Intelligent power device safe working area test equipment and system for realizing data judgment

Abstract

The invention provides an intelligent power device safe working area testing device and system for realizing data judgment, and relates to the technical field of power device testing, and the intelligent power device safe working area testing device comprises a first testing module for determining a current boundary value, and a second testing module for obtaining characteristic data, the data frame selection module is used for determining normal state feature data based on a recursive expansion density clustering algorithm and an expert scoring method introducing random numbers, the data expansion module is used for expanding data based on the normal state feature data, and the data simulation module is used for capturing the relation between the feature data and outputting a safe working area in combination with a safe working standard. According to the method, abnormal values are eliminated through a recursive expansion density clustering algorithm and an expert scoring method for introducing random numbers, so that the accuracy of subsequent analysis of the safe working area is improved, the actual test workload is reduced through cooperative setting of the adversarial network and the data simulation module, and the test efficiency of the safe working area is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power device testing, in particular to an intelligent power device safe working area testing device and system for realizing data judgment. BACKGROUND

[0002] In the field of inverters in the power electronics field, such as frequency converters, photovoltaic, electric vehicles, etc., power devices IGBT are the core components of inverters, known as the heart of inverters. If the power device fails, the inverter cannot work normally, and the electric vehicle will also lose power. If this happens while driving at high speed, the consequences are unpredictable, and there is even a risk of car damage and death.

[0003] Therefore, power device manufacturers have strict failure rate requirements for various components, generally aiming to control at 50ppm (i.e. 50 parts per million), and in some special application scenarios, even 30ppm (i.e. 30 parts per million). Since the production process cannot be 100% safe, the importance of out-of-factory testing becomes even greater. In order to test and age during the out-of-factory stage, problems are intercepted in the factory, and various device manufacturers have been working hard.

[0004] In the prior art, a kind of IGBT intermittent life test method based on simulation modeling and short-time test (classification number G06F) with publication number "CN107861040A" includes the following steps: step one: determine the heat dissipation condition;Step two: determine the parameter control method;Step three: determine the failure criterion;Step four: determine the device power size and junction temperature control range;Step five: safe working area and maximum allowable junction temperature analysis;Step six: implement short-time test;Step seven: simulation modeling analysis;Step eight: optimization of intermittent life test scheme. The prior art comprehensively considers factors such as heat dissipation conditions, parameter control methods, power size, temperature range, safe working area and failure mechanism, pre-selects several sets of intermittent life test schemes, obtains the temperature rise and fall time of the device in a single cycle through short-time power cycle test, obtains the power cycle number before device failure using simulation method, estimates the actual power cycle test time of each pre-selected scheme, and then optimizes the best intermittent life test scheme, achieving the effect of safe working area testing of power devices.

[0005] However, the prior art still has great defects, such as: during long-time high-intensity performance testing of power devices, short-time failures are prone to occur, resulting in distorted data collection, the accuracy of the safe working area determined based on the distorted data is poor, and the prior art needs to perform a large number of test experiments during safe working area testing, which has the problem of slow efficiency.

[0006] The above information disclosed in the Background section is only for strengthening the understanding of the background of the present disclosure, and thus it can include information that does not constitute the prior art known to those of ordinary skill in the art. SUMMARY

[0007] The purpose of the present application is to provide an intelligent power device safe working area test equipment and system to realize data judgment, so as to solve the problems raised in the background.

[0008] To achieve the above purpose, the present application provides the following technical solutions:

[0009] An intelligent power device safe working area test equipment to realize data judgment, comprising:

[0010] A first test module for applying a test condition with increasing current to the power device under test, determining a current demarcation value based on the temperature of the power device under test under each test condition;

[0011] A second test module for applying a test condition with increasing current and constant voltage to the power device under test to obtain characteristic data of the power device under test under each test condition, the characteristic data including current value, voltage value, temperature, accuracy rate and response time when performing data judgment, and the amplitude of the increasing current is determined based on the current demarcation value;

[0012] A data frame selection module for clustering and analyzing the characteristic data based on the recursive expansion density clustering algorithm with temperature-current as the clustering reference, determining multiple clusters and isolated characteristic data, introducing random numbers based on expert scoring method, analyzing isolated characteristic data to determine fault characteristic data and frame selection characteristic data, and frame selecting the clusters based on the frame selection characteristic data to determine normal characteristic data;

[0013] A data augmentation module for generating multiple sets of simulated characteristic data based on the data augmentation module constructed based on the adversarial network, and the data augmentation module is trained based on the normal characteristic data;

[0014] A data simulation module, the input of the data simulation module constructed based on the deep neural network is voltage value and current value, and the output is temperature, accuracy rate and response time, the data simulation module is trained based on the simulated characteristic data and the normal characteristic data, and based on the safe working standard, the safe working area under each voltage is output.

[0015] Further, the first test module comprises a power source for applying different currents to the power device under test, a temperature sensor for detecting the surface temperature of the power device under test, and a control software for controlling the power source to output different currents;

[0016] The determination logic of the current demarcation value is that after the power device under test is stabilized, the temperature of the power device under test is collected, if the temperature exceeds a preset temperature threshold, the current value after the current increment is determined as the current demarcation value, otherwise, it is determined that it is not the current demarcation value.

[0017] Further, the second test module comprises a power source for applying different currents to the power device under test, a temperature sensor for detecting the surface temperature of the power device under test, control software for controlling the power source to output different currents, and an adjustable load for keeping the voltage of the power device under test unchanged.

[0018] Further, the logic for determining the isolated feature data is as follows:

[0019] All feature data with the same voltage value are divided into a group, for the feature data with the same voltage value, based on the temperature and the current, a neighborhood search radius for clustering analysis of the feature data in the group is determined, and the calculation formula of the neighborhood search radius is as follows:

[0020]

[0021] In the formula, T i and T j are the temperature of the i-th feature data and the temperature of the j-th feature data in the same group of feature data, I i and I j are the current value of the i-th feature data and the current value of the j-th feature data in the same group of feature data, i and j are indexes of the feature data in the same group, and m is the total number of feature data in the same group.

[0022] In the formula, r T-I (i,j) is a temperature-current coupling factor, μ T-I is the mean of the temperature-current coupling factor, σ T-I is the standard deviation of the temperature-current coupling factor, R T-I is the neighborhood search radius, and δ is a margin adjustment factor.

[0023] Based on the neighborhood search radius, a recursively expanded density clustering algorithm is used to perform clustering analysis on the feature data in the same group to obtain multiple clusters and isolated feature data.

[0024] Further, the logic for dividing clusters and isolated feature data based on the recursively expanded density clustering algorithm is as follows:

[0025] Feature data with the same voltage value are merged to form a feature data set, and the initial state of each feature data in the set is marked as unvisited.

[0026] Randomly pick an unvisited feature data p from the feature data set, p is the index of the feature data in the feature data set, and p∈[1, m];

[0027] Based on the neighborhood search radius, traverse the feature data set to determine the neighborhood N(p) of the feature data p ∈ (p), specifically: if there is a feature data with unvisited state in the feature data set, and the temperature-current coupling factor between the feature data and the feature data p is less than the neighborhood search radius, then the feature data is included in the neighborhood N(p) of the feature data p;

[0028] If the number of feature data in the neighborhood N(p) of the feature data p is less than the minimum neighborhood node number, then reselect an unvisited feature data from the feature data set, otherwise take the feature data p as the core point and create a cluster C p , the cluster C p is composed of the feature data p and the neighborhood N ∈ (p), and the state of the feature data in the feature data set is updated in real time based on the cluster C p , specifically: the state of the feature data included in the cluster C p in the feature data set is modified to visited, wherein the value of the minimum neighborhood node number is between 5% and 10% of the total number of the feature data in the set;

[0029] For the feature data in the neighborhood N ∈ (p), the neighborhood of the feature data is determined by the same method, if the number of feature data in the neighborhood is less than the minimum neighborhood node number, then the neighborhood is not included in the cluster C p , otherwise the neighborhood is included in the cluster C p and the state of the feature data is updated in real time, and the recursive expansion continues until the cluster C p cannot continue to expand;

[0030] Randomly select an unvisited feature data from the feature data set again, and construct a new cluster according to the same method until no cluster can be constructed, and for the feature data not included in any cluster, mark it as isolated feature data.

[0031] Further, the logic for determining fault feature data and frame selection feature data is as follows:

[0032] Based on expert scoring method, score the isolated feature data with voltage, current and accuracy matching rationality as scoring benchmark, generate accuracy credibility probability for evaluating whether the accuracy of the isolated feature data is credible;

[0033] The isolated feature data is scored based on expert scoring method, and a response time length credibility probability for evaluating whether the response time length in the isolated feature data is credible is generated, with voltage, current and response time length rationality as scoring criteria;

[0034] The accuracy rate abnormality is marked on the isolated feature data with the accuracy rate credibility probability less than the preset threshold, and the response time length abnormality is marked on the isolated feature data with the response time length credibility probability less than the preset threshold, and the isolated feature data marked with the accuracy rate abnormality or the response time length abnormality is defined as fault feature data;

[0035] The isolated feature data with both the accuracy rate credibility probability and the response time length credibility probability not less than the preset threshold is screened out as to-be-determined feature data, the credibility probability is corrected based on the distribution of the credibility probability, and the to-be-determined feature data is screened out as selected feature data in combination of the random number and the corrected credibility probability, and the specific screening logic is as follows:

[0036]

[0037] In the formula, is the accuracy rate credibility probability of the kth to-be-determined feature data, is the response time length credibility probability of the kth to-be-determined feature data, k is the index of the to-be-determined feature data, and k ∈ [1, K], K is the total number of to-be-determined feature data, γ1 k represents a first correction factor for correcting , γ2 k represents a second correction factor for correcting ;

[0038] In the formula, X1 k represents a first random number corresponding to , X1 k ~ U(0, 1) represents that the first random number X1 k is randomly selected in the interval [0, 1], k X2 k represents a second random number corresponding to , X2 k ~ U(0, 1) represents that the second random number X2 k is randomly selected in the interval [0, 1];

[0039] In the formula, τ k is a logical judgment value, τ k = 1 represents that the accuracy rate and the response time length of the kth to-be-determined feature data are both correct and reasonable, and then the k to-be-determined feature data is selected as the selected feature data, τ k = 0 represents that the accuracy rate and the response time length of the kth to-be-determined feature data are not both correct and reasonable, and then the k to-be-determined feature data is not selected as the selected feature data.

[0040] Furthermore, the logic for determining the normal characteristic data is as follows:

[0041] The feature data within a cluster is defined as non-isolated feature data. Based on the voltage value, current value, and accuracy of the selected feature data, the first radius of the neighborhood search is determined for performing a single cluster analysis on the non-isolated feature data. The formula for calculating the first radius of the neighborhood search is as follows:

[0042]

[0043] In the formula, η x η y P represents the accuracy of selecting feature data for the x-th box and the accuracy of selecting feature data for the y-th box out of all bounding boxes. x P y P represents the power value of the x-th feature data and the power value of the y-th feature data in the selected feature data of the entire box. x =I x ×U x I x U x P represents the current value and voltage value of the x-th feature data in the selected feature data of the entire box. y =I y ×U y I y U y Here, x and y are the current and voltage values ​​of the yth feature data in the entire box-selected feature data, respectively, where x and y are the indices of the box-selected feature data, and n is the total number of box-selected feature data.

[0044] In the formula, r η-P (x,y) is the accuracy-power coupling factor, μ η-P The mean of the accuracy-power coupling factor, σ η-P R represents the standard deviation of the accuracy-power coupling factor. η-P δ1 is the first radius of the neighborhood search, and δ1 is the first margin adjustment factor;

[0045] Based on the first radius of neighborhood search, a recursive expansion density clustering algorithm is used to perform cluster analysis on non-isolated feature data to obtain multiple first clusters and first isolated feature data.

[0046] Based on the voltage, current, and response time of the selected feature data, a second radius for neighborhood search is determined for secondary clustering analysis of non-isolated feature data. The formula for calculating the second radius for neighborhood search is as follows:

[0047]

[0048] In the formula, tx , t y respectively are the response time of the xth frame selection feature data and the response time of the yth frame selection feature data, r η-P (x,y) is the response time-power coupling factor, μ t-P is the mean of the response time-power coupling factor, σ t-P is the standard deviation of the response time-power coupling factor, R t-P is the neighborhood search second radius, δ2 is the second margin adjustment factor;

[0049] Based on the neighborhood search second radius, using the recursively extended density clustering algorithm, the non-isolated feature data is clustered and analyzed to obtain a plurality of second clusters and second isolated feature data;

[0050] The non-isolated feature data belonging to the first cluster and the second cluster at the same time is defined as normal feature data.

[0051] Further, the data simulation module includes three sub-models of temperature prediction model, accuracy prediction model and response time prediction model, the input of the temperature prediction model is the voltage value and the current value, and the output is the corresponding temperature, the input of the accuracy prediction model is the temperature, the voltage value and the current value, and the output is the corresponding accuracy, the input of the response time prediction model is the temperature, the voltage value and the current value, and the output is the corresponding response time, the temperature prediction model, the accuracy prediction model and the response time prediction model all use the deep learning framework of TensorFlow.

[0052] Further, the safe working standard includes: the temperature is not higher than the highest temperature threshold, the accuracy is not lower than the lowest accuracy threshold, and the response time is not higher than the longest response time threshold;

[0053] The acquisition logic of the safe working area is: keeping the voltage value input in the data simulation module unchanged, gradually increasing the current value input, defining the maximum current value that meets the temperature not higher than the highest temperature threshold as the maximum safe current of the power device under test under the voltage value, defining the maximum current value that meets the temperature not higher than the highest temperature threshold, the accuracy not lower than the lowest accuracy threshold and the response time not higher than the longest response time threshold as the maximum suitable current of the power device under test under the voltage value, defining the region from 0 to the maximum safe current as the first safe working area of the power device under test under the corresponding voltage value without damage, and defining the region from 0 to the maximum suitable current as the second safe working area of the power device under test under the corresponding voltage value suitable for working.

[0054] The application discloses an intelligent power device safety working area test system for realizing data judgment, and relates to the technical field of intelligent power device safety working area test.

[0055] Compared with the prior art, the application has the beneficial effects that:

[0056] The intelligent power device safety working area test equipment and system for realizing data judgment of the application first screen out isolated characteristic data with abnormal temperature through a recursive expansion density clustering algorithm, then screen out frame selection characteristic data with reasonable accuracy and response time length from the isolated characteristic data by combining with an expert scoring method with random numbers, frame selection normal characteristic data with reasonable indexes from the clustering based on the frame selection characteristic data, so as to achieve the purpose of eliminating abnormal values, then train the generative adversarial network by using the reasonable normal characteristic data, so as to obtain a large number of reasonable simulation characteristic data, train the data simulation module based on the simulation characteristic data, so as to simulate the nonlinear relationship among the voltage value, the current value, the temperature, the accuracy and the response time length, finally control the data simulation module to output the safety working area based on the safety working standard, the setting of eliminating abnormal values improves the accuracy of subsequent analysis of the safety working area, the cooperative setting of the generative adversarial network and the data simulation module reduces the actual test workload, and the efficiency of the safety working area test is greatly improved. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 It is a module unit diagram of the whole equipment of the application;

[0058] Figure 2 It is a current-temperature scatter plot;

[0059] Figure 3 It is a power-accuracy scatter plot;

[0060] Figure 4 It is a power-response time scatter plot;

[0061] Figure 5 It is a comprehensive confidence probability control diagram. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical scheme and advantages of the application more clear and apparent, the application is further described in detail below in combination with specific embodiments.

[0063] It should be noted that the technical terms or scientific terms used in the present application should be understood as the general meaning understood by those skilled in the art to which the present application belongs, unless otherwise defined. The "first", "second" and similar words used in the present application do not represent any order, quantity or importance, but are only used to distinguish different components. "Include" or "contain" and similar words mean that the elements or objects before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connected" or "connected" and similar words are not limited to physical or mechanical connection, but can include electrical connection, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to represent relative positional relationship, which may change accordingly when the absolute position of the described object changes.

[0064] Embodiment one:

[0065] Please refer to Figures 1-5 The present application provides a kind of intelligent power device safety working area test equipment for realizing data judgment, comprising:

[0066] First test module, for applying current increment test working condition to the power device to be tested, based on the temperature of the power device to be tested under each test working condition, determine the current demarcation value;

[0067] Wherein, the first test module includes power source for applying different current to the power device to be tested, temperature sensor for detecting the surface temperature of the power device to be tested, control software for controlling the power source to output different current, and the specific test method is as follows:

[0068] Set the power source to the lowest current value, which can be set to 0, and set the test environment temperature of the power device to be tested to its normal working environment temperature, to ensure the rationality of the test process;

[0069] Start from the lowest current value, increase the current according to the set current increment amplitude (the current increment amplitude can be set between 0.1A-0.5A), and collect the temperature of the power device to be tested after the power device to be tested is stable;

[0070] It should be noted that after each current increment, wait for a certain time (specifically 30 seconds-2 minutes) to ensure that the power device to be tested reaches a stable state, and then collect the temperature of the power device to be tested, to ensure the accuracy of data collection;

[0071] If the current value after the current increment is the current demarcation value, the test is stopped, and the current value after the current increment is taken as the current demarcation value; if not, the current is continuously increased according to the set current increment until the current demarcation value is determined, and the specific determination method is as follows:

[0072] After the to-be-tested power device is stabilized, the temperature of the to-be-tested power device is collected. If the temperature exceeds the preset temperature threshold, the current value after the current increment is determined as the current demarcation value; otherwise, it is determined that it is not the current demarcation value. The temperature threshold can be specifically set to 80 degrees Celsius or the maximum working temperature recommended by the manufacturer. When the temperature exceeds the temperature threshold, it indicates that the current of the to-be-tested power device at this time is at risk. If the current continues to increase, it will adversely affect the normal operation of the to-be-tested power device. Therefore, this current is taken as the current demarcation value.

[0073] Further, after the to-be-tested power device is stabilized, the temperature is collected multiple times according to the preset collection frequency, and the average temperature rise rate is calculated. The current value when the average temperature rise rate first exceeds the temperature rise rate threshold is recorded. The current value is compared with the current demarcation value in the above, and the minimum value of the two is taken as the final current demarcation value.

[0074] It should be noted that the collection frequency can be set to 10 seconds per collection, a total of five times. The temperature rise rate is the ratio of the temperature change amount to the time change amount. The multiple temperature collection and average temperature rise rate calculation are common technical means in the art, which will not be repeated here. The temperature rise rate threshold can be set to 1 degree Celsius per second. When the average temperature rise rate exceeds the temperature rise rate threshold, it indicates that the to-be-tested power device is still abnormally heated after being stabilized, that is, the heat dissipation capacity of the to-be-tested power device at this time cannot meet the heat dissipation demand. The minimum value of the two is taken as the final current demarcation value, which takes into account the temperature and temperature rise of the to-be-tested power device, ensures the accuracy of the selection of the current demarcation value, and avoids the problem that the to-be-tested power device is damaged in the subsequent test process due to the too large value of the current demarcation value.

[0075] Further, the first test module further includes a voltmeter for detecting the voltage of the to-be-tested power device. When different current values are applied to the to-be-tested power device, the voltage of the to-be-tested power device is measured synchronously, and a current-voltage characteristic curve is drawn. The point in the curve where the slope first falls below the slope threshold is recorded. The minimum value of the current value corresponding to the point, the current value when the average temperature rise rate first exceeds the temperature rise rate threshold, and the current demarcation value in the above is selected as the final current demarcation value.

[0076] It should be noted that the current-voltage characteristic curve with the horizontal axis as the current value and the vertical axis as the voltage value is drawn based on software such as MATLAB, and the slope threshold value can be set to 0.01-0.1, and when the slope of the curve from a certain point is lower than the slope threshold value, it indicates that the curve tends to be flat, that is, as the current increases, the voltage no longer continues to increase, and the power device under test reaches a saturated state, and if the current continues to increase, it will cause damage to the power device under test, and in the technical solution, the minimum value of the three is taken as the final current demarcation value, which takes into account the temperature, temperature rise and saturation state of the power device under test, ensures the accuracy of the selection of the current demarcation value, and avoids the problem of damage to the power device under test caused by the selection of a large current demarcation value;

[0077] As an embodiment, the power source can be a programmable DC power supply with a model number of Keysight E3634A, the voltmeter can be a high-precision digital multimeter with a model number of Fluke 87V, the temperature sensor can be an infrared thermometer or a thermocouple with a model number of Fluke 568, and the control software can be computer software such as LabVIEW, and the current value applied by the power source to the power device under test can be directly read from the programmable DC power supply or the control software.

[0078] The second test module is configured to apply a test condition with an unchanged voltage and an increasing current to the power device under test to obtain characteristic data of the power device under test under each test condition, the characteristic data including the current value, the voltage value, the temperature, the accuracy rate when the data is executed, and the response time length of the power device under test, and the increasing amplitude of the current is determined based on the current demarcation value;

[0079] The logic for determining the current increasing amplitude based on the current demarcation value is as follows: starting from the lowest current value, increasing the current according to a set first increasing amplitude until the current value is first not less than the current demarcation value, and then increasing the current according to a set second increasing amplitude, and the second increasing amplitude is smaller than the first increasing amplitude.

[0080] It should be noted that at a lower voltage, the power device to be tested can withstand a larger current, and vice versa at a higher voltage, the current value that the power device to be tested can withstand will decrease, that is, due to the difference in voltage value, the upper limit of the current that the power device to be tested can withstand is also different, and in the calculation of the current demarcation value above, the voltage of the power device to be tested increases with the increase of the input current, so the current demarcation value is the upper limit of the current of the power device to be tested at a higher voltage, that is, the power device to be tested generally will not be damaged before reaching the current demarcation value, and it is possible to be damaged after exceeding the current demarcation value, so here the current demarcation value is used as the adjustment index of the increase amplitude, before the current reaches the current demarcation value, the current is increased at a larger first increase amplitude, the test working condition is reduced to improve the test efficiency, and after the current reaches the current demarcation value, the current is increased at a smaller second increase amplitude, so as to avoid the risk of damage to the power device to be tested due to the large increase amplitude, the first increase amplitude can be specifically set to be between 0.1A-0.5A, and the second increase amplitude can be specifically set to be between 0.01A-0.05A, which is set by the staff according to the actual situation, and is not limited here;

[0081] Among them, the second test module includes a power source for applying different currents to the power device to be tested, a temperature sensor for detecting the surface temperature of the power device to be tested, control software for controlling the power source to output different currents, an adjustable load for keeping the voltage of the power device to be tested unchanged, and the specific test method is as follows:

[0082] A plurality of voltage test points lower than the breakdown voltage are determined, for example, for a power device to be tested of MOSFET type, the breakdown voltage is 300V, five voltage test points of 50V, 100V, 150V, 200V and 250V are determined;

[0083] Starting from the first voltage test point (such as 50V), the power source applies the voltage to the power device to be tested, and starts from the lowest current value, increases the current according to the set first increase amplitude, and after the power device to be tested is stable, the temperature of the power device to be tested is collected, and the accuracy and response time of the data judgment performed after the power device to be tested is stable;

[0084] Based on the temperature, it is judged whether to continue to increase the current, if yes, stop testing, if not, continue to increase the current according to the set first increase amplitude or second increase amplitude, until the temperature exceeds the preset temperature threshold, stop increasing the current, avoid damage to the power device to be tested, the temperature threshold can be set to be consistent with the above;

[0085] Further, the same method as above can be used to introduce temperature rise to determine the node at which the current increase is stopped, specifically: after the power device under test stabilizes, the temperature is collected multiple times at a preset collection frequency and the average temperature rise rate is calculated, when the average temperature rise rate exceeds the temperature rise rate threshold, the current increase is stopped to avoid damage to the power device under test;

[0086] The same method is used to test and obtain the characteristic data of the power device under test at other voltage test points in turn;

[0087] The accuracy and response time are obtained by: after the power device under test stabilizes, inputting a plurality of sample data with known true labels to the power device under test, obtaining the predicted labels of the power device under test for each sample data, counting the number of sample data predicted correctly by the power device under test, and taking the ratio of the number of sample data predicted correctly to the total number of sample data as the accuracy of the power device under test under the corresponding voltage and current conditions. The response time of the power device under test for processing each sample data is calculated and the mean value is obtained as the response time of the power device under test under the corresponding voltage and current conditions.

[0088] As an example, the power device under test is used to determine the charging state as normal or fault according to the voltage, current and temperature data of the electric vehicle during charging, then the sample data is voltage, current and temperature data, the true label is the charging state (normal or fault) corresponding to the sample data, 50 groups of sample data with known true labels are input to the power device under test, the predicted labels of the power device under test for each sample data are obtained, and if the number of sample data predicted correctly by the power device under test is 45 groups, the accuracy is 90%.

[0089] As an example, the adjustable load can be an electronic load with model number Chroma 63000, the second test module and the first test module can share the power source, temperature sensor and control software, and the current value and voltage value applied by the power source to the power device under test can be directly read from the programmable DC power supply, control software and electronic load.

[0090] The data frame selection module is used to cluster and analyze the characteristic data based on the recursive expansion density clustering algorithm with temperature-current as the clustering reference, determine a plurality of clusters and isolated characteristic data, and introduce random numbers on the basis of expert scoring method to analyze isolated characteristic data to determine fault characteristic data and frame selection characteristic data, frame select the clusters based on the frame selection characteristic data to determine normal characteristic data;

[0091] The logic for determining isolated characteristic data is as follows:

[0092] All feature data with the same voltage value are divided into a group, and for the feature data with the same voltage value, a neighborhood search radius for clustering analysis of the group of feature data is determined based on temperature and current, and the calculation formula of the neighborhood search radius is as follows:

[0093]

[0094] In the formula, T i and T j are the temperature of the i-th feature data and the temperature of the j-th feature data in the same group of feature data, respectively, I i and I j are the current value of the i-th feature data and the current value of the j-th feature data in the same group of feature data, respectively, i and j are indexes of the feature data in the same group, and the feature data in the same group is sorted in order of current value from small to large, and m is the total number of feature data in the same group of feature data.

[0095] In the formula, r T-I (i,j) is a temperature-current coupling factor, which is used to reflect the temperature change resistance of the power device to be tested when the current value changes from I i to the current value I j under the same voltage environment, and the square of the current is proportional to the heat, and here is used to represent the temperature change resistance of the power device to be tested, and the larger the ratio , the greater the temperature change amplitude of the power device to be tested when the current value changes from I i to the current value I j , and the worse the temperature change resistance.

[0096] In the formula, μ T-I is the mean of the temperature-current coupling factor, is set to ensure that any two feature data in the same group are calculated once, avoiding the problem of unnecessary calculation power waste caused by repeated calculation, is a mathematical statistical symbol, such as σ T-I is the standard deviation of the temperature-current coupling factor, is also set to avoid repeated calculation;

[0097] In the formula, R T-IFor the neighborhood search radius, for the power device to be tested under normal working condition, because the current is proportional to the square of the heat, if the collected characteristic data is accurate, the quadratic current-temperature scatter plot drawn with the square of the current as the horizontal axis and the temperature as the vertical axis should have the following characteristics: the slope of the line connecting any two points (i.e. the temperature-current coupling factor in the above) will have some difference, but for a plurality of consecutive points, the slope of the line connecting any two points in the plurality of consecutive points will fluctuate up and down at a certain value (see Figure 2 ), if the slope of the line connecting two points deviates from the certain value to a large extent, it means that at least one of the temperature values corresponding to the two points has an abnormal value, ignoring the case of low reading error and power source failure, only considering the following two cases: case one, the power device to be tested has a sudden failure due to long-term high-intensity testing, such as element fatigue work leading to failure of the power device to be tested, under the same voltage and current test conditions, the temperature, accuracy and response time of the power device to be tested in this state will deviate from the normal state of the power device to be tested, case two, the external environmental temperature and other factors change to interfere with the reading accuracy of the temperature sensor, and ignore the extremely rare case of both cases, the common point of case one and case two is that the temperature is abnormal, the difference between the two is that: in case one, the power device to be tested is in a fault state, the accuracy and response time are also different from the normal power device to be tested under the same current and voltage, in case two, the power device to be tested is in a normal state, the accuracy and response time are the same as the normal power device to be tested under the same current and voltage;

[0098] Based on the above analysis, the recursive expanding density clustering algorithm is used to select the isolated points, and the neighborhood search radius is selected as the mean value μ T-I of the temperature-current coupling factor as the reference value, δ×σ T-I is used to increase the search margin, and the neighborhood search radius is represented in the form of μ T-I + δ × σ T-I , δ is the margin adjustment factor, generally set between 1-3, of course, it can also be set by the staff according to the actual situation, but a too large value is easy to ignore the abnormal characteristic data, otherwise a too small value is easy to misjudge the normal characteristics as abnormal, the margin adjustment factor needs to be reasonably selected;

[0099] Based on the neighborhood search radius, the recursive expanding density clustering algorithm is used to cluster and analyze the characteristic data in the same group to obtain a plurality of clusters and isolated characteristic data, the specific logic is as follows:

[0100] The characteristic data with the same voltage value is combined to form a characteristic data set, and the initial state of each characteristic data in the set is marked as unvisited;

[0101] Randomly select an unvisited feature data p from the feature data set, where p is the index of the feature data in the feature data set and p∈[1,m];

[0102] Based on the neighborhood search radius, the feature data set is traversed to determine the neighborhood N of feature data p. ∈ (p), specifically: if there exists a feature data in the feature data set that is in an unvisited state, and the temperature-current coupling factor between this feature data and feature data p is less than the neighborhood search radius, then this feature data is included in the neighborhood N of feature data p. ∈ (p) in;

[0103] If the neighborhood of feature data p is N ∈ If the number of feature data points within (p) is less than the minimum number of neighboring nodes, then unvisited feature data points are reselected from the feature data set; otherwise, feature data point p is used as the core point and a cluster C is created. p Clustering C p The feature data p and the neighborhood N ∈ (p) Composition, and based on clustering C p The status of feature data within the feature dataset is updated in real time, specifically by including features in cluster C. p The status of the feature data is changed to "visited", where the minimum number of neighboring nodes is between 5% and 10% of the total number of feature data in this group.

[0104] For the neighborhood N ∈ For the feature data within (p), the same method is used to determine its neighborhood. If the number of feature data within that neighborhood is less than the minimum number of neighborhood nodes, then that neighborhood is not included in cluster C. p Otherwise, include the neighborhood in cluster C. p It also updates the status of the feature data in real time and continues to recursively expand until cluster C is reached. p Unable to expand further;

[0105] Again, randomly select an unvisited feature from the feature data set and construct a new cluster using the same method until no cluster can be constructed. For feature data that does not belong to any cluster, mark it as isolated feature data.

[0106] It should be noted that, based on the current-temperature coupling factor and the recursive expansion density clustering algorithm mentioned above, if a feature data is an isolated feature data, it means that the temperature in the feature data is an outlier. However, it is still uncertain whether the accuracy and response time are outliers. Therefore, the isolated feature data is marked as having an abnormal temperature.

[0107] The logic for determining fault characteristic data and box selection characteristic data is as follows:

[0108] The voltage, the current and the accuracy rate rationality are taken as the scoring benchmarks, the isolated characteristic data is scored based on the expert scoring method, and an accuracy rate credibility probability for evaluating whether the accuracy rate in the isolated characteristic data is credible is generated;

[0109] It should be noted that the value of the accuracy rate credibility probability is between 0 and 1, 0 represents complete uncredibility, and 1 represents complete credibility. When scoring based on the expert scoring method, at least 5 experts or technical personnel in the relevant field are invited, and working reports of the same type of power device are provided for the experts to refer to. The working reports at least include the current, the voltage and the accuracy rate of the power device under normal working state and abnormal working state. After each expert gives the accuracy rate credibility probability, the mean value is taken after removing the maximum and minimum value, as the accuracy rate credibility probability for evaluating whether the accuracy rate in the isolated characteristic data is credible;

[0110] The voltage, the current and the response time length rationality are taken as the scoring benchmarks, the isolated characteristic data is scored based on the expert scoring method, and a response time length credibility probability for evaluating whether the response time length in the isolated characteristic data is credible is generated;

[0111] It should be noted that the value of the response time length credibility probability is between 0 and 1, 0 represents complete uncredibility, and 1 represents complete credibility. When scoring based on the expert scoring method, at least 5 experts or technical personnel in the relevant field are invited, and working reports of the same type of power device are provided for the experts to refer to. The working reports at least include the current, the voltage and the response time length of the power device under normal working state and abnormal working state. After each expert gives the response time length credibility probability, the mean value is taken after removing the maximum and minimum value, as the response time length credibility probability for evaluating whether the response time length in the isolated characteristic data is credible;

[0112] The accuracy rate abnormality is marked on the isolated characteristic data with the accuracy rate credibility probability less than a preset threshold, and the response time length abnormality is marked on the isolated characteristic data with the response time length credibility probability less than a preset threshold. The isolated characteristic data marked with the accuracy rate abnormality or the response time length abnormality is defined as the fault characteristic data. At least one of the accuracy rate abnormality and the response time length abnormality is marked on the fault characteristic factor, and the temperature abnormality is marked;

[0113] The preset threshold corresponding to the accuracy rate credibility probability can be specifically set as 0, 1-0.4, so as to screen out the isolated characteristic data with obviously unreasonable accuracy rate as the fault characteristic data. Similarly, the preset threshold corresponding to the response time length credibility probability can be specifically set as 0, 1-0.4, so as to screen out the isolated characteristic data with obviously unreasonable response time length as the fault characteristic data. The fault characteristic data corresponds to the case one in the foregoing description. Of course, the specific value of the preset threshold can also be set by the staff according to the actual situation, which is not described here;

[0114] Screening the isolated feature data whose accuracy confidence probability and response duration confidence probability are not less than the preset threshold value as the to-be-determined feature data, correcting the confidence probability based on the distribution of the confidence probability, and screening the frame selection feature data from the to-be-determined feature data in combination with the random number and the corrected confidence probability, the specific screening logic is as follows:

[0115]

[0116] In the formula, is the accuracy confidence probability of the kth to-be-determined feature data, is the response duration confidence probability of the kth to-be-determined feature data, k is the index of the to-be-determined feature data, and k [1, K], K is the total number of to-be-determined feature data, is used to represent the comprehensive confidence probability of the accuracy and response duration of the kth to-be-determined feature data, which meets the conventional mathematical statistical logic in a multiplicative manner;

[0117] In the formula, X1 k represents the first random number corresponding to X1 k ~ U(0, 1) represents that the first random number X1 k is randomly selected in the interval [0, 1]; k represents the second random number corresponding to X2 k ~ U(0, 1) represents that the second random number X2 k is randomly selected in the interval [0, 1];

[0118] It should be noted that the accuracy confidence probability only represents the possibility that the accuracy is correct and reasonable, when is a decimal (i.e. the interval is (0, 1)), it is difficult to determine the appropriate threshold value through a single conventional threshold value to determine whether the accuracy is correct and reasonable, and if the threshold value is not reasonably selected, it will cause subsequent classification problems, and a single conventional threshold value is difficult to achieve accurate judgment, for example, if the threshold value is set to 0.8, but there is an accuracy rate of 0.7, which is actually correct and reasonable, and there is an accuracy rate of 0.9, which is actually incorrect and unreasonable;

[0119] Here, the technical solution creatively uses a random value countermeasure judgment method, that is, a first random number X1 k is set for each accuracy confidence probability , by comparing the accuracy confidence probability and the first random number X1 k , it is judged whether the accuracy is reliable, when the accuracy confidence probability is not less than the first random number X1k If the accuracy of the k-th feature data to be determined is considered reliable, then it is considered unreliable. This method is equivalent to introducing a random dynamic threshold for judgment, and the random dynamic threshold is more in line with probability characteristics. Compared with the single conventional threshold judgment, it solves the problem of difficulty in determining a suitable threshold.

[0120] Furthermore, compared to other accuracy probabilities, the higher the accuracy probability, i.e., the greater the difference between the accuracy probability and other accuracy probabilities, the higher the probability that the accuracy is correct and reasonable. Therefore, when comparing accuracy probabilities... And the first random number X1 k At that time, the accuracy and reliability probability should be improved. Not less than the first random number X1 k The tendency is to adjust the confidence probability based on the distribution of the confidence probability, hence the first adjustment factor γ1 is set. k To correct the accuracy confidence probability First correction factor γ1 k Consists of benchmark item 1 and correction item composition, This represents the mean confidence probability of the accuracy of all undetermined feature data. The settings are used to characterize the accuracy confidence probability. Relative to accuracy and mean confidence probability The degree of excellence, The larger the value, the more reliable the accuracy probability. The better the performance, the higher the accuracy and reliability probability in comparison. And the first random number X1 k At that time, the accuracy and reliability probability should be improved. Not less than the first random number X1 k The tendency to use Characterized by the first correction factor γ1 k ,by The form of accuracy and reliability probability After making the corrections, use the corrected results. With the first random number X1 k To conduct analysis and judgment, For the simplified γ2 k To be used to correct the reliability probability of response time The second correction factor is calculated using the same logic as the first correction factor, and will not be elaborated here.

[0121] It should be noted that the above text uses Based on the comprehensive reliability probability of the accuracy and response time of the k-th feature data to be determined, this section compares... The magnitude relationship is used to determine whether the accuracy and response time of the k-th feature data to be determined are both correct and reasonable. k τ is a logical judgment value used to determine whether the accuracy and response time of the k-th feature data to be determined are both correct and reasonable. k =1 indicates that the accuracy and response time of the k-th feature data to be determined are both correct and reasonable. Therefore, the k-th feature data to be determined are selected as the bounding box feature data. τ k =0 indicates that the accuracy and response time of the k-th feature data to be determined are not both correct and reasonable. Therefore, the k-th feature data to be determined will not be used as the bounding box feature data. The bounding box feature data will only be marked with the label of temperature anomaly.

[0122] Please refer to the following for details. Figure 5 Based on a conventional threshold judgment method (threshold set to 0.8), multiple reference samples are selected from the feature data to be determined. Reference samples are feature data whose accuracy confidence probability and response time confidence probability are both greater than the threshold. Then, based on an expert scoring method, the accuracy confidence probability and response time confidence probability of the reference samples are given. The accuracy confidence probability and response time confidence probability are multiplied together to obtain the comprehensive confidence probability of each reference sample. A plot is then drawn based on the comprehensive confidence probability of each reference sample. Figure 5 The red line graph;

[0123] After determining the bounding box selection feature data based on this method, multiple bounding box selection feature data are selected as samples. Then, based on the expert scoring method, the accuracy confidence probability and response time confidence probability of each sample are given. The accuracy confidence probability and response time confidence probability are multiplied to obtain the comprehensive confidence probability of each sample. Based on the comprehensive confidence probability of each sample, a graph is plotted. Figure 5 The blue line graph shows that, by comparing the two line graphs, the red line fluctuates between 0.65 and 1, while the blue line fluctuates between 0.85 and 1. This indicates that the bounding box selection feature data determined by this method is closer to the real data.

[0124] The logic for determining normal characteristic data is as follows:

[0125] The feature data within a cluster is defined as non-isolated feature data. Based on the voltage value, current value, and accuracy of the selected feature data, the first radius of the neighborhood search is determined for performing a single cluster analysis on the non-isolated feature data. The formula for calculating the first radius of the neighborhood search is as follows:

[0126]

[0127] In the formula, η x η yrespectively are the accuracy of the xth frame selection feature data and the accuracy of the yth frame selection feature data in all frame selection feature data, P x , y respectively are the power value of the xth feature data and the power value of the yth feature data in all frame selection feature data, P x = I x × U x , I x , U x respectively are the current value and the voltage value of the xth feature data in all frame selection feature data, P y = I y × U y , I y , U y respectively are the current value and the voltage value of the yth feature data in all frame selection feature data, x and y are both indexes of frame selection feature data, and all frame selection feature data is sorted in order of power value from small to large, and if there are two or more frame selection feature data with the same power value, one frame selection feature data is randomly retained for numbering sorting, and n is the total number of frame selection feature data;

[0128] In the formula, r η-P (x,y) is an accuracy-power coupling factor, which is used to reflect the accuracy stability performance of the power device when the power value changes from P x to the power value P y , and in this case, the accuracy stability performance of the power device is characterized by The larger the ratio is, the greater the change range of the accuracy of the power device is when the power value changes from P x to the power value P y , and the worse the accuracy stability performance is;

[0129] In the formula, μ η-P is the mean of the accuracy-power coupling factor, The setting of is used to ensure that any two frame selection feature data are calculated once, avoiding the problem of wasting computing power caused by repeated calculation, is a mathematical statistical symbol, and η-P is the standard deviation of the accuracy-power coupling factor, The setting of

[0130] In the formula, R η-PFor the first radius of neighborhood search, for the power device to be tested under normal working condition, if the collected characteristic data is accurate, the power-accuracy scatter plot with power as horizontal axis and accuracy as vertical axis should have the following characteristics: the slope of the line connecting any two points (i.e. the accuracy-power coupling factor in the foregoing) will have certain difference, but for continuous multiple points, the slope of the line connecting any two points in the continuous multiple points will fluctuate up and down at a certain value (see Figure 3 ), if the slope of the line connecting two points deviates from the certain value to a large extent, it indicates that at least one of the accuracy values corresponding to the two points is an abnormal value, and in this case, the low-level reading error is ignored, and the abnormal value occurs: the sudden failure of the power device to be tested due to long-time high-intensity test, such as element fatigue work leading to failure of the power device to be tested, under the same test condition of power, the accuracy of the power device to be tested under this state deviates from the accuracy of the power device to be tested under normal state;

[0131] According to the above analysis, the recursive expanding density clustering algorithm is selected to screen out isolated points, and the neighborhood search radius is selected as the mean value μ η-P of the accuracy-power coupling factor η-P as the reference value, and δ1×σ T-I is used to increase the search margin, and the neighborhood search radius is represented in the form of μ T-I + δ × σ, δ1 is the first margin adjustment factor, which is generally set between 1-3, of course, it can also be set by the staff according to the actual situation, but too large value is easy to cause the problem of ignoring abnormal characteristic data, and vice versa, too small value is easy to misjudge the normal characteristic as abnormal, so the margin adjustment factor needs to be reasonably selected;

[0132] Based on the first radius of neighborhood search, the recursive expanding density clustering algorithm is used to perform clustering analysis on the non-isolated characteristic data to obtain multiple first clusters and first isolated characteristic data, the accuracy of the non-isolated characteristic data in the first cluster is normal, and the accuracy of the first isolated characteristic data is abnormal, the specific clustering method has been discussed in the foregoing, and will not be repeated here;

[0133] Based on the frame selected characteristic data, the voltage value, the current value and the response time are determined to determine the second radius of neighborhood search for secondary clustering analysis of the non-isolated characteristic data, and the calculation formula of the second radius of neighborhood search is as follows:

[0134]

[0135] In the formula, t x and t y are the response time of the xth frame selected characteristic data and the response time of the yth frame selected characteristic data, respectively, r η-P(x, y) is a response time-power coupling factor, which is used to reflect the power value from P x to the power value P y , the response time stability performance of the power device to be tested is characterized by , and the greater the ratio , the greater the response time change range of the power device to be tested when the power value changes from P x to the power value P y , and the worse the response time stability performance.

[0136] In the formula, μ t-P is the average of the response time-power coupling factor, is set to ensure that any two box-selected feature data are calculated twice to avoid unnecessary waste of computing power caused by repeated calculations, is a mathematical statistical symbol, such as σ t-P is the standard deviation of the response time-power coupling factor, is also set to avoid repeated calculations.

[0137] In the formula, R t-P is the second radius of neighborhood search. For the power device to be tested under normal working conditions, if the collected feature data is accurate, the power-response time scatter plot with power as the horizontal axis and response time as the vertical axis should have the following characteristics: the slope of the line connecting any two points (i.e., the response time-power coupling factor mentioned above) will have some differences, but for a plurality of consecutive points, the slope of the line connecting any two points in the plurality of consecutive points will fluctuate up and down at a constant value (see Figure 4 for details). If the slope of the line connecting two points deviates from the constant value to a large extent, it means that at least one of the response time values corresponding to the two points is an outlier. Ignoring low-level reading errors, the occurrence of outliers is: the sudden failure of the power device to be tested due to long-term high-intensity testing, such as element fatigue work leading to failure of the power device to be tested, and under the same power test conditions, the response time of the power device to be tested in this state will deviate from the response time of the power device to be tested under normal conditions.

[0138] Based on the above analysis, the recursive expanding density clustering algorithm is used to screen out outliers, and the response time-power coupling factor average μ t-P is used as the reference value, and δ2×σ t-P is used to increase the search margin, and μ t-P + δ2×σ t-Pcharacterize the neighborhood search radius in the form of δ 1, and δ 2 is a second margin adjustment factor, which is generally set between 1-3, of course, it can also be set by the staff according to the actual situation, but too large value is easy to cause the problem of ignoring the abnormal feature data, otherwise, too small value is easy to misjudge the normal feature as abnormal, and the margin adjustment factor needs to be reasonably selected;

[0139] Based on the second radius of neighborhood search, the recursive expanding density clustering algorithm is used for clustering analysis of non-isolated feature data to obtain a plurality of second clusters and second isolated feature data. The non-isolated feature data in the second cluster is normal in response duration, and the response duration of the second isolated feature data is abnormal. The specific clustering method has been discussed in the foregoing, and will not be repeated here.

[0140] The non-isolated feature data belonging to the first cluster and the second cluster at the same time is defined as normal feature data, so as to complete the framing of the normal feature data. The temperature value, accuracy rate and response time in the normal feature data are all normal, that is, the normal feature data is not marked with an abnormality.

[0141] The data expansion module is used for generating a plurality of groups of simulated feature data based on the data expansion module constructed based on the adversarial network, and the data expansion module is trained based on the normal feature data;

[0142] The data expansion module is composed of a generator and a discriminator. The generator is a neural network with random noise as input and simulated feature data (i.e. voltage value, current value, temperature, accuracy rate and response duration) as output. The neural network includes at least two fully connected layers. The discriminator is a neural network for discriminating whether the input data is real data or generated data. The final output is a judgment probability value. The data expansion module can use the existing GAN adversarial generation network framework. The specific training process is a prior art, and the details are as follows:

[0143] A predetermined training round (which can be set between 30-50) is set. In each training round, the discriminator is trained first and then the generator is trained until the predetermined training round is reached. The logic for training the discriminator is as follows: a certain number (such as 30) of normal feature data is randomly selected, the same number of false feature data is generated using the generator, and the discriminator is used to judge the two batches of data respectively, and the loss value is calculated. The label of the real data (i.e. normal feature data) is 1, and the label of the generated data (i.e. false feature data) is 0. The optimization goal of the discriminator is to maximize the judgment probability of the real data and minimize the judgment probability of the generated data. Specifically, the binary cross-entropy formula can be used as the loss function of the discriminator. The logic for training the generator is as follows: random noise is input into the generator, and the loss of the generated data is calculated. The optimization goal is to make the discriminator think that these generated data are real. Specifically, the binary cross-entropy formula can also be used as the loss function of the generator.

[0144] After the data augmentation module is trained, the generator of the data augmentation module is used to randomly generate multiple sets of simulated feature data to ensure the sufficiency of the samples during the subsequent training of the data simulation module and to avoid the problem of inaccurate prediction performance of the data simulation module due to the lack of training samples.

[0145] The data simulation module is based on a deep neural network. The input of the data simulation module is voltage and current values, and the output is temperature, accuracy and response time. The data simulation module is trained based on simulated feature data and normal feature data, and based on the safety working standard, the safety working area under each voltage is output.

[0146] The data simulation module uses the deep learning framework of TensorFlow, and includes an input layer for receiving voltage and current values, multiple hidden layers for data processing of input data, and ReLU as the activation function on the hidden layer, and an output layer for outputting temperature, accuracy and response time. The training process of the data simulation module is as follows:

[0147] The simulated feature data and the normal feature data are summarized to form a sample set, the sample set is divided into a training set, a test set and a validation set according to a ratio of 70:15:15, and the batch size and the training period are set. The current and voltage values in the training set are used as input, and the corresponding temperature, accuracy and response time are used as output labels to train the data simulation module. The root mean square error is used as the loss function. In the training process, the model parameters are updated through back propagation to minimize the loss function value. Specifically, the model parameters can be updated by optimization algorithms such as Adam and SGD. In the training process, the model hyperparameters (such as learning rate, batch size, number of hidden layers, etc.) are adjusted by the validation set to optimize the model performance. After reaching the predetermined training period, the test set is input into the data simulation module for performance testing. If the deviation between the data simulation module and the true value is less than 3%, the training is considered complete. Otherwise, the training is restarted.

[0148] Further, the temperature of the power device to be tested is mainly affected by the voltage and current values, while the accuracy and response time are simultaneously affected by the voltage, current and temperature. Therefore, the data simulation module is divided into three sub-models: temperature prediction model, accuracy prediction model and response time prediction model. The input of the temperature prediction model is voltage and current values, and the output is the corresponding temperature. The input of the accuracy prediction model is temperature, voltage and current values, and the output is the corresponding accuracy. The input of the response time prediction model is temperature, voltage and current values, and the output is the corresponding response time. In this way, the three models can capture the non-linear relationship between the corresponding input and output, achieving accurate prediction.

[0149] It should be noted that the temperature prediction model, the accuracy prediction model and the response time prediction model all use the deep learning framework of TensorFlow, and the specific training process is the same as the method described above, which will not be repeated here.

[0150] The safety working standard includes: the temperature is not higher than the maximum temperature threshold, the accuracy is not lower than the minimum accuracy threshold, and the response time is not longer than the maximum response time threshold. The maximum temperature threshold, the minimum accuracy threshold and the maximum response time threshold are set by the staff according to the actual situation, such as the maximum temperature threshold can be set to 85 degrees Celsius, the minimum accuracy threshold can be set to 95%, and the maximum response time threshold can be set to 10 milliseconds, and the like, which will not be limited here.

[0151] The acquisition logic of the safety working area is: keeping the voltage value input in the data simulation module unchanged, gradually increasing the current value input, the maximum current value that meets the condition that the temperature is not higher than the maximum temperature threshold is defined as the maximum safe current of the power device under test at the voltage value, the maximum current value that meets the conditions that the temperature is not higher than the maximum temperature threshold, the accuracy is not lower than the minimum accuracy threshold and the response time is not longer than the maximum response time threshold is defined as the maximum suitable current of the power device under test at the voltage value, the region from 0 to the maximum safe current is defined as the first safety working area of the power device under test at the corresponding voltage value, and the region from 0 to the maximum suitable current is defined as the second safety working area of the power device under test at the corresponding voltage value, so as to complete the test division of the safety working area.

[0152] Embodiment two provides an intelligent power device safety working area test system for realizing data judgment, which comprises the intelligent power device safety working area test device for realizing data judgment described above, and further comprises a power device under test for safety working area test.

[0153] The above formulas are all dimensionless numerical calculations, and the formulas are obtained by collecting a large amount of data to simulate the most recent real situation, and the preset parameters in the formula are set by the person skilled in the art according to the actual situation.

[0154] The above embodiments can be realized by software, hardware, firmware or any combination thereof. When realized by software, the above embodiments can be realized in the form of a computer program product. Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are executed by hardware or software depends on the specific application and design constraints of the technical solution.

[0155] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, and may be located in one place, or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiment of the present application according to actual needs.

[0156] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A test device for the safe operating area of ​​intelligent power devices that enables data judgment, characterized in that, include: The first test module is used to apply a test condition with increasing current to the power device under test, and to determine the current threshold value based on the temperature of the power device under test under each test condition. The second test module is used to apply a test condition with constant voltage and increasing current to the power device under test in order to obtain characteristic data of the power device under test under each test condition. The characteristic data includes the current value, voltage value, temperature, accuracy and response time of the power device under test when performing data judgment, and the magnitude of the current increase is determined based on the current threshold value. The data selection module is used to perform cluster analysis on feature data based on temperature-current as the clustering benchmark and a recursive expansion density clustering algorithm to determine multiple clusters and isolated feature data. It also introduces random numbers based on the expert scoring method to analyze isolated feature data to determine fault feature data and selection feature data. Based on the selection feature data, it performs selection of clusters to determine normal feature data. The data augmentation module, built on an adversarial network, is used to generate multiple sets of simulated feature data, and is trained based on normal feature data. The data simulation module, built on a deep neural network, takes voltage and current values ​​as input and outputs temperature, accuracy, and response time. The data simulation module is trained based on simulated feature data and normal feature data, and outputs the safe operating range for each voltage based on safe operating standards.

2. The intelligent power device safe operating area test equipment for data judgment according to claim 1, characterized in that: The first test module includes a power source for applying different currents to the power device under test, a temperature sensor for detecting the surface temperature of the power device under test, and control software for controlling the power source to output different currents; The current threshold determination logic is as follows: after the power device under test stabilizes, the temperature of the power device under test is collected. If the temperature exceeds the preset temperature threshold, the current value after the current increase is determined to be the current threshold value; otherwise, it is determined not to be the current threshold value.

3. The intelligent power device safe operating area test equipment for data judgment according to claim 1, characterized in that: The second test module includes a power source for applying different currents to the power device under test, a temperature sensor for detecting the surface temperature of the power device under test, control software for controlling the power source to output different currents, and an adjustable load for keeping the voltage of the power device under test constant.

4. The intelligent power device safe operating area test equipment for data judgment according to claim 1, characterized in that: The logic for identifying isolated feature data is as follows: All feature data with the same voltage value are grouped together. For feature data with the same voltage value, a neighborhood search radius is determined based on temperature and current to perform cluster analysis on this group of feature data. The formula for calculating the neighborhood search radius is as follows: In the formula, T i T j Let I represent the temperature of the i-th feature data and the temperature of the j-th feature data in the same set of feature data. i I j Let i and j be the current values ​​of the i-th and j-th feature data in the same set of feature data, respectively, where i and j are the indices of the feature data in the same set, and m is the total number of feature data in the same set of feature data. In the formula, r T-I (i,j) is the temperature-current coupling factor, μ T-I The mean temperature-current coupling factor, σ T-I R is the standard deviation of the temperature-current coupling factor. T-I The neighborhood search radius is δ, and the margin adjustment factor is δ. Based on the neighborhood search radius, a recursively extended density clustering algorithm is used to perform cluster analysis on the same group of feature data to obtain multiple clusters and isolated feature data.

5. The intelligent power device safe operating area test equipment for data judgment according to claim 4, characterized in that, The logic of the density clustering algorithm based on recursive expansion to divide clusters and isolated feature data is as follows: Feature data with the same voltage value are merged into a feature data set, and the initial state of each feature data in the set is marked as unvisited. Randomly select an unvisited feature data p from the feature data set, where p is the index of the feature data in the feature data set and p∈[1,m]; Based on the neighborhood search radius, the feature data set is traversed to determine the neighborhood N of feature data p. ∈ (p), specifically: if there exists a feature data in the feature data set that is in an unvisited state, and the temperature-current coupling factor between this feature data and feature data p is less than the neighborhood search radius, then this feature data is included in the neighborhood N of feature data p. ∈ (p) in; If the neighborhood of feature data p is N ∈ If the number of feature data points within (p) is less than the minimum number of neighboring nodes, then unvisited feature data points are reselected from the feature data set; otherwise, feature data point p is used as the core point and a cluster C is created. p Clustering C p The feature data p and the neighborhood N ∈ (p) Composition, and based on clustering C p The status of feature data within the feature dataset is updated in real time, specifically by including features in cluster C. p The status of the feature data is changed to "visited", where the minimum number of neighboring nodes is between 5% and 10% of the total number of feature data in this group. For the neighborhood N ∈ For the feature data within (p), the same method is used to determine its neighborhood. If the number of feature data within that neighborhood is less than the minimum number of neighborhood nodes, then that neighborhood is not included in cluster C. p Otherwise, include the neighborhood in cluster C. p It also updates the status of the feature data in real time and continues to recursively expand until cluster C is reached. p Unable to expand further; Again, randomly select an unvisited feature from the feature data set and construct a new cluster using the same method until no cluster can be constructed. For feature data that does not belong to any cluster, mark it as isolated feature data.

6. The intelligent power device safe operating area test equipment for data judgment according to claim 1, characterized in that, The logic for determining fault characteristic data and box selection characteristic data is as follows: Using the reasonableness of matching voltage, current and accuracy as the scoring benchmark, isolated feature data is scored based on the expert scoring method, and an accuracy confidence probability is generated to evaluate whether the accuracy in the isolated feature data is reliable. Using the reasonableness of the matching of voltage, current and response time as the scoring benchmark, isolated feature data is scored based on the expert scoring method, and a response time reliability probability is generated to evaluate whether the response time in the isolated feature data is reliable. Mark accuracy anomalies on isolated feature data where the accuracy confidence probability is less than a preset threshold, and mark response time anomalies on isolated feature data where the response time confidence probability is less than a preset threshold. Define isolated feature data marked with accuracy anomalies or response time anomalies as fault feature data. Isolated feature data with both accuracy and response time confidence probabilities not less than preset thresholds are selected as feature data to be determined. The confidence probabilities are corrected based on their distribution. Combining random numbers and the corrected confidence probabilities, bounding box selection feature data is selected from the feature data to be determined. The specific selection logic is as follows: In the formula, Let k be the accuracy confidence probability of the k-th feature data to be determined. Let γ1 be the confidence probability of the response time of the k-th feature data to be determined, where k is the index of the feature data to be determined, and k∈[1,K], and K is the total number of feature data to be determined. k Indicates that it is used for correction The first correction factor, γ2 k Indicates that it is used for correction The second correction factor; In the formula, X1 k Indicates and The corresponding first random number, X1 k ~U(0,1) means randomly selecting the first random number X1 within the interval [0,1]. k X2 k Indicates and The corresponding second random number, X2 k ~U(0,1) means randomly selecting a second random number X2 within the interval [0,1]. k ; In the formula, τ k τ is the logical judgment value. k =1 indicates that the accuracy and response time of the k-th feature data to be determined are both correct and reasonable. Therefore, the k-th feature data to be determined are selected as the bounding box feature data. τ k =0 indicates that the accuracy and response time of the k-th feature data to be determined are not both correct and reasonable, so the k-th feature data to be determined will not be selected as the bounding box feature data.

7. The intelligent power device safe operating area test equipment for data judgment according to claim 1, characterized in that, The logic for determining normal characteristic data is as follows: The feature data within a cluster is defined as non-isolated feature data. Based on the voltage value, current value, and accuracy of the selected feature data, the first radius of the neighborhood search is determined for performing a single cluster analysis on the non-isolated feature data. The formula for calculating the first radius of the neighborhood search is as follows: In the formula, η x η y P represents the accuracy of selecting feature data for the x-th box and the accuracy of selecting feature data for the y-th box out of all bounding boxes. x P y P represents the power value of the x-th feature data and the power value of the y-th feature data in the selected feature data of the entire box. x =I x ×U x I x U x P represents the current value and voltage value of the x-th feature data in the selected feature data of the entire box. y =I y ×U y I y U y Here, x and y are the current and voltage values ​​of the yth feature data in the entire box-selected feature data, respectively, where x and y are the indices of the box-selected feature data, and n is the total number of box-selected feature data. In the formula, r η-P (x,y) is the accuracy-power coupling factor, μ η-P The mean of the accuracy-power coupling factor, σ η-P R represents the standard deviation of the accuracy-power coupling factor. η-P δ1 is the first radius of the neighborhood search, and δ1 is the first margin adjustment factor; Based on the first radius of neighborhood search, a recursive expansion density clustering algorithm is used to perform cluster analysis on non-isolated feature data to obtain multiple first clusters and first isolated feature data. Based on the voltage, current, and response time of the selected feature data, a second radius for neighborhood search is determined for secondary clustering analysis of non-isolated feature data. The formula for calculating the second radius for neighborhood search is as follows: In the formula, t x t y Let r be the response time of the x-th selected feature data and the y-th selected feature data in the total selected feature data. η-P (x,y) is the response time-power coupling factor, μ t-P To represent the mean of the response time-power coupling factor, σ t-P For the response time-power coupling factor standard deviation, R t-P δ2 is the second radius for neighborhood search, and δ2 is the second margin adjustment factor. Based on the second radius of neighborhood search, a recursively extended density clustering algorithm is used to perform cluster analysis on non-isolated feature data to obtain multiple second clusters and second isolated feature data. Non-isolated feature data that simultaneously belongs to both the first and second clusters are defined as normal feature data.

8. The intelligent power device safe operating area test equipment for data judgment according to claim 1, characterized in that: The data simulation module includes three sub-models: a temperature prediction model, an accuracy prediction model, and a response time prediction model. The temperature prediction model takes voltage and current values ​​as inputs and outputs the corresponding temperature. The accuracy prediction model takes temperature, voltage, and current values ​​as inputs and outputs the corresponding accuracy. The response time prediction model takes temperature, voltage, and current values ​​as inputs and outputs the corresponding response time. All three models use the TensorFlow deep learning framework.

9. The intelligent power device safe operating area test equipment for data judgment according to claim 1, characterized in that, The safety operating standards include: temperature not exceeding the highest temperature threshold, accuracy not lower than the lowest accuracy threshold, and response time not exceeding the longest response time threshold. The logic for obtaining the safe operating area is as follows: keeping the voltage input in the data simulation module constant, gradually increasing the current input, defining the maximum current value that satisfies the condition that the temperature is not higher than the highest temperature threshold as the maximum safe current of the power device under test at that voltage value, defining the maximum current value that simultaneously satisfies the conditions that the temperature is not higher than the highest temperature threshold, the accuracy is not lower than the lowest accuracy threshold, and the response time is not higher than the longest response time threshold as the maximum suitable current of the power device under test at that voltage value, defining the region from 0 to the maximum safe current as the first safe operating area where the power device under test will not be damaged at the corresponding voltage value, and defining the region from 0 to the maximum suitable current as the second safe operating area where the power device under test can operate appropriately at the corresponding voltage value.

10. A test system for the safe operating area of ​​intelligent power devices that realizes data judgment, comprising the test equipment for the safe operating area of ​​intelligent power devices that realizes data judgment according to any one of claims 1-9, characterized in that... It also includes power devices under test for safe operating area testing.

Images