarity between the vector to be classified and the support vector. Represents the current cycle loss feature vector to be classified, Indicates the The eigenvectors corresponding to the support vectors are represents the Lagrange multiplier coefficient used to construct the support vector machine decision function, represents the width parameter of the Gaussian kernel function, Indicates the The support vector corresponds to the complexity coefficient of the air duct structure of the equipment represents the bias term, represents the subgroup category imbalance weight coefficient, represents the current subgroup entropy value, represents the regularization penalty coefficient, represents the total number of selected support vectors, Represents the final calculated value of the structure grouping decision function.

10. The system for calculating air heat dissipation loss of a DC power electronic switch according to claim 1, characterized in that: The error correction module includes: Error collection submodule: Based on the group consistent loss benchmark, it collects the loss deviation value sequence of each cycle, removes outliers through filtering and smoothing, traverses adjacent data points to calculate the differential amplitude, locates the timestamp corresponding to the differential value peak and marks it, extracts deviation samples before and after the time period and organizes the data dimensions to establish a deviation feature set; Weight reorganization submodule: Based on the deviation feature set, read the estimated weight corresponding to each sample, reduce and normalize the weight of the change point sample in sequence, splice the weighted estimated value sequence in descending order, summarize the weighted results at each moment and update the estimation matrix to obtain the dynamic correction loss result.

Images