WO2022227018A1 - 一种功率半导体器件的测试系统、云服务器及测试机 - Google Patents

一种功率半导体器件的测试系统、云服务器及测试机 Download PDF

Info

Publication number
WO2022227018A1
WO2022227018A1 PCT/CN2021/091550 CN2021091550W WO2022227018A1 WO 2022227018 A1 WO2022227018 A1 WO 2022227018A1 CN 2021091550 W CN2021091550 W CN 2021091550W WO 2022227018 A1 WO2022227018 A1 WO 2022227018A1
Authority
WO
WIPO (PCT)
Prior art keywords
power semiconductor
artificial intelligence
testing machine
intelligence model
semiconductor device
Prior art date
Application number
PCT/CN2021/091550
Other languages
English (en)
French (fr)
Inventor
杜若阳
唐诗
龙纲
邱辉
Original Assignee
华为数字能源技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为数字能源技术有限公司 filed Critical 华为数字能源技术有限公司
Priority to CN202180007976.XA priority Critical patent/CN115398251A/zh
Priority to PCT/CN2021/091550 priority patent/WO2022227018A1/zh
Publication of WO2022227018A1 publication Critical patent/WO2022227018A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G01R31/26—Testing of individual semiconductor devices
    • G01R31/2601—Apparatus or methods therefor
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G01R31/26—Testing of individual semiconductor devices
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G01R31/26—Testing of individual semiconductor devices
    • G01R31/2607—Circuits therefor
    • G01R31/2637—Circuits therefor for testing other individual devices
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • H—ELECTRICITY
    • H01—ELECTRIC ELEMENTS
    • H01L—SEMICONDUCTOR DEVICES NOT COVERED BY CLASS H10
    • H01L22/00—Testing or measuring during manufacture or treatment; Reliability measurements, i.e. testing of parts without further processing to modify the parts as such; Structural arrangements therefor

Definitions

  • the present application relates to the technical field of power semiconductors, and in particular, to a testing system, cloud server and testing machine for power semiconductor devices.
  • a power semiconductor device generally refers to a semiconductor device that realizes the function of circuit switching, for example, it can be applied to power conversion circuits of electric vehicles, photovoltaic systems or data centers.
  • power semiconductor devices can be used in electric vehicle powertrain control circuits.
  • performance testing of the power semiconductor device is required after production to reduce the risk of failure of the power semiconductor device.
  • the quality test of power semiconductor devices mainly sets the upper and lower limit values of the performance parameters of the power semiconductor devices through the testing machine, and eliminates the power semiconductor devices whose performance parameters exceed the upper and lower limit values.
  • the present application provides a power semiconductor device testing system, cloud server and testing machine, which can more comprehensively screen defective power semiconductor devices.
  • the test system for power semiconductor devices includes a cloud server and at least one test machine; the cloud server is used for pre-training with historical data of multiple power semiconductor devices to obtain an artificial intelligence (AI, Artificial Intelligence) model;
  • the test machine is used to obtain the data of the power semiconductor device to be tested; the cloud server or the test machine can use the trained artificial intelligence model to diagnose the power semiconductor device to be tested, that is, the cloud server or the test machine inputs the data of the power semiconductor device to be tested.
  • Artificial intelligence model the output of the artificial intelligence model is the diagnosis result of the power semiconductor device to be tested.
  • a relatively accurate artificial intelligence model can be trained by using the powerful computing power and storage capacity of the cloud server, and then the pre-trained artificial intelligence model can be used to more accurately diagnose the power semiconductor device to be tested.
  • the high-risk power semiconductor devices can be used to make up for the limitation of traditional test systems that only use upper limit data to screen high-risk power semiconductor devices, that is, the power semiconductor devices can be diagnosed more accurately and comprehensively.
  • the cloud server is used to input the data of the power semiconductor device to be tested into the artificial intelligence model, the output of the artificial intelligence model is the diagnosis result of the power semiconductor device to be tested, and the diagnosis result is sent to the test machine.
  • the cloud server inputs the data into the artificial intelligence model, obtains the diagnosis result, and sends the diagnosis result to the test machine, so the test machine can directly obtain the test result, which saves the computing resources of the test machine.
  • the cloud server in the embodiment of the present application is used to send the pre-trained artificial intelligence model to the testing machine; the testing machine inputs the data of the power semiconductor device to be tested into the artificial intelligence model, and the output of the artificial intelligence model It is the diagnostic result of the power semiconductor device to be tested.
  • the cloud server sends the trained artificial intelligence model to the testing machine, which then inputs data and obtains diagnostic results. Therefore, the test machine can complete the test of the power semiconductor device to be tested by a single machine, and it is not necessary to obtain the diagnosis result from the cloud server, which reduces the dependence on the network between the cloud server and the test machine, and improves the test of the power semiconductor device in the embodiment of the present application.
  • the system obtains real-time diagnosis results.
  • the testing machine in this embodiment of the present application is further configured to collect historical data of multiple power semiconductor devices, and send the collected historical data of multiple power semiconductor devices to a cloud server;
  • the historical data includes At least one of the chip test data of multiple power semiconductor devices, the functional test data of the packaged module, or the relevant data when the actual application is abnormal;
  • the cloud server is also used to utilize the historical data of multiple power semiconductor devices to utilize supervised learning At least one of the model or the unsupervised learning model is trained to obtain an artificial intelligence model.
  • the cloud server uses unsupervised learning model training to obtain an artificial intelligence model, which is specifically used to obtain failure types of multiple power semiconductor devices, and uses expert knowledge to extract each failure in the test items of power semiconductor devices. All test items corresponding to the type, and all test items corresponding to each failure type form a test item subset, and use an unsupervised learning model for each test item subset to distinguish abnormal subsets from normal subsets, and obtain abnormal subsets and normal subsets.
  • the subsets are recorded with the first score and the second score respectively, and the sum of the first score and the second score of all test item subsets of each power semiconductor device is obtained as the total score of abnormal level; the total score of abnormal level is greater than or equal to the preset score When the threshold value is reached, the power semiconductor device is judged to be an abnormal device.
  • the embodiment of the present application adopts an unsupervised learning model to train an AI model, and can make full use of expert knowledge to jointly perform outlier detection on a subset of test items including multiple test items, and can use the same test item subset. , or the physical basis corresponding to the same failure type, to detect the risk of power semiconductor device failure from multiple dimensions, so that the risk of device failure can be more accurately assessed.
  • the cloud server in this embodiment of the present application is specifically configured to use expert knowledge to record different first scores for abnormal devices in different test item subsets. It should be understood that different test item subsets may have different weights for judging abnormal devices, and by recording abnormal devices in different test item subsets with different first scores, a more accurate diagnosis result can be obtained.
  • the cloud server uses a supervised learning model to train to obtain an artificial intelligence model, which is specifically used to obtain relevant data when the power semiconductor device is abnormal as a data label, and use the supervised learning model to extract data.
  • the data feature of the tag is used to obtain the abnormal level total score of each power semiconductor device. When the abnormal level total score is greater than the preset score threshold, it is determined that the power semiconductor device is an abnormal device.
  • the testing machine in the embodiment of the present application is further configured to compare the data of the power semiconductor device to be tested with the preset upper and lower limit values, and when the data of the power semiconductor device to be tested exceeds the upper and lower limit values , judging that the power semiconductor device to be tested is an abnormal device.
  • the test system for power semiconductor devices provided by the embodiments of the present application can combine traditional test methods with artificial intelligence models, and use the artificial intelligence models in the embodiments of the present application to perform secondary tests on the basis of traditional test methods. Detection, improve the accuracy of detection, realize the complementarity between traditional methods and AI model testing methods, and can diagnose high-risk power semiconductor devices more comprehensively.
  • the embodiment of the present application does not limit the number of testing machines included in the testing system, which may include one testing machine, or may include two or more testing machines.
  • the cloud server can first train the global AI model, and then obtain a local AI model suitable for each test machine based on the global AI model according to the differences between the test machines. That is, in a possible implementation manner, the test system provided by the embodiment of the present application includes at least the following two test machines: a first test machine and a second test machine; the cloud server is specifically used for sending the first test machine according to the first test machine.
  • Adjust to obtain a second artificial intelligence model use the first artificial intelligence model to test the power semiconductor device to be tested corresponding to the first testing machine to obtain a first diagnosis result, and use the second artificial intelligence model to test the corresponding to-be-tested device of the second testing machine.
  • the power semiconductor device is tested to obtain a second diagnostic result, the first diagnostic result is sent to the first testing machine, and the second diagnostic result is sent to the second testing machine.
  • the embodiment of the present application uses the first historical data and the second historical data to train the global artificial intelligence model, which can make full use of all the historical data and improve the testing accuracy of the global artificial intelligence model.
  • the implementation of this application also uses the first historical data to adjust the global artificial intelligence model to obtain the first artificial intelligence model, and uses the second historical data to adjust the global artificial intelligence model.
  • the artificial intelligence model is adjusted to obtain a second artificial intelligence model, so that a local artificial intelligence model (first artificial intelligence model) for the characteristics of the first test machine and a local artificial intelligence model for the characteristics of the second test machine (the second artificial intelligence model) can be obtained. model), which further improves the test accuracy of the artificial intelligence model used in actual detection.
  • each test machine corresponds to a global AI model
  • the diagnosis result is obtained by the cloud server.
  • the test system provided by the embodiment of the present application includes at least the following two test machines: a first test machine and a second test machine;
  • a global artificial intelligence model is obtained by training a historical data and the second historical data sent by the second testing machine, and the global artificial intelligence intelligent model is used to test the power semiconductor device to be tested corresponding to the first testing machine to obtain the first diagnosis result.
  • the artificial intelligence model tests the power semiconductor device to be tested corresponding to the second testing machine to obtain a second diagnostic result, sends the first diagnostic result to the first testing machine, and sends the second diagnostic result to the second testing machine.
  • the cloud server uses the first historical data and the second historical data to train the global artificial intelligence model, and can make full use of all historical data, thereby improving the generalization ability of the global artificial intelligence model, and using the global artificial intelligence model with higher accuracy
  • the artificial intelligence model diagnoses the power semiconductor device to be tested, which can improve the universality of the artificial intelligence model and increase the accuracy of the diagnosis of new devices.
  • the cloud server obtains the local AI model corresponding to each test machine and sends it to each test machine respectively, and each test machine completes the diagnosis. That is, in a possible implementation manner, the testing system for the power semiconductor device in the embodiment of the present application includes at least the following two testing machines: a first testing machine and a second testing machine; the cloud server is specifically used for testing according to the first testing machine.
  • the first historical data sent and the second historical data sent by the second testing machine are trained to obtain a global artificial intelligence model, and the first historical data is used to adjust the global artificial intelligence model to obtain the first artificial intelligence model and send it to the first testing machine, Using the second historical data to adjust the global artificial intelligence model to obtain the second artificial intelligence model and send it to the second testing machine; the first testing machine is specifically configured to use the first artificial intelligence model to diagnose the corresponding power semiconductor device to be tested; The second testing machine is specifically configured to use the second artificial intelligence model to diagnose the corresponding power semiconductor device to be tested.
  • the testing machine of the embodiment of the present application can directly use the artificial intelligence model to complete the task of detecting the power semiconductor device to be tested, and does not need to communicate with the cloud server in the cloud.
  • the network between the testing machine on the board and the cloud server in the cloud can be avoided.
  • the testing system for the power semiconductor device in the embodiment of the present application cannot obtain the diagnostic result of the power semiconductor device to be tested.
  • the power semiconductor device testing system provided by the embodiments of the present application can reduce the network dependence between the cloud server and the testing machine, and improve the real-time performance of the diagnostic results obtained by the power semiconductor device testing system in the embodiments of the present application.
  • the cloud server sends the global AI model to each test machine, and each test machine uses the global AI model to complete the diagnosis.
  • the testing system of the power semiconductor device in the embodiment of the present application includes at least the following two testing machines: a first testing machine and a second testing machine; The first historical data sent by the testing machine and the second historical data sent by the second testing machine are trained to obtain a global artificial intelligence model, and the global artificial intelligence model is sent to the first testing machine and the second testing machine; the first testing machine is used for The corresponding power semiconductor device to be tested is diagnosed by using the global artificial intelligence intelligent model; the second testing machine is used for diagnosing the corresponding power semiconductor device to be tested by using the global artificial intelligence model.
  • the testing machine in the embodiment of the present application can complete the testing work of the power semiconductor device to be tested by a single machine.
  • the cloud server does not need to deliver the diagnostic result to the testing machine. Therefore, the diagnostic result obtained by the testing machine is not affected by the network failure, and the dependence on the network is low.
  • the testing machine in the embodiment of the present application is further configured to send update data to the cloud server; the cloud server is further configured to update the artificial intelligence model according to the update data.
  • the test machine is also used to fine-tune the pre-trained AI model in combination with the data on the test machine side, use the fine-tuned model to diagnose the power semiconductor device to be tested, and output the diagnosis result. It should be understood that the testing machine in the embodiment of the present application can fine-tune the obtained AI model based on its own data or characteristics of the data, which can further improve the accuracy of the AI model in the embodiment of the present application.
  • the test machine can send the fine-tuned AI model to the cloud server for unified management.
  • the embodiments of the present application further provide a cloud server.
  • the advantages of the above embodiments of the test system are also applicable to the following servers, which will not be repeated.
  • the server includes: a first transceiver device and a first controller; a first transceiver device for receiving test data of a power semiconductor device to be tested sent by a testing machine; a first controller for pre-using multiple power semiconductor devices
  • the artificial intelligence model is obtained by training with historical data; it is also used to input the data of the power semiconductor device to be tested into the artificial intelligence model, and the output of the artificial intelligence model is the diagnosis result of the power semiconductor device to be tested; the first transceiver device is also used to The diagnosis result is sent to the testing machine; or, the first controller is used for sending the artificial intelligence model to the testing machine, so that the testing machine uses the artificial intelligence model to diagnose the power semiconductor device to be tested.
  • the first controller in this embodiment of the present application is specifically used to obtain the failure types of multiple power semiconductor devices when an artificial intelligence model is obtained by training an unsupervised learning model, using expert knowledge All test items corresponding to each failure type are extracted from the test items of power semiconductor devices, all test items corresponding to each failure type form a test item subset, and an unsupervised learning model is used to detect abnormal levels of each test item subset. , get the abnormal device and the normal device, respectively record the first score and the second score, and obtain the sum of the first score and the second score of all test item subsets of each power semiconductor device as the total score of abnormal level; the total score of abnormal level is greater than When the score threshold is preset, it is determined that the power semiconductor device is an abnormal device.
  • the cloud server in this embodiment of the present application corresponds to the following at least two testing machines: a first testing machine and a second testing machine;
  • the first historical data and the second historical data sent by the second testing machine are used for training to obtain a global artificial intelligence model, the first historical data is used to adjust the global artificial intelligence model to obtain the first artificial intelligence model, and the second historical data is used to analyze the global artificial intelligence model.
  • Adjust the model to obtain a second artificial intelligence model use the first artificial intelligence model to test the power semiconductor device to be tested corresponding to the first testing machine to obtain a first diagnosis result, and use the second artificial intelligence model to test the waiting power semiconductor device corresponding to the second testing machine.
  • the power semiconductor device is tested to obtain a second diagnosis result;
  • the first transceiver device is specifically configured to send the first diagnosis result to the first testing machine, and is also used to send the second diagnosis result to the second testing machine.
  • the cloud server in this embodiment of the present application corresponds to the following at least two testing machines: a first testing machine and a second testing machine;
  • a global artificial intelligence model is obtained by training a historical data and the second historical data sent by the second testing machine, and the global artificial intelligence intelligent model is used to test the power semiconductor device to be tested corresponding to the first testing machine to obtain the first diagnosis result.
  • the artificial intelligence model tests the power semiconductor device to be tested corresponding to the second test machine to obtain a second diagnosis result;
  • the first transceiver device is specifically used to send the first diagnosis result to the first test machine, and is also used to send the second diagnosis result The result is sent to the second test machine.
  • the first transceiver device in the embodiment of the present application is further configured to receive update data sent by the testing machine; the first controller is further configured to update the artificial intelligence model according to the update data.
  • the embodiments of the present application further provide a test machine, and the advantages of the above embodiments of the test system are also applicable to the following test machines, which will not be repeated.
  • the test machine includes: a second transceiver device and a second controller; the second transceiver device is used to receive an artificial intelligence model sent by a cloud server, and the artificial intelligence model is obtained by the cloud server through pre-training with historical data of multiple power semiconductor devices
  • the second controller is used to obtain the data of the power semiconductor device to be tested; the data of the power semiconductor device to be tested is input into the artificial intelligence model, and the output of the artificial intelligence model is the diagnosis result of the power semiconductor device to be tested.
  • the second controller in the embodiment of the present application is further configured to collect historical data of multiple power semiconductor devices, and send the collected historical data of multiple power semiconductor devices to a cloud server; historical data
  • the data includes at least one of chip test data or packaged module function test data of a plurality of power semiconductor devices.
  • the second controller in the embodiment of the present application is further configured to compare the data of the power semiconductor device to be tested with the preset upper and lower limit values, and the data of the power semiconductor device to be tested exceeds the upper and lower limits When the value is detected, it is judged that the power semiconductor device to be tested is an abnormal device.
  • the embodiments of the present application have the following advantages:
  • the test system provided by the embodiment of the present application includes a cloud server and a test machine.
  • the computing power and storage capacity of the cloud server are higher than those of the test machine. Therefore, the artificial intelligence model is obtained by pre-training by using the powerful computing power and storage capacity of the cloud server. Since the cloud server uses a large amount of historical data of power semiconductor devices to train the artificial intelligence model in advance, the parameters of the artificial intelligence model can be accurately obtained. Therefore, the cloud server or test machine can use the trained artificial intelligence model to accurately diagnose the power semiconductor device to be tested. , so as to obtain accurate diagnosis results, and eliminate abnormal power semiconductor devices, so as to avoid failures in the actual product use process and cause the entire circuit to fail.
  • the test system can make up for the shortcomings of the traditional method, that is, traditionally only use the upper and lower limit values to screen high-risk power semiconductor devices, and some cannot be screened out. Power semiconductor devices can be diagnosed more comprehensively for risks. This solution can make up for the inaccuracy of the testing machine simply using the upper and lower limit values to screen the defective power semiconductor devices.
  • FIG. 1 is a schematic diagram of a test system for a power semiconductor device provided by an embodiment of the present application
  • FIG. 2 is a schematic diagram of a data source of a power semiconductor device provided in the implementation of this application;
  • 3A is a flowchart of a method for testing a power semiconductor device provided by an embodiment of the present application
  • 3B is a flowchart of another power semiconductor device testing method provided by an embodiment of the present application.
  • FIG. 4 is a schematic diagram of a testing system for another power semiconductor device provided by an embodiment of the present application.
  • FIG. 5 is a schematic diagram of yet another testing system for a power semiconductor device provided by an embodiment of the present application.
  • FIG. 6 is a schematic diagram of another testing system for a power semiconductor device provided by an embodiment of the present application.
  • FIG. 7 is a schematic diagram of yet another testing system for a power semiconductor device provided by an embodiment of the present application.
  • FIG. 8 is a schematic diagram of a testing system for a power semiconductor device including a plurality of testing machines according to an embodiment of the present application;
  • FIG. 9 is a schematic diagram of another testing system of a power semiconductor device including multiple testing machines provided by an embodiment of the present application.
  • FIG. 10 is a schematic diagram of yet another testing system for a power semiconductor device including multiple testing machines provided by an embodiment of the application;
  • FIG. 11 is a schematic diagram of yet another testing system for a power semiconductor device including multiple testing machines provided by an embodiment of the present application;
  • FIG. 12 is a schematic structural diagram of a cloud server according to an embodiment of the present application.
  • FIG. 13 is a schematic structural diagram of a testing machine provided by an embodiment of the present application.
  • directional terms such as “upper” and “lower” may include, but are not limited to, definitions relative to the schematic placement of components in the drawings. It should be understood that these directional terms may be relative concepts, They are used for relative description and clarification, which may vary accordingly depending on the orientation in which the components are placed in the drawings.
  • connection should be understood in a broad sense.
  • connection may be a fixed connection, a detachable connection, or an integrated body; it may be directly connected, or Can be indirectly connected through an intermediary.
  • coupled may be a manner of electrical connection that enables signal transmission.
  • Coupling can be a direct electrical connection or an indirect electrical connection through an intermediate medium.
  • Embodiments of the present application relate to a testing system for power semiconductor devices, wherein the testing system for power semiconductor devices includes: a cloud server and at least one testing machine.
  • Cloud servers have powerful computing energy and storage space. Therefore, cloud servers can train artificial intelligence models using a large amount of historical data of power semiconductor devices. Since the AI module is obtained through a large amount of historical data training, the cloud server or test machine can use the trained AI model to accurately diagnose the power semiconductor devices to be tested, and remove abnormal power semiconductor devices to prevent them from being used in the real process. A fault occurs that causes the entire circuit to fail.
  • the solution can solve the inaccuracy of screening defective products of the power semiconductor device to be tested simply by using the upper and lower limit values preset by the testing machine in the traditional solution.
  • test system provided by the embodiments of the present application is described in detail below with reference to the accompanying drawings, and the test system includes a cloud server and at least one test machine.
  • the embodiment of the present application does not specifically limit the number of testing machines, which may be set according to the number of actual product lines.
  • One testing machine may correspond to one product line, and one product line may also correspond to multiple testing machines.
  • the types of power semiconductor devices corresponding to multiple product lines may be the same or different.
  • three test machines are used as an example to introduce in conjunction with FIG. 1. It should be understood that it may be one test machine, two test machines, or more test machines.
  • FIG. 1 this figure is an architectural diagram of a testing system for a power semiconductor device provided by an embodiment of the present application.
  • the power semiconductor device testing system includes a cloud server 100 , a first testing machine 201 , a second testing machine 202 , and a third testing machine 203 .
  • the cloud service 100 is located on a cloud platform, and the cloud platform may include multiple cloud servers, and only one cloud server is illustrated in the figure as an example.
  • the first testing machine 201 , the second testing machine 202 and the third testing machine 203 on the board side can respectively communicate with the cloud server 100 in the cloud to perform data exchange or transmission.
  • the cloud server 100 is used to obtain an AI model by pre-training with historical data of multiple power semiconductor devices.
  • the embodiments of the present application do not specifically limit the specific model used by the cloud server 100 for AI model training, for example, a supervised learning model or an unsupervised learning model may be used.
  • the historical data used by the cloud server 100 to train the AI model may be the data uploaded by the first testing machine 201, the second testing machine 202, and the third testing machine 203, or may be data uploaded by more testing machines. No specific limitation is made. It should be understood that when the cloud server 100 performs AI model training, the more historical data is used, the more accurate the parameters of the AI model obtained by training.
  • the first testing machine 201 , the second testing machine 202 and the third testing machine 203 are used to obtain data of the respective corresponding power semiconductor devices to be tested.
  • the cloud server 100 may perform the test, or the cloud server 100 may deliver the trained AI model to the test machine, and the test machine may perform the test.
  • the cloud server 100 is tested.
  • the cloud server 100 is configured to input the data of the power semiconductor device to be tested into the pre-trained artificial intelligence model, and the output of the artificial intelligence model is the diagnosis result of the power semiconductor device to be tested. Send the diagnostic results to the testing machine.
  • the power semiconductor device to be tested here generally refers to any power semiconductor device that needs to be tested; the testing machine here generally refers to any testing machine.
  • the first testing machine 201, the second testing machine 202 and the third testing machine 203 are respectively used to input the data of the respective corresponding power semiconductor devices to be tested into the pre-trained artificial intelligence model, and obtain the output of the pre-trained model respectively as Corresponding diagnostic results of the power semiconductor device to be tested.
  • This embodiment of the present application does not limit whether the first testing machine 201, the second testing machine 202, and the third testing machine 203 test the corresponding power semiconductor devices to be tested at the same time, or perform tests in sequence, because the testing action of each testing machine can be Completed independently, without cross-interference.
  • an artificial intelligence model is used to test the performance of the power semiconductor devices.
  • the test system for power semiconductor devices provided by the embodiments of the present application utilizes the powerful computing and storage capabilities of cloud servers to obtain an artificial intelligence model through historical data training of multiple power semiconductor devices, and then uses the artificial intelligence model to test the power Semiconductor devices are tested.
  • the AI model trained by the cloud server can be sent to the test machine, and the test machine uses the AI model trained by the cloud server to test the power semiconductor devices to be tested, and finally outputs the diagnosis result.
  • the testing system for power semiconductor devices provided by the embodiments of the present application can use the artificial intelligence model to diagnose the performance of power semiconductors, and to screen defective products in the power semiconductor devices. And considering the limited computing and storage capabilities of the testing machine on the board side, the present application uses a cloud server with powerful computing and storage capabilities to train the artificial intelligence model, thereby obtaining an artificial intelligence model with high diagnostic accuracy.
  • FIG. 2 is a schematic diagram of a data source of a power semiconductor device provided in the implementation of the present application.
  • the cloud server uses a large amount of historical data for training to obtain an AI model.
  • Historical data can include production line data and application-side data.
  • the source of the production line data D1 may include data obtained from the following tests: chip (CP, Chip Probe) test, functional test (FT, Functional Test), single-board test and complete machine test.
  • the CP test can be the test of a single chip
  • the FT test can be the test of the packaged module, and the module has a certain circuit function.
  • the single-board test can be the upper-board test, and the whole-machine test can be the upper-board test. It can be understood that the chip test data and the power test data can be used separately for AI model training, or can be used together for AI model training.
  • the CP test is mainly aimed at a single chip, and a single chip can be a separate insulated gate bipolar transistor (IGBT, Insulated Gate Bipolar Transistor,), a separate metal-oxide semiconductor field effect transistor (MOSFET, Metal-Oxide -Semiconductor Field-Effect Transistor) chips, diodes (Doide), triodes (Bipolar Junction Transistor), thyristor (Thyristor), integrated gate commutated thyristor (IGCT, Integrated Gate-Commutated Thyristor) and other power semiconductor devices, can also be A chip formed by a plurality of IGBTs can also be a chip formed by IGBTs and diodes.
  • the FT test is mainly aimed at a packaged module,
  • the application-side data D2 mainly includes the relevant data corresponding to the failure of the power semiconductor device in the actual use process after the production of the power semiconductor device, that is, the actual working condition data of the power semiconductor device fed back by the customer side, which can include the test data source and the verification data.
  • test data of the power semiconductor device to be tested in the embodiments of the present application may include, but not limited to, the following breakdown voltage, leakage current, on-voltage drop, parasitic capacitance, gate charge, turn-on loss and turn-off loss.
  • the embodiments of the present application provide an AI model obtained by training a cloud server, which may be obtained by training an unsupervised integrated learning model, or may be obtained by training a supervised learning model; or may be obtained by using a supervised learning model and an unsupervised learning model obtained in combination, not specifically limited in the examples of this application
  • an AI model provided by an embodiment of the present application is obtained by training an unsupervised ensemble learning model.
  • the unsupervised learning model in the embodiment of the present application may be an unsupervised ensemble learning model. It can be understood that the unsupervised model in the embodiment of the present application adopts an unsupervised learning algorithm and expert knowledge to combine to construct an unsupervised integrated learning model, which can make full use of expert experience and reduce the impact of too few failed samples on the accuracy of the artificial intelligence model.
  • the cloud server uses the unsupervised learning model to train to obtain the artificial intelligence model
  • the cloud server is specifically used to obtain the failure types of multiple power semiconductor devices, and the expert knowledge is used to obtain the failure types of the power semiconductor devices.
  • All test items corresponding to each failure type are extracted from the test items, and all test items corresponding to each failure type form a test item subset, and an unsupervised learning model is used for each test item subset to distinguish abnormal subsets from normal subsets , the abnormal subset and the normal subset are respectively recorded as the first score and the second score, and the sum of the first score and the second score of all test item subsets of each power semiconductor device is obtained as the total score of abnormal level; When the score is greater than or equal to the preset score threshold, it is determined that the power semiconductor device is an abnormal device.
  • the failure types may specifically be various failure types that cannot be intercepted by the card control during the single-item test of the power semiconductor device. For example, early failures of power semiconductor devices in actual use were not screened out during controlled testing.
  • a failure type may have problems with many test items, such as gate leakage, which may be a problem with the physical layer. Many test items involved are caused by problems with the physical layer. Therefore, all test items associated with a failure type are extracted to form a subset of test items corresponding to the test item.
  • Subset1 Subset2 Subset3 ... Abnormal Level Total Score Device A 0 0 0 0 0 Device B 0 1 0 1 Device C 0 0 0 0 ... ...
  • device A, device B and device C are respectively different power semiconductor devices.
  • Subset1, Subset2 and Subset3 are different test item subsets respectively; that is, each device includes at least three test item subsets.
  • the content and number of test items included in each test item subset varies.
  • the abnormal subset is denoted as 1
  • the normal subset is denoted as 0
  • the sum of the first and second scores of all test item subsets of each power semiconductor device is obtained as the total abnormal level score ; That is, the abnormal level of device A has a total score of 0, the abnormal level of device B has a total score of 1, and the abnormal level of device C has a total score of 0.
  • the preset score threshold is 1, the test device B is an abnormal device.
  • FIG. 3A is a flowchart of a method for testing a power semiconductor device according to an embodiment of the present application.
  • failure types such as Gate failure, PN junction failure, or pressure ring failure can be obtained. It is understandable that, when there are more historical data of the same power semiconductor device, the more failure types are collected and the more comprehensive.
  • Gate failure corresponds to multiple test items such as IGES, VTH, VCESAT, and Eon, and these test items form a subset1 of test items.
  • PN junction failure corresponds to test items such as VCESAT and ICES, and these test items form a subset2 of test items.
  • the failure of the withstand voltage ring corresponds to test items such as ICES, and these test items form a subset3 of test items.
  • S303 Use an unsupervised learning model to distinguish abnormal subsets from normal subsets for each test item subset, and obtain a first score and a second score for the abnormal subset and normal subset, respectively.
  • an unsupervised learning model is used to detect abnormal points on the subset of test items corresponding to the gate failure, and the obtained normal device A, normal device B, and normal device C are respectively recorded as second scores.
  • An unsupervised learning model is used to detect abnormal points on the subset of test items corresponding to PN junction failures, and the obtained normal device A and normal device C are respectively recorded as the second score, and the abnormal device B is recorded as the first score.
  • the first score is 1 and the second score is 0.
  • the cloud server also uses expert knowledge to record different first scores for abnormal devices in different test item subsets. That is, the weights corresponding to the test item subsets are added on the basis of the first score and the second score. For example, there are three failure types in total, and the three failure types correspond to the three test item subsets a, b and c respectively.
  • the weight of the a test subset is 0.2
  • the weight of the b test subset is 0.3
  • the weight of the c test subset is 0.3.
  • the weight is 0.5.
  • S304 Obtain the sum of the first score and the second score of all test item subsets of each power semiconductor device as a total abnormal level score.
  • the abnormal level of device A is divided into the second score of 0 for subset1 of test items, the second score of 0 for subset2 of test items, and the second score of subset3 of test items
  • the sum of 0 is 0.
  • the total abnormal level of device B is the sum of the second score 0 of the test item subset subset1, the first score 1 of the test item subset subset2 and the second score 0 of the test item subset subset3, which is 1.
  • the abnormal level of the device C is divided into the sum of the second score 0 of the test item subset subset1, the second score 0 of the test item subset subset2 and the second score 0 of the test item subset subset3, which is 0.
  • the preset score threshold may be 1. Of course, other values may also be obtained according to the actual situation, which is not limited in the implementation of this application.
  • the preset score threshold is 1, since the abnormal level total score of device B is greater than or equal to the preset score threshold, it is determined that device B is an abnormal device.
  • the solution introduced in the above embodiment is that the cloud server obtains the artificial intelligence model through the training of the unsupervised learning model.
  • the following describes the artificial intelligence model obtained by the cloud server through the training of the supervised learning model.
  • FIG. 3B is a flowchart of another method for testing a power semiconductor device provided by an embodiment of the present application.
  • S311 Obtain relevant data when the power semiconductor device is abnormal as a data label. For example, a very small number of failure samples can be selected as data labels.
  • S312 Extract data features of data labels by using a supervised learning model.
  • the method for screening defective power semiconductor devices by using an artificial intelligence model in the embodiments of the present application is an improved solution based on the traditional upper and lower limit numerical diagnosis methods. That is, the power semiconductor device testing method in the embodiment of the present application is that the testing machine uses the preset upper and lower limit values to first diagnose the power semiconductor device to be tested, that is, the testing machine, and is also used to compare the data of the power semiconductor device to be tested with the data. The preset upper and lower limit values are compared, and when the data of the power semiconductor device to be tested exceeds the upper and lower limit values, it is determined that the power semiconductor device to be tested is an abnormal device.
  • the test machine or cloud server uses the AI model to diagnose the power semiconductor device to be tested, and perform a logical OR operation on the diagnosis results of the two schemes. Therefore, the test system provided by the embodiments of the present application can complement the two solutions, thereby more comprehensively diagnosing the power semiconductor device to be tested.
  • the power semiconductor device testing system while using the artificial intelligence model to diagnose the power semiconductor device, it can also use the traditional method to compare the data of the power semiconductor device to be tested with the preset upper and lower limits. Values are compared to diagnose power semiconductor devices.
  • the power semiconductor device is judged to be failure, that is, the two diagnosis results are logically ORed.
  • the power semiconductor device testing system provided by the embodiments of the present application can combine the traditional testing method with the artificial intelligence model to further improve the detection accuracy.
  • the data of the power semiconductor device X does not exceed the preset upper and lower limit data, but after the data of the power semiconductor device X is input into the artificial intelligence model, the diagnosis result output by the artificial intelligence model is invalid, at this time, the power semiconductor device X is judged to be invalid.
  • the data of the power semiconductor device Y exceeds the preset upper and lower limit data, but after the data of the power semiconductor device Y is input into the artificial intelligence model, the diagnosis result output by the artificial intelligence model is not invalid, at this time, the power semiconductor device Y is judged to be invalid.
  • the data of the power semiconductor device Z exceeds the preset upper and lower limit data, and after the data of the power semiconductor device Z is input into the artificial intelligence model, the diagnosis result output by the artificial intelligence model is invalid, and at this time, the power semiconductor device Z is judged to be invalid.
  • the testing machine in order to make the diagnosis of the artificial intelligence model more accurate, as a possible implementation manner, is also used to send updated data to the cloud server; the cloud server also uses for updating AI models based on updated data.
  • the step of inputting the data of the power semiconductor device to be tested into the pre-trained artificial intelligence model and obtaining the diagnosis result in the embodiment of the present application can be completed by a cloud server in the cloud or by a test machine on the board side.
  • the embodiments of the present application are not limited herein. The following will specifically introduce two different implementation manners through embodiments.
  • FIG. 4 is a schematic diagram of another testing system of a power semiconductor device provided by an embodiment of the present application.
  • the testing machine 200 is also used to collect historical data of a plurality of power semiconductor devices, and send the collected historical data of the plurality of power semiconductor devices to the cloud server 100, and the cloud server 100 can train artificial intelligence through the historical data of the plurality of power semiconductor devices Model.
  • the cloud server 100 is further configured to use historical data of a plurality of power semiconductor devices to perform training using at least one of a supervised learning model or an unsupervised learning model to obtain an artificial intelligence model.
  • the historical data may include at least one of chip test data of a plurality of power semiconductor devices, functional test data of a packaged module, or related data when the actual application is abnormal.
  • the historical data may include one or more items of the training data in FIG. 2 .
  • the artificial intelligence model obtained by training in the embodiment of the present application is stored in the cloud server.
  • the artificial intelligence model needs to be used to detect the power semiconductor device to be tested the artificial intelligence model to be tested is used by the cloud server.
  • the power semiconductor device is tested, and the test result is obtained.
  • FIG. 5 is a schematic diagram of yet another testing system for a power semiconductor device provided by an embodiment of the present application.
  • the cloud server 100 is configured to input the data of the power semiconductor device to be tested obtained from the testing machine into a pre-trained artificial intelligence model, and the output of the pre-trained artificial intelligence model is the diagnosis result of the power semiconductor device to be tested. Then, the cloud server 100 sends the diagnosis result to the testing machine 200, and the testing machine 200 finally outputs the diagnosis result.
  • the cloud server 100 will store the artificial intelligence model trained by the cloud server 100 .
  • the cloud server 100 can obtain the data of the power semiconductor device to be tested from the testing machine 200, and input the data into the pre-trained artificial intelligence model to obtain the diagnosis result .
  • the cloud server 100 sends the obtained diagnostic result to the testing machine 200, so that the testing machine 200 on the end board outputs the diagnostic result, that is, the testing machine 200 only receives the diagnostic result and does not test the power semiconductor device itself.
  • the step of inputting the data of the power semiconductor to be tested into the artificial intelligence model for pre-selection and training and obtaining the diagnosis result in the embodiment of the present application can also be completed by the testing machine on the board side. Let's first introduce the training process of the AI model.
  • FIG. 6 is a schematic diagram of another testing system for a power semiconductor device provided by an embodiment of the present application.
  • the testing machine 200 is further configured to collect historical data of a plurality of power semiconductor devices, and send the collected historical data of the plurality of power semiconductor devices to the cloud server 100, so that the cloud server 100 can be trained by the historical data of the plurality of power semiconductor devices artificial intelligence model.
  • the historical data includes at least one of chip test data or packaged module function test data of the plurality of power semiconductor devices in the above embodiments. It should be noted that, after training the obtained artificial intelligence model in the embodiment of the present application, the cloud server 100 sends the obtained artificial intelligence model to the testing machine 200 on the board side, and the testing machine 200 stores the artificial intelligence model.
  • FIG. 7 is a schematic diagram of yet another testing system for a power semiconductor device provided by an embodiment of the present application.
  • the cloud server 100 is used for sending the pre-trained artificial intelligence model to the testing machine 200 .
  • the testing machine 200 is configured to input the data of the power semiconductor device to be tested into a pre-trained artificial intelligence model, and the output of the pre-trained artificial intelligence model is the diagnosis result of the power semiconductor device to be tested.
  • the testing machine 200 will receive the artificial intelligence model sent by the cloud server 100, and store the artificial intelligence model.
  • the testing machine 200 inputs the data of the power semiconductor device to be tested into the artificial intelligence model, and obtains and outputs a diagnosis result.
  • the test machine since the artificial intelligence model is stored in the test machine, the test machine can directly use the artificial intelligence model to complete the task of detecting the power semiconductor device to be tested, and does not need to communicate with the cloud server in the cloud. Therefore, it can be avoided that the power semiconductor device testing system in the embodiment of the present application cannot obtain the diagnostic result of the power semiconductor device to be tested when the network is interrupted between the on-board testing machine and the cloud server in the cloud. In this way, the power semiconductor device testing system provided by the embodiments of the present application can reduce the network dependence between the cloud server and the testing machine, and improve the real-time performance of the diagnostic results obtained by the power semiconductor device testing system in the embodiments of the present application.
  • the cloud server in the cloud can input the data of the power semiconductor to be tested into the artificial intelligence model of pre-selected training and obtain the diagnosis result, or the test machine on the board side can input the data of the power semiconductor to be tested into the artificial intelligence model. Data from power semiconductors is fed into pre-selected trained artificial intelligence models and diagnostic results are obtained.
  • the cloud server can directly store the trained artificial intelligence model, and does not need to be sent to the test machine on the board.
  • the cloud server needs to send the pre-trained artificial intelligence model to the test machine, and the test machine can complete the test work of the power semiconductor device to be tested by a single machine.
  • the cloud server does not need to deliver the diagnostic result to the testing machine. Therefore, the diagnostic result obtained by the testing machine is not affected by the network failure, and the dependence on the network is low.
  • the testing system provided by the embodiments of the present application may include one testing machine, or may include multiple testing machines.
  • the working principle of the test system provided by the embodiments of the present application including multiple test machines will be described below with reference to the accompanying drawings.
  • FIG. 8 is a schematic diagram of a testing system for a power semiconductor device including a plurality of testing machines according to an embodiment of the present application.
  • the testing system for power semiconductor devices includes at least the following two testing machines: a first testing machine 201 and a second testing machine 202;
  • the cloud server 100 is specifically configured to perform training according to the first historical data sent by the first testing machine 201 and the second historical data sent by the second testing machine 202 to obtain a global artificial intelligence model, and use the global artificial intelligence intelligent model to perform training on the first testing machine 202.
  • the power semiconductor device to be tested corresponding to 201 is tested to obtain a first diagnosis result
  • the power semiconductor device to be tested corresponding to the second testing machine 202 is tested to obtain a second diagnosis result by using the global artificial intelligence model
  • the first diagnosis result is sent to the first diagnosis result.
  • a testing machine 201 sends the second diagnosis result to the second testing machine 202 .
  • the cloud server uses the first historical data and the second historical data to train the global artificial intelligence model, and can make full use of all historical data, thereby improving the generalization ability of the global artificial intelligence model, and using accurate data.
  • a more robust global artificial intelligence model for diagnosing the power semiconductor devices to be tested can improve the universality of the artificial intelligence model and increase the accuracy of diagnosing new devices.
  • the cloud server provided in the embodiment of the present application can not only directly send the diagnosis result to the testing machine on the board side, but also send the trained global artificial intelligence model to the testing machine to complete the test by each testing machine, which is not limited in the embodiment of the present application. .
  • FIG. 9 is a schematic diagram of another testing system of a power semiconductor device including multiple testing machines provided by an embodiment of the present application.
  • the power semiconductor device testing system includes at least the following two testing machines: a first testing machine 201 and a second testing machine 202 .
  • the cloud server 100 is specifically configured to perform training according to the first historical data sent by the first testing machine 201 and the second historical data sent by the second testing machine 202 to obtain a global artificial intelligence model, and send the global artificial intelligence model to the first testing machine 202.
  • Test machine 201 and second test machine 202 are specifically configured to perform training according to the first historical data sent by the first testing machine 201 and the second historical data sent by the second testing machine 202 to obtain a global artificial intelligence model, and send the global artificial intelligence model to the first testing machine 202.
  • Test machine 201 and second test machine 202 are specifically configured to perform training according to the first historical data sent by the first testing machine 201 and the second historical data sent by the second testing machine 202 to obtain a global artificial intelligence model, and send the global artificial intelligence model to the first testing machine 202.
  • Test machine 201 and second test machine 202 are specifically configured to perform training according to the first historical data sent by the first testing machine 201 and the second historical data sent by the second testing machine 202 to obtain a global artificial intelligence model, and
  • the first testing machine 201 is used to test the corresponding power semiconductor device to be tested by using the global artificial intelligence intelligent model; the second testing machine 202 is used to test the corresponding power semiconductor device to be tested by using the global artificial intelligence model.
  • the test machine can directly use the stored global artificial intelligence model to complete the task of detecting the power semiconductor device to be tested by a single machine, without the need for the cloud server in the cloud. to communicate.
  • the testing system for the power semiconductor device in the embodiment of the present application cannot obtain the diagnostic result of the power semiconductor device to be tested.
  • the test system for power semiconductor devices provided by the embodiments of the present application can reduce the dependence on network stability between the cloud server and the test machine, and improve the real-time performance of the diagnostic results obtained by the test systems for power semiconductor devices in the embodiments of the present application. sex.
  • the testing system for power semiconductor devices can obtain a global artificial intelligence model through all historical data, so as to obtain detection results by using the global artificial intelligence model.
  • the testing machine may have certain systematic errors, the systematic errors in the measured data between different testing machines are different, and this error may be related to the testing machine itself. Therefore, the embodiments of the present application also provide a testing system for power semiconductor devices. After obtaining the global artificial intelligence model, the system fine-tunes the global artificial intelligence model according to historical data obtained by different testing machines, and obtains different The local artificial intelligence model corresponding to the test machine.
  • FIG. 10 is a schematic diagram of yet another testing system of a power semiconductor device including a plurality of testing machines according to an embodiment of the present application.
  • the testing system for a power semiconductor device includes at least the following two testing machines: a first testing machine 201 and a second testing machine 202 .
  • the cloud server 100 is specifically configured to perform training according to the first historical data sent by the first testing machine 201 and the second historical data sent by the second testing machine 202 to obtain a global artificial intelligence model, and use the first historical data to analyze the global artificial intelligence model. Adjust the model to obtain a first artificial intelligence model, and use the second historical data to adjust the global artificial intelligence model to obtain a second artificial intelligence model; use the first artificial intelligence model to test the power semiconductor device to be tested corresponding to the first testing machine 201 Obtain the first diagnosis result, use the second artificial intelligence model to test the power semiconductor device to be tested corresponding to the second tester 202 to obtain the second diagnosis result, send the first diagnosis result to the first tester 201, and send the second diagnosis The results are sent to the second tester 202 .
  • the first historical data sent by the first testing machine 201 is historical data corresponding to the first testing machine 201 . That is, the first historical data includes data directly or indirectly measured by the first testing machine 201 .
  • the second historical data sent by the first testing machine 202 is historical data corresponding to the second testing machine 202 .
  • the embodiment of the present application uses the first historical data and the second historical data to train the global artificial intelligence model, which can make full use of all the historical data and improve the testing accuracy of the global artificial intelligence model.
  • the implementation of this application also uses the first historical data to adjust the global artificial intelligence model to obtain the first artificial intelligence model, and uses the second historical data to adjust the global artificial intelligence model.
  • the artificial intelligence model is adjusted to obtain a second artificial intelligence model, so that a local artificial intelligence model (first artificial intelligence model) for the characteristics of the first test machine and a local artificial intelligence model for the characteristics of the second test machine (the second artificial intelligence model) can be obtained. model), which further improves the test accuracy of the artificial intelligence model used in actual detection.
  • the testing system for power semiconductor devices includes: a testing machine Q and a testing machine W, the testing machine Q sends the historical data of the testing machine Q to the cloud server, and the testing machine W sends the historical data of the testing machine W to the cloud server.
  • Cloud Server uses the historical data of the test machine Q and the historical data of the test machine W to train a global artificial intelligence model. Then, the cloud server uses the historical data of the test machine Q to adjust the global artificial intelligence model, obtains the local artificial intelligence model corresponding to the test machine Q, and uses the local artificial intelligence model to obtain the power semiconductor device to be tested corresponding to the test machine Q. test.
  • the cloud server uses the historical data of the testing machine W to adjust the global artificial intelligence model, obtains the local artificial intelligence model corresponding to the testing machine W, and uses the local artificial intelligence model to obtain the power semiconductor device to be tested corresponding to the testing machine W for testing.
  • the cloud server provided by the embodiment of the present application can not only directly send the diagnosis results to the test machine on the board side, but also can send the trained artificial intelligence model to the test machine on the board side, and the test machine on the board side uses the trained artificial intelligence model to analyze the power semiconductors.
  • the device is tested to obtain a diagnosis result, which will be described in detail below with reference to the accompanying drawings.
  • FIG. 11 is a schematic diagram of yet another testing system of a power semiconductor device including a plurality of testing machines provided by an embodiment of the present application.
  • the testing system for power semiconductor devices includes at least the following two testing machines: a first testing machine 201 and a second testing machine 202;
  • the cloud server 100 is specifically configured to perform training according to the first historical data sent by the first testing machine 201 and the second historical data sent by the second testing machine 202 to obtain a global artificial intelligence model, and use the first historical data to perform training on the global artificial intelligence model.
  • the first artificial intelligence model obtained by adjustment is sent to the first testing machine 201
  • the global artificial intelligence model is adjusted by using the second historical data to obtain a second artificial intelligence model and sent to the second testing machine 202 .
  • the first test machine 201 is specifically used to test the corresponding power semiconductor device to be tested by using the first artificial intelligence model; the second test machine 202 is specifically used to use the second artificial intelligence model to test the corresponding power semiconductor device to be tested. test.
  • the cloud server directly delivers the first artificial intelligence model and the second artificial intelligence model to the first testing machine and the second testing machine, respectively, the testing machine can directly use the corresponding artificial intelligence model, and the single machine completes the detection pending.
  • the task of testing power semiconductor devices does not require communication with the cloud server in the cloud, which avoids that when the network between the test machine on the board and the cloud server in the cloud is interrupted, the power semiconductor device test system in the embodiment of the present application cannot be tested. diagnostic results of power semiconductor devices.
  • test system for power semiconductor devices provided by the embodiments of the present application can reduce the dependence on network stability between the cloud server and the test machine, and improve the real-time performance of the diagnostic results obtained by the test systems for power semiconductor devices in the embodiments of the present application. sex.
  • the testing system for power semiconductor devices provided by the embodiments of the present application can either directly use all historical data to obtain a global artificial intelligence model and use it directly, or use different testing machines to correspond to the global artificial intelligence model after obtaining the global artificial intelligence model.
  • the global artificial intelligence model is adjusted to obtain a local artificial intelligence model, and the local artificial intelligence model is used to detect the power semiconductor devices of its corresponding testing machine.
  • a follow-up test can also be performed on the power semiconductor device diagnosed as failure to verify whether the power semiconductor device fails, and then the AI model can be verified. Is it accurate? If it is verified that the power semiconductor device does not fail, the AI model is inaccurate and the AI model needs to be adjusted.
  • the operation data on the application side is continuously collected, and then the data of the above two kinds of power semiconductor devices are uploaded to the cloud server, and the cloud server uses these two kinds of data to adjust the AI model.
  • the testing machine in the embodiment of the present application is also used to fine-tune the pre-trained model in combination with the data on the testing machine side , using the fine-tuned model to diagnose the power semiconductor device to be tested, and output the diagnosis result.
  • the fine-tuning in this embodiment of the present application may include parameter optimization of the pre-trained model.
  • a pre-trained AI model can be fine-tuned using transfer learning.
  • the fine-tuned AI model can be used to test whether the wafer fails.
  • the data on the test machine side in the embodiment of the present application is also Can include data not uploaded to the cloud server.
  • the testing machine in the embodiment of the present application usually uploads the collected data to the cloud server on a regular basis, and the testing machine in the embodiment of the present application usually stores a certain amount of historical data that has not been uploaded. Therefore, the testing machine can use these data.
  • the AI model is fine-tuned with the unuploaded data, and then power semiconductor devices are inspected based on the fine-tuned AI model. After the testing machine in the embodiment of the present application fine-tunes the AI model, the fine-tuned AI model can also be uploaded to the cloud server, so that the cloud server can manage the AI model in a unified manner.
  • the testing system for power semiconductor devices uses a cloud server with relatively large computing and storage capabilities to train an artificial intelligence model, which can be run in the cloud.
  • the artificial intelligence model can be used in the cloud server to directly obtain the diagnosis results, and the artificial intelligence model can also be sent to the test machine on the board side, and the test machine can use the artificial intelligence model to obtain the diagnosis results, or only the artificial intelligence model that has not been fully trained can be sent.
  • the testing machine on the board side completes the training of the artificial intelligence model according to its stored data, and uses the trained artificial intelligence model to obtain the diagnosis result.
  • the embodiment of the present application further provides a cloud server.
  • FIG. 12 this figure is a schematic structural diagram of a cloud server provided by an embodiment of the present application.
  • the cloud server provided by this embodiment of the present application includes: a first transceiver device 1201 and a first controller 1202 .
  • the first transceiver 1201 is configured to receive test data of the power semiconductor device to be tested sent by the tester.
  • the first controller 1202 is used for pre-training with historical data of multiple power semiconductor devices to obtain an artificial intelligence model; it is also used to input the data of the power semiconductor devices to be tested into the artificial intelligence model, and the output of the pre-trained AI model is: The diagnosis result of the power semiconductor device to be tested; the first transceiver device is also used for sending the diagnosis result to the testing machine;
  • the first controller 1202 is configured to send the pre-trained artificial intelligence model to the testing machine, so that the testing machine uses the pre-trained artificial intelligence model to diagnose the power semiconductor device to be tested.
  • the first controller is specifically used to obtain the failure types of multiple power semiconductor devices when an artificial intelligence model is obtained by training an unsupervised learning model, and to use expert knowledge in the power semiconductor device. All the test items corresponding to each failure type are extracted from the test items of the device, and all the test items corresponding to each failure type form a test item subset, and the unsupervised learning model is used to detect the abnormal level of each test item subset, and the abnormality is obtained.
  • the device and the normal device are recorded with the first score and the second score respectively, and the sum of the first score and the second score of all test item subsets of each power semiconductor device is obtained as the total score of abnormal level; the total score of abnormal level is greater than the preset score When the threshold value is reached, the power semiconductor device is judged to be an abnormal device.
  • the cloud server corresponds to at least two of the following testing machines: a first testing machine and a second testing machine.
  • the first controller is specifically configured to perform training according to the first historical data sent by the first testing machine and the second historical data sent by the second testing machine to obtain a global artificial intelligence model, and use the first historical data to perform training on the global artificial intelligence model.
  • Adjust to obtain the first artificial intelligence model use the second historical data to adjust the global artificial intelligence model to obtain the second artificial intelligence model; use the first artificial intelligence model to test the power semiconductor device to be tested corresponding to the first testing machine to obtain the first artificial intelligence model.
  • the first transceiver device is specifically configured to send the first diagnosis result to the first test machine, It is also used for sending the second diagnosis result to the second testing machine.
  • the cloud server corresponds to at least two of the following testing machines: a first testing machine and a second testing machine.
  • the first controller is specifically configured to perform training according to the first historical data sent by the first testing machine and the second historical data sent by the second testing machine to obtain a global artificial intelligence model, and use the global artificial intelligence intelligent model to test the first
  • the first diagnostic result is obtained by testing the power semiconductor device to be tested corresponding to the test machine
  • the second diagnostic result is obtained by using the global artificial intelligence model to test the power semiconductor device to be tested corresponding to the second test machine
  • the first diagnostic result is sent to the first testing machine, and is also used for sending the second diagnostic result to the second testing machine.
  • the first transceiver device is further configured to receive update data sent by the testing machine; the first controller is further configured to update the artificial intelligence model according to the update data.
  • the cloud server provided in this embodiment of the present application may include, in addition to the first transceiver device and the first controller, a first storage device (not shown in the figure), and the first storage device may be used to store the history uploaded by the tester Data and update data are convenient for the first controller to perform subsequent processing.
  • the first controller uses historical data to train the AI model, in order to make the parameters of the trained AI model more accurate, some interfering data can be eliminated, and specifically, data cleaning can be performed.
  • the cloud server in the embodiment of the present application can use the cloud server with its powerful computing capability and storage capability to train an artificial intelligence model, thereby obtaining an artificial intelligence model with high diagnostic accuracy.
  • the artificial intelligence model can be directly used in the cloud server to obtain the diagnosis result, or the artificial intelligence model can only be provided for the test machine without providing the direct diagnosis result.
  • an embodiment of the present application further provides a testing machine, which will be described in detail below with reference to the accompanying drawings.
  • FIG. 13 this figure is a schematic structural diagram of a testing machine provided by an embodiment of the present application.
  • the testing machine in this embodiment of the present application includes: a second transceiver 1301 and a second controller 1302;
  • the second transceiver device 1301 is configured to receive an artificial intelligence model sent by the cloud server, and the artificial intelligence model is obtained by training the cloud server in advance using historical data of multiple power semiconductor devices;
  • the second controller 1302 is configured to obtain data of the power semiconductor device to be tested; input the data of the power semiconductor device to be tested into a pre-trained artificial intelligence model, and the output of the pre-trained AI model is the diagnostic result of the power semiconductor device to be tested.
  • the testing machine provided in the embodiments of the present application generally refers to any testing machine, and one cloud server can correspond to multiple testing machines provided in the embodiments of the present application.
  • one cloud server can correspond to multiple testing machines provided in the embodiments of the present application.
  • For the specific working methods of the testing machines please refer to the introduction of the testing system embodiments above for the testing machines. It is only briefly explained here.
  • the second controller is further configured to collect historical data of multiple power semiconductor devices, and send the collected historical data of multiple power semiconductor devices to a cloud server; the historical data includes multiple At least one of the chip test data of each power semiconductor device or the packaged module function test data.
  • the second controller is further configured to compare the data of the power semiconductor device to be tested with the preset upper and lower limit values, and when the data of the power semiconductor device to be tested exceeds the upper and lower limit values , judging that the power semiconductor device to be tested is an abnormal device.
  • the testing machine can directly obtain the diagnosis result sent from the cloud server, or obtain the artificial intelligence model sent by the cloud server, and obtain the diagnosis result through the artificial intelligence model in the testing machine.
  • the computing power and storage capacity of the test machine provided by the embodiments of the present application are limited, so the artificial intelligence model is trained by using the cloud server with powerful computing power and storage capacity of the cloud server by sending the data to the cloud server, thus, an artificial intelligence model with higher diagnostic accuracy can be obtained.
  • At least one (item) refers to one or more, and "a plurality” refers to two or more.
  • “And/or” is used to describe the relationship between related objects, indicating that there can be three kinds of relationships, for example, “A and/or B” can mean: only A, only B, and both A and B exist , where A and B can be singular or plural.
  • the character “/” generally indicates that the associated objects are an “or” relationship.
  • At least one item(s) below” or similar expressions thereof refer to any combination of these items, including any combination of single item(s) or plural items(s).
  • At least one (a) of a, b or c can mean: a, b, c, "a and b", “a and c", “b and c", or "a and b and c" ", where a, b, c can be single or multiple.

Landscapes

  • Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Manufacturing & Machinery (AREA)
  • Computer Hardware Design (AREA)
  • Microelectronics & Electronic Packaging (AREA)
  • Power Engineering (AREA)
  • Testing Of Individual Semiconductor Devices (AREA)

Abstract

一种功率半导体器件的测试系统、云服务器及测试机，该系统包括：云服务器（100）和至少一台测试机（200）；云服务器（100），用于预先利用多个功率半导体器件的历史数据进行训练获得人工智能模型；测试机（200），用于获得待测试功率半导体器件的数据；云服务器（100）或测试机（200），用于将待测试功率半导体器件的数据输入人工智能模型，人工智能模型的输出为待测试功率半导体器件的诊断结果。利用云服务器（100）强大的运算能力和存储能力可以预先训练获得准确的人工智能模型，从而云服务器（100）或者测试机（200）利用训练好的人工智能模型对待测试功率半导体器件进行诊断，可以获得准确的诊断结果，将异常的功率半导体器件剔除，以免在实际产品使用过程中发生故障致使整个电路故障。

Description

一种功率半导体器件的测试系统、云服务器及测试机 技术领域
本申请涉及功率半导体技术领域，尤其涉及一种功率半导体器件的测试系统、云服务器及测试机。
背景技术
功率半导体器件一般是指实现电路开关功能的半导体器件，例如可以应用于电动汽车、光伏系统或者数据中心的功率变换电路中。例如，在电动汽车中，功率半导体器件可以应用于电动汽车的动力总成控制电路中。在实际应用中，为了保证功率半导体器件所应用的电路的安全性，功率半导体器件在生产后需要进行性能检测以减少功率半导体器件的失效风险。
目前，对功率半导体器件的质量测试主要通过测试机对功率半导体器件的性能参数设置上下限数值，将性能参数超过上下限数值的功率半导体器件淘汰。
但是，目前采用上下限数值进行测试的方法需要通过大量历史产品进行确认，且通过上下限数值的筛选方法对经验的依赖程度较高，无法与器件的失效原理建立完整的关联关系，仅能识别出一部分存在问题的功率半导体器件。
发明内容
本申请提供了一种功率半导体器件的测试系统、云服务器及测试机，能够更全面筛选不良的功率半导体器件。
本申请实施例提供的功率半导体器件的测试系统，包括云服务器和至少一台测试机；云服务器用于预先利用多个功率半导体器件的历史数据进行训练获得人工智能(AI，Artificial Intelligence)模型；测试机用于获得待测试功率半导体器件的数据；云服务器或测试机均可以利用训练好的人工智能模型对待测试功率半导体器件进行诊断，即云服务器或测试机将待测试功率半导体器件的数据输入人工智能模型，人工智能模型的输出为待测试功率半导体器件的诊断结果。
本申请实施例所提供的方案，利用云服务器强大的运算能和存储能力可以训练得到较为精准的人工智能模型，然后利用预先训练好的人工智能模型可以较为精准地诊断出待测试功率半导体器件中的高风险功率半导体器件，从而弥补传统测试系统仅用上限数据来筛选高风险功率半导体器件的局限性，即可以更加精准和全面地诊断功率半导体器件是否存在风险。
在一种可能的实现方式中，本申请实施例中，云服务器用于将待测试功率半导体器件的数据输入人工智能模型，人工智能模型的输出为待测试功率半导体器件的诊断结果，将诊断结果发送给测试机。应该理解，本申请实施例中云服务器将数据输入人工智能模型，获得诊断结果，将诊断结果发送给测试机，因此测试机可以直接获得测试结果，节省了测试机的计算资源。
在一种可能的实现方式中，本申请实施例中云服务器用于将预先训练的人工智能模型发送给测试机；测试机将待测试功率半导体器件的数据输入人工智能模型，人工智能模型的输出为待测试功率半导体器件的诊断结果。应该理解，云服务器将训练好的人工智能 模型发送给测试机，然后由测试机来输入数据并获得诊断结果。因此，测试机单机完成待测试功率半导体器件的测试工作，不需要从云服务器获得诊断结果，减少了云服务器和测试机之间对于网络的依赖，提高了本申请实施例中功率半导体器件的测试系统获得诊断结果的实时性。
在一种可能的实现方式中，本申请实施例中的测试机还用于收集多个功率半导体器件的历史数据，并将收集的多个功率半导体器件的历史数据发送给云服务器；历史数据包括多个功率半导体器件的芯片测试数据、封装后的模组功能测试数据或实际应用异常时的相关数据中的至少一项；云服务器还用于利用多个功率半导体器件的历史数据利用有监督学习模型或无监督学习模型中的至少一种进行训练获得人工智能模型。
在一种可能的实现方式中，云服务器利用无监督学习模型训练获得人工智能模型，具体用于获得多个功率半导体器件的失效类型，利用专家知识在功率半导体器件的测试项中提取每个失效类型对应的所有测试项，每个失效类型对应的所有测试项形成测试项子集，对每个测试项子集利用无监督学习模型进行异常子集和正常子集区分，得到异常子集和正常子集分别记第一分数和第二分数，获得每个功率半导体器件的所有测试项子集的第一分数和第二分数之和作为异常水平总分；异常水平总分大于或等于预设分数阈值时，判断该功率半导体器件为异常器件。应该理解，本申请实施例采用无监督学习模型训练AI模型，可以充分利用专家知识，将包含多个测试项的测试项子集共同进行异常点检测，可以利用同一个测试项子集中器件之间的联系，或同一个失效类型对应的物理基理，从多种维度检测功率半导体器件失效的风险，从而可以更加准确地评估器件失效的风险。
在一种可能的实现方式中，本申请实施例中的云服务器具体用于利用专家知识为不同的测试项子集中的异常器件记不同的第一分数。应该理解，不同的测试项子集对判断异常器件的权重可能不同，将不同的测试项子集中的异常器件记不同的第一分数，可以获得更为精准的诊断结果。
在一种可能的实现方式中，本申请实施例中云服务器利用有监督学习模型训练获得人工智能模型，具体用于获得功率半导体器件异常时的相关数据作为数据标签，利用有监督学习模型提取数据标签的数据特征，利用数据特征获得每个功率半导体器件的异常水平总分，异常水平总分大于预设分数阈值时，判断该功率半导体器件为异常器件。
在一种可能的实现方式中，本申请实施例中的测试机还用于将待测试功率半导体器件的数据与预设的上下限数值进行比较，待测试功率半导体器件的数据超过上下限数值时，判断该待测试功率半导体器件为异常器件。应该理解，本申请实施例所提供的功率半导体器件的测试系统，可以将传统的测试方法与人工智能模型相结合，在传统测试方法的基础上利用本申请实施例中的人工智能模型进行二次检测，提高检测的准确率，实现传统方式和AI模型测试方式之间的互补，可以更全面诊断出高风险的功率半导体器件。
本申请实施例不限定测试系统中包括的测试机的数量，可以包括一台测试机，也可以包括两台或更多台测试机。当测试系统包括多台测试机时，云服务器可以先训练处全局AI模型，然后根据各个测试机之间的差异，在全局AI模型的基础上获得适用于各个测试机的局部AI模型。即在一种可能的实现方式中，本申请实施例提供的测试 系统至少包括以下两台测试机：第一测试机和第二测试机；云服务器具体用于根据第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，利用第一历史数据对全局人工智能模型进行调整获得第一人工智能模型，利用第二历史数据对全局人工智能模型进行调整获得第二人工智能模型；利用第一人工智能模型对第一测试机对应的待测试功率半导体器件进行测试获得第一诊断结果，利用第二人工智能模型对第二测试机对应的待测试功率半导体器件进行测试获得第二诊断结果，将第一诊断结果发送给第一测试机，将第二诊断结果发送给第二测试机。本申请实施例利用第一历史数据和第二历史数据训练全局人工智能模型，可以充分的利用所有的历史数据，提高全局人工智能模型的测试精准性。在此基础上，考虑到不同的测试机之间的测试性能可能存在差异，本申请实施还利用第一历史数据对全局人工智能模型进行调整获得第一人工智能模型，利用第二历史数据对全局人工智能模型进行调整获得第二人工智能模型，从而可以得到针对第一测试机特性的局部人工智能模型(第一人工智能模型)和针对第二测试机特性的局部人工智能模型(第二人工智能模型)，进一步提高了实际检测时使用的人工智能模型的测试精准性。
本实施例中，各个测试机对应的均为全局AI模型，且由云服务器获得诊断结果。即在一种可能的实现方式中，本申请实施例提供的测试系统至少包括以下两台测试机：第一测试机和第二测试机；云服务器，具体用于根据第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，利用全局人工智能智能模型对第一测试机对应的待测试功率半导体器件进行测试获得第一诊断结果，利用全局人工智能模型对第二测试机对应的待测试功率半导体器件进行测试获得第二诊断结果，将第一诊断结果发送给第一测试机，将第二诊断结果发送给第二测试机。本申请实施例提供的云服务器利用第一历史数据和第二历史数据训练全局人工智能模型，可以充分利用所有的历史数据，从而提高全局人工智能模型的泛化能力，利用精准性更高的全局人工智能模型对待测试功率半导体器件进行诊断，可以提高人工智能模型的普适性，并增加对新器件诊断的精准性。
本实施例中，云服务器获得各个测试机对应的局部AI模型分别发送给各个测试机，由各个测试机完成诊断。即在一种可能的实现方式中，本申请实施例中的功率半导体器件的测试系统至少包括以下两台测试机：第一测试机和第二测试机；云服务器具体用于根据第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，利用第一历史数据对全局人工智能模型进行调整获得第一人工智能模型发送给第一测试机，利用第二历史数据对全局人工智能模型进行调整获得第二人工智能模型发送给第二测试机；第一测试机，具体用于利用第一人工智能模型对对应的待测试功率半导体器件进行诊断；第二测试机，具体用于利用第二人工智能模型对对应的待测试功率半导体器件进行诊断。本申请实施例的测试机可以直接利用人工智能模型，单机完成检测待测试功率半导体器件的任务，不需要与云端的云服务器进行通信，因此可以避免板端的测试机与云端的云服务器之间网络中断时，本申请实施例中的功率半导体器件的测试系统无法得到待测试的功率半导体器件的诊断结果。如此，本申请实施例所提供的功率半导体器件 的测试系统，可以减少云服务器和测试机之间的网络依赖，提高了本申请实施例中功率半导体器件的测试系统获得诊断结果的实时性。
本实施例中，云服务器将全局AI模型发送给各个测试机，由各个测试机利用全局AI模型完成诊断。即在一种可能的实现方式中，本申请实施例中的功率半导体器件的测试系统至少包括以下两台测试机：第一测试机和第二测试机；云服务器，具体用于根据第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，将全局人工智能模型发送给第一测试机和第二测试机；第一测试机，用于利用全局人工智能智能模型对对应的待测试功率半导体器件进行诊断；第二测试机，用于利用全局人工智能模型对对应的待测试功率半导体器件进行诊断。本申请实施例中的测试机根据训练好的全局人工智能模型，可以单机完成待测试功率半导体器件的测试工作。当由测试机获得诊断结果时，云服务器不需要下发诊断结果给测试机，因此，测试机获得诊断结果不受网络故障的影响，对网络的依赖性较低。
在一种可能的实现方式中，本申请实施例中的测试机还用于向云服务器发送更新数据；云服务器还用于根据更新数据更新人工智能模型。另外，测试机还用于结合测试机侧的数据对预先训练的AI模型进行微调，利用微调后的模型对待测试功率半导体器件进行诊断，输出诊断结果。应该理解，本申请实施例中测试机可以基于自身的数据或数据的特性对获得的AI模型进行微调，如此可进一步提高本申请实施例中的AI模型的准确性。测试机可以将微调后的AI模型发送给云服务器，由云服务器进行统一管理。
基于上述实施例提供的功率半导体器件的测试系统，本申请实施例还提供了一种云服务器，以上测试系统各个实施例的优点同样适用于以下的服务器，不再赘述。该服务器包括：第一收发设备和第一控制器；第一收发设备，用于接收测试机发送的待测试功率半导体器件的测试数据；第一控制器，用于预先利用多个功率半导体器件的历史数据进行训练获得人工智能模型；还用于利用将待测试功率半导体器件的数据输入人工智能模型，人工智能模型的输出为待测试功率半导体器件的诊断结果；第一收发设备，还用于将诊断结果发送给测试机；或，第一控制器，用于将人工智能模型发送给测试机，以使测试机利用人工智能模型对待测试功率半导体器件进行诊断。
在一种可能的方式中，本申请实施例中的第一控制器，具体用于利用无监督学习模型训练获得人工智能模型时，具体用于获得多个功率半导体器件的失效类型，利用专家知识在功率半导体器件的测试项中提取每个失效类型对应的所有测试项，每个失效类型对应的所有测试项形成测试项子集，对每个测试项子集利用无监督学习模型进行异常水平检测，得到异常器件和正常器件分别记第一分数和第二分数，获得每个功率半导体器件的所有测试项子集的第一分数和第二分数之和作为异常水平总分；异常水平总分大于预设分数阈值时，判断该功率半导体器件为异常器件。
在一种可能的方式中，本申请实施例中的云服务器对应以下至少两台测试机：第一测试机和第二测试机；第一控制器，具体用于根据第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，利用第一历史数据对全局人工智能模型进行调整获得第一人工智能模型，利用第二历史数据对全局人工智能模型进行调 整获得第二人工智能模型；利用第一人工智能模型对第一测试机对应的待测试功率半导体器件进行测试获得第一诊断结果，利用第二人工智能模型对第二测试机对应的待测试功率半导体器件进行测试获得第二诊断结果；第一收发设备，具体用于将第一诊断结果发送给第一测试机，还用于将第二诊断结果发送给第二测试机。
在一种可能的方式中，本申请实施例中的云服务器对应以下至少两台测试机：第一测试机和第二测试机；第一控制器，具体用于根据第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，利用全局人工智能智能模型对第一测试机对应的待测试功率半导体器件进行测试获得第一诊断结果，利用全局人工智能模型对第二测试机对应的待测试功率半导体器件进行测试获得第二诊断结果；第一收发设备，具体用于将第一诊断结果发送给第一测试机，还用于将第二诊断结果发送给第二测试机。
在一种可能的方式中，本申请实施例中的第一收发设备，还用于接收测试机发送的更新数据；第一控制器，还用于根据更新数据更新人工智能模型。
基于以上实施例提供的功率半导体器件的测试系统和服务器，本申请实施例还提供了一种测试机，以上测试系统各个实施例的优点同样适用于以下的测试机，不再赘述。该测试机包括：第二收发设备和第二控制器；第二收发设备，用于接收云服务器发送的人工智能模型，人工智能模型由云服务器预先利用多个功率半导体器件的历史数据进行训练获得；第二控制器，用于获得待测试功率半导体器件的数据；将待测试功率半导体器件的数据输入人工智能模型，人工智能模型的输出为待测试功率半导体器件的诊断结果。
在一种可能的方式中，本申请实施例中的第二控制器，还用于收集多个功率半导体器件的历史数据，并将收集的多个功率半导体器件的历史数据发送给云服务器；历史数据包括多个功率半导体器件的芯片测试数据或封装后的模组功能测试数据中的至少一项。
在一种可能的方式中，本申请实施例中的第二控制器，还用于将待测试功率半导体器件的数据与预设的上下限数值进行比较，待测试功率半导体器件的数据超过上下限数值时，判断该待测试功率半导体器件为异常器件。
从以上技术方案可以看出，本申请实施例具有以下优点：
本申请实施例提供的测试系统包括云服务器和测试机，云服务器的计算能力和存储能力均高于测试机，因此利用云服务器强大的运算能力和存储能力预先训练获得人工智能模型。由于云服务器预先利用功率半导体器件大量的历史数据训练人工智能模型，可以准确获得人工智能模型的参数，因此，云服务器或者测试机利用训练好的人工智能模型可以准确地对待测试功率半导体器件进行诊断，从而获得准确的诊断结果，将异常的功率半导体器件剔除，以免在实际产品使用过程中发生故障致使整个电路故障。该测试系统可以弥补传统方式的弊端，即传统仅利用上下限数值来筛选高风险功率半导体器件，有些无法筛选出来，利用本申请实施例提供的测试系统可以进一步筛选出高风险功率半导体器件，即可以更全面地诊断功率半导体器件是否存在风险。该方案可以弥补测试机利用上下限数值简单地对功率半导体器件进行不良品筛选的不准确性。
附图说明
图1为本申请实施例提供的一种功率半导体器件的测试系统的架构图；
图2为本申请实施提供的一种功率半导体器件的数据来源示意图；
图3A为本申请实施例提供的一种功率半导体器件的测试方法流程图；
图3B为本申请实施例提供的另一种功率半导体器件测试方法的流程图；
图4为本申请实施例提供的另一种功率半导体器件的测试系统示意图；
图5为本申请实施例提供的又一种功率半导体器件的测试系统的示意图；
图6为本申请实施例提供的另一种功率半导体器件的测试系统的示意图；
图7为本申请实施例提供的又一种功率半导体器件的测试系统的示意图；
图8为本申请实施例提供的一种包含多个测试机的功率半导体器件的测试系统示意图；
图9为本申请实施例提供的另一种包含多个测试机的功率半导体器件的测试系统示意图；
图10为本申请实施例提供的再一种包含多个测试机的功率半导体器件的测试系统示意图；
图11为本申请实施例提供的再一种包含多个测试机的功率半导体器件的测试系统示意图；
图12为本申请实施例提供的一种云服务器的结构示意图；
图13为本申请实施例提供的一种测试机的结构示意图。
具体实施方式
以下说明中的“第一”、“第二”等用词仅用于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。由此，限定有“第一”、“第二”等的特征可以明示或者隐含地包括一个或者更多该特征。在本申请的描述中，除非另有说明，“多个”的含义是两个或两个以上。
此外，本申请中，“上”、“下”等方位术语可以包括但不限于相对附图中的部件示意置放的方位来定义的，应当理解到，这些方向性术语可以是相对的概念，它们用于相对于的描述和澄清，其可以根据附图中部件附图所放置的方位的变化而相应地发生变化。
在本申请中，除非另有明确的规定和限定，术语“连接”应做广义理解，例如，“连接”可以是固定连接，也可以是可拆卸连接，或成一体；可以是直接相连，也可以通过中间媒介间接相连。此外，术语“耦接”可以是实现信号传输的电性连接的方式。“耦接”可以是直接的电性连接，也可以通过中间媒介间接电性连接。
本申请实施例涉及一种功率半导体器件的测试系统，其中功率半导体器件的测试系统包括：云服务器和至少一台测试机。云服务器具有强大的计算能量和存储空间，因此，云服务器可以利用功率半导体器件的大量历史数据训练人工智能模型。由于AI模块是经过大量历史数据训练得到的，因此，云服务器或者测试机利用训练好的AI模型可以准确对待测试功率半导体器件进行诊断，将异常的功率半导体器件剔除，以免其在真实使用过程中发生故障致使整个电路故障。该方案可以解决传统方案中仅利用测试机预先设定的上下限数值简单地对待测试功率半导体器件进行不良品筛选的不准确性。
为了使本领域技术人员更好地理解本申请实施例提供的技术方案，下面结合附图详 细介绍本申请实施例提供的测试系统，该测试系统包括云服务器和至少一台测试机。
本申请实施例不具体限定测试机的数量，可以根据实际产品线的数量来设置，一台测试机可以对应一个产品线，一个产品线也可以对应多个测试机。其中，多个产品线对应的功率半导体器件的类型可以相同，也可以不相同。下面结合附图1以三台测试机为例进行介绍，应该理解，也可以为一台测试机，也可以为两台测试机，也可以为更多台测试机。
系统实施例
参见图1，该图为本申请实施例提供的一种功率半导体器件的测试系统架构图。
如图1所示，本申请实施例提供的功率半导体器件的测试系统中，包括云服务器100、第一测试机201、第二测试机202和第三测试机203。其中，云服务100位于云平台，云平台可以包括多个云服务器，图中仅示意一个云服务器为例。板端的第一测试机201、第二测试机202和第三测试机203分别可以与云端的云服务器100通信，进行数据交换或传输。
其中，云服务器100，用于预先利用多个功率半导体器件的历史数据进行训练获得AI模型。
应该理解，本申请实施例不具体限定云服务器100进行AI模型训练利用的具体模型，例如可以利用有监督学习模型，也可以利用无监督学习模型。云服务器100训练AI模型利用的历史数据可以为第一测试机201、第二测试机202和第三测试机203上传的数据，也可以为更多测试机上传的数据，本申请实施例中均不做具体限定。应该理解，云服务器100进行AI模型训练时，使用的历史数据越多，则训练得到的AI模型的参数越准确。
第一测试机201、第二测试机202和第三测试机203，用于获得各自对应的待测试功率半导体器件的数据。
具体测试时，可以由云服务器100来进行测试，也可以云服务器100将训练好的AI模型下发给测试机，由测试机来进行测试。
第一种，云服务器100进行测试。
云服务器100，用于将待测试功率半导体器件的数据输入预先训练的人工智能模型，人工智能模型的输出为待测试功率半导体器件的诊断结果。将诊断结果发送给测试机。
应该理解，此处的待测试功率半导体器件泛指任意一个需要测试的功率半导体器件；此处的测试机泛指任意一个测试机。
第二种：测试机进行测试。
第一测试机201、第二测试机202和第三测试机203，分别用于将各自对应的待测试功率半导体器件的数据输入预先训练的人工智能模型，分别获得预先训练的模型的输出为各自对应的待测试功率半导体器件的诊断结果。
本申请实施例不限定第一测试机201、第二测试机202和第三测试机203同时对各自对应的待测试功率半导体器件进行测试，还是分别依次进行测试，因为每个测试机测试动作可以独立完成，互不交叉干扰。
可以理解的是，本申请实施例为了提高对功率半导体器件不良品进行拦截的精准度，利用人工智能模型来测试功率半导体器件的性能。但由于板端的测试机的计算能力和存储 能力有限，如果利用板端的测试机训练人工智能模型耗时较长且会占用过多的计算资源，影响测试机的其他功能。因此，本申请实施例提供的功率半导体器件的测试系统，利用云服务器强大的运算能力和存储能力，通过多个功率半导体器件的历史数据训练得到人工智能模型，然后通过该人工智能模型对待测试功率半导体器件进行检测。云服务器训练的AI模型，可以下发给测试机，由测试机利用云服务器已经训练好的AI模型对待测试功率半导体器件进行测试，最终输出诊断结果。
由此可知，本申请实施例提供的功率半导体器件的测试系统，可以利用人工智能模型诊断功率半导体的性能，筛选功率半导体器件中的不良品。且考虑到板端的测试机的计算能力和存储能力有限，本申请利用具有强大计算能力和存储能力的云服务器对人工智能模型进行训练，从而获得诊断精准度较高的人工智能模型。
上述的内容主要介绍本申请实施例提供的功率半导体器件的测试系统的架构，下面结合附图介绍云服务器进行AI模型训练的具体过程。
参见图2，该图为本申请实施提供的一种功率半导体器件的数据来源示意图。
本申请实施例中，云服务器利用大量历史数据用于训练获得AI模型。历史数据可以包括产线数据和应用侧数据。
其中，产线数据D1的来源可以包括以下几种测试获得的数据：芯片(CP，Chip Probe)测试、功能测试(FT，Functional Test)、单板测试和整机测试。其中，CP测试可以为单个芯片的测试，FT测试可以为封装后的模组的测试，该模组具有一定的电路功能。单板测试可以为上板测试，整机测试可以为上机测试。可以理解，芯片测试数据和功率测试数据可以分别单独使用进行AI模型训练，也可以联合一起使用进行AI模型训练。
需要说明的是，CP测试主要针对单个芯片，单个芯片可以为单独的绝缘栅双极型晶体管(IGBT，Insulated Gate Bipolar Transistor，)、单独的金属-氧化物半导体场效应晶体管(MOSFET，Metal-Oxide-Semiconductor Field-Effect Transistor)的芯片、二极管(Doide)、三极管(Bipolar Junction Transistor)、晶闸管(Thyristor)、集成门极换流晶闸管(IGCT，Integrated Gate-Commutated Thyristor)等功率半导体器件，还可以为多个IGBT形成的芯片，还可以为IGBT和二极管形成的芯片。FT测试主要针对于一个封装后的模组，
应用侧数据D2主要包括功率半导体器件生产后，在实际使用过程中出现失效时对应的相关数据，即客户侧反馈的功率半导体器件的实际工况数据，可以包括测试数据来源，也可以包括验证数据来源
作为一个示例，本申请实施例中待测试功率半导体器件的测试数据可以包括但不限于以下的击穿电压、漏电流、导通压降、寄生电容、门级电荷、开通损耗和关断损耗。
下面介绍本申请所提供的功率半导体器件的测试系统利用测试数据获得诊断结果的原理。
本申请实施例提供云服务器训练获得的AI模型，可以为利用无监督集成学习模型训练得到的，也可以为利用有监督学习模型训练得到的；也可以为利用有监督学习模型和无监督学习模型联合得到的，本申请实施例中不做具体限定
下面先介绍本申请实施例提供AI模型利用无监督集成学习模型训练获得。
作为一个示例，本申请实施例中的无监督学习模型可以为无监督集成学习模型。可以理解的是，本申请实施例中的无监督模型采用无监督学习算法和专家知识结合构建无监督集成学习模型，可以充分利用专家经验，降低失效样本过少对人工智能模型准确性的影响。
本申请实施例中云服务器利用无监督学习模型训练获得人工智能模型时，作为一种可能的实施方式，云服务器具体用于获得多个功率半导体器件的失效类型，利用专家知识在功率半导体器件的测试项中提取每个失效类型对应的所有测试项，每个失效类型对应的所有测试项形成测试项子集，对每个测试项子集利用无监督学习模型进行异常子集和正常子集区分，得到异常子集和正常子集分别记第一分数和第二分数，获得每个功率半导体器件的所有测试项子集的第一分数和第二分数之和作为异常水平总分；异常水平总分大于或等于预设分数阈值时，判断该功率半导体器件为异常器件。
其中，失效类型具体可以为功率半导体器件单项测试时卡控无法拦截的各种失效类型。例如，实际使用中功率半导体器件出现了早期故障，但是可控测试时并未筛选出来。
另外，一个失效类型可能对应很多测试项均存在问题，例如门极漏电，可能是物理层出现问题，涉及很多个测试项都由物理层出现问题引起。因此，将一个失效类型关联的所有测试项均提取出来，形成该测试项对应的测试项子集。
作为一个示例，下面结合表1和表2具体介绍本申请实施例提供的利用无监督集成学习模型训练获得人工智能模型的方案。
表1
Figure PCTCN2021091550-appb-000001
表2
  Subset1 Subset2 Subset3 …… 异常水平总分
器件A 0 0 0   0
器件B 0 1 0   1
器件C 0 0 0   0
……         ……
其中，Gate失效、PN结失效或耐压环失效分别为不同的失效类型；IGES(GE漏电流)、VTH(阈值电压)、VCESAT(C极和E极之间的导通压降)、ICES(CE漏电流)和Eon(开通损耗)等为不同的测试项，从表1可以看出，不同的失效类型包括不同的测试项，而且每个失效类型包括多个测试项，因此，本申请实施例在AI模型训练时，使用的数据均为多维数据，并不是传统的一维数据，因此，训练获得的AI模型更加准确，更加全面，最终利用训练好的AI模型进行诊断时，输出的诊断结果也更加准确。
其中，器件A、器件B和器件C分别为不同的功率半导体器件。Subset1、Subset2和 Subset3分别为不同的测试项子集；即每个器件至少包括三个测试项子集。每个测试项子集包括的测试项的内容和数量有所区别。表中以异常子集记第二分数为1，正常子集记第一分数为0，获得每个功率半导体器件的所有测试项子集的第一分数和第二分数之和作为异常水平总分；即器件A的异常水平总分为0，器件B的异常水平总分为1，器件C的异常水平总分为0。如果异常水平总分大于或等于预设分数阈值时，判断该功率半导体器件为异常器件。例如预设分数阈值为1，则测试器件B为异常器件。
参见图3A，该图为本申请实施例提供的一种功率半导体器件的测试方法流程图。
本申请实施例提供的一种功率半导体器件的测试方法，包括：
S301：获得多个功率半导体器件的失效类型。
作为一个示例，可以获得Gate失效、PN结失效或耐压环失效等失效类型。可以理解的是，当同一种功率半导体器件的历史数据越多时，收集的失效类型越多，越全面。
S302：利用专家知识在功率半导体器件的测试项中提取每个失效类型对应的所有测试项，每个失效类型对应的所有测试项形成测试项子集。
在该示例中，Gate失效对应着IGES、VTH、VCESAT和Eon等多个测试项，这些测试项形成一个测试项子集subset1。PN结失效对应着VCESAT和ICES等测试项，这些测试项形成一个测试项子集subset2。耐压环失效对应着ICES等测试项，这些测试项形成一个测试项子集subset3。
S303：对每个测试项子集利用无监督学习模型进行异常子集和正常子集区分，得到异常子集和正常子集分别记第一分数和第二分数。
例如对Gate失效对应的测试项子集利用无监督学习模型进行异常点检测，得到的正常器件A、正常器件B和正常器件C分别记第二分数。对PN结失效对应的测试项子集利用无监督学习模型进行异常点检测，得到的正常器件A和正常器件C分别记第二分数，异常器件B记第一分数。在表2的示例中，第一分数为1，第二分数为0。
在实际的应用中，为了更加精准地评估每个器件失效的风险，作为一种可能的实施方式，云服务器还利用专家知识为不同的测试项子集中的异常器件记不同的第一分数。即在第一分数和第二分数的基础上添加测试项子集对应的权重。例如，一共有三个失效类型，三个失效类型分别对应a、b和c三个测试项子集，a测试子集的权重为0.2，b测试子集的权重为0.3，c测试子集的权重为0.5。相应地，如果b测试子集中的器件A和器件B正常，那么器件A和器件B分别可以记第二分数0。器件C出现异常，器件C可以记初始分数1*b测试子集的权重0.3＝第一分数0.3。
S304：获得每个功率半导体器件的所有测试项子集的第一分数和第二分数之和作为异常水平总分。
在表1和表2对应的示例中，器件A的异常水平总分为测试项子集subset1的第二分数0、测试项子集subset2的第二分数0和测试项子集subset3的第二分数0之和，为0。器件B的异常水平总分为测试项子集subset1的第二分数0、测试项子集subset2的第一分数1测试项子集subset3的第二分数0之和，为1。器件C的异常水平总分为测试项子集subset1的第二分数0、测试项子集subset2的第二分数0和测试项子集subset3的第二分数0之和， 为0。
S305：异常水平总分大于或等于预设分数阈值时，判断该功率半导体器件为异常器件。
在该示例中，预设分数阈值可以为1，当然，也可以根据实际情况取得其他数值，本申请实施在此不做限定。当预设分数阈值为1，由于器件B的异常水平总分为1大于或等于预设分数阈值，则判断器件B为异常器件。
由此可知，本申请实施例通过将每种失效类型对应多个测试项，将包含多个测试项的测试项子集共同进行异常点检测，可以利用同一个测试项子集中的器件之间的联系，或同一个失效类型对应的物理基理，从多种维度检测功率半导体器件失效的风险，从而可以更加准确地评估器件失效的风险。
上述实施例中介绍的方案是云服务器通过无监督学习模型训练得到人工智能模型，作为另一种可能的实施方式，接下来介绍云服务器通过有监督学习模型训练得到人工智能模型。
参见图3B，该图为本申请实施例提供的另一种功率半导体器件测试方法的流程图。
本申请实施例提供的功率半导体器件的测试方法，包括：
S311：获得功率半导体器件异常时的相关数据作为数据标签。例如，可以选择极少量的失效样本作为数据标签。
S312：利用有监督学习模型提取数据标签的数据特征。
S313：利用数据特征获得每个功率半导体器件的异常水平总分。
S314：异常水平总分大于预设分数阈值时，判断该功率半导体器件为异常器件。由此可知，本申请实施例提供的方案可以通过有监督学习模型或无监督学习模型，获得诊断功率半导体器件的人工智能模型。
应该理解，本申请实施例中的利用人工智能模型进行对功率半导体器件进行不良品筛选的方法，是在传统设置的上下限数值诊断方法的基础上又改进的方案。即本申请实施例中的功率半导体器件测试方法，是测试机利用预先设定的上下限数值先对对待测试功率半导体器件进行诊断，即测试机，还用于将待测试功率半导体器件的数据与预设的上下限数值进行比较，待测试功率半导体器件的数据超过上下限数值时，判断该待测试功率半导体器件为异常器件。然后测试机或云服务器利用AI模型再对待测试功率半导体器件进行诊断，两种方案的诊断结果进行逻辑或运算。因此，本申请实施例所提供的测试系统可以将两种方案进行互补，进而更全面地诊断待测试功率半导体器件。
可以理解的是，本申请实施例所提供的功率半导体器件的测试系统，在利用人工智能模型诊断功率半导体器件的同时，还可以利用传统方法将待测试功率半导体器件的数据与预设的上下限数值进行比较，从而诊断功率半导体器件。作为一种可能的实施方式，当人工智能模型的方案和上下限数值的方案中至少一个方案诊断功率半导体器件失效时，则判断该功率半导体器件失效，即两种诊断结果进行逻辑或运算。如此，本申请实施例所提供的功率半导体器件的测试系统，可以将传统的测试方法与人工智能模型相结合，进一步提高检测的准确率。
例如，功率半导体器件X的数据未超出预设的上下限数据，但功率半导体器件X的数 据输入人工智能模型后，人工智能模型输出的诊断结果为失效，此时则判断该功率半导体器件X失效。功率半导体器件Y的数据超出预设的上下限数据，但功率半导体器件Y的数据输入人工智能模型后，人工智能模型输出的诊断结果为未失效，此时则判断该功率半导体器件Y失效。功率半导体器件Z的数据超出预设的上下限数据，且功率半导体器件Z的数据输入人工智能模型后，人工智能模型输出的诊断结果为失效，此时则判断该功率半导体器件Z失效。
在本申请实施例提供的功率半导体器件的测试系统中，为了使人工智能模型的诊断更加精准，作为一种可能的实施方式，测试机还用于向云服务器发送更新数据；云服务器，还用于根据更新数据更新人工智能模型。
需要说明的是，本申请实施例中将待测试的功率半导体器件的数据输入预先训练的人工智能模型并获得诊断结果的步骤，可以由云端的云服务器完成，也可以由板端的测试机完成，本申请实施例在此不做限定。下面将通过实施例具体介绍两种不同的实现方式。
参见图4，该图为本申请实施例提供的另一种功率半导体器件的测试系统示意图。
测试机200还用于收集多个功率半导体器件的历史数据，并将收集的多个功率半导体器件的历史数据发送给云服务器100，云服务器100可以通过多个功率半导体器件的历史数据训练人工智能模型。
云服务器100还用于利用多个功率半导体器件的历史数据使用有监督学习模型或无监督学习模型中的至少一种进行训练获得人工智能模型。
本申请实施例中，历史数据可以包括多个功率半导体器件的芯片测试数据、封装后的模组功能测试数据或实际应用异常时的相关数据中的至少一项。具体地，作为一种可能的实施方式，历史数据可以包括图2中的训练数据中的一项或多项。
需要说明的是，在本申请实施例中训练获得的人工智能模型存储在云服务器中，当需要利用该人工智能模型检测待测试的功率半导体器件时，由云服务器利用该人工智能模型对待测试的功率半导体器件进行检测，并获得检测的结果。
下面将结合附图具体介绍在该示例中云服务器利用人工智能模型检测待测试的功率半导体器件的具体方案。
参见图5，该图为本申请实施例提供的又一种功率半导体器件的测试系统的示意图。
云服务器100用于将从测试机获得的待测试功率半导体器件的数据输入预先训练的人工智能模型，预先训练的人工智能模型的输出为待测试功率半导体器件的诊断结果。然后，云服务器100将诊断结果发送给测试机200，最终由测试机200输出诊断结果。
可以理解的是，在本申请实施例提供的方案中，云服务器100将存储其训练的人工智能模型。当需要利用人工智能模型检测待测试的功率半导体器件时，云服务器100可以从测试机200处获得待测试的功率半导体器件的数据，并将这些数据输入预先训练的人工智能模型中，获得诊断结果。然后，云服务器100再将所获得的诊断结果发送给测试机200，使得在端板的测试机200输出诊断结果，即测试机200仅接收诊断结果，自身并不对功率半导体器件进行测试。
作为另一种可能的实施方式，本申请实施例中将待测试的功率半导体的数据输入预选 训练的人工智能模型并获得诊断结果的步骤，也可以由板端的测试机完成。下面先介绍AI模型的训练过程。
参见图6，该图为本申请实施例提供的另一种功率半导体器件的测试系统的示意图。
测试机200还用于收集多个功率半导体器件的历史数据，并将收集的多个功率半导体器件的历史数据发送给云服务器100，从而使得云服务器100可以通过多个功率半导体器件的历史数据训练人工智能模型。历史数据包括上述实施例中的多个功率半导体器件的芯片测试数据或封装后的模组功能测试数据中的至少一项。需要说明的是，在本申请实施例中训练获得的人工智能模型后，云服务器100将获得的人工智能模型发送给板端的测试机200，测试机200对人工智能模型进行存储。
下面将结合附图具体介绍在该示例中测试机利用人工智能模型检测待测试的功率半导体器件的具体方案。
参见图7，该图为本申请实施例提供的又一种功率半导体器件的测试系统的示意图。
云服务器100用于将预先训练的人工智能模型发送给测试机200。测试机200用于将待测试功率半导体器件的数据输入预先训练的人工智能模型，预先训练的人工智能模型的输出为待测试功率半导体器件的诊断结果。
可以理解的是，在本申请实施例提供的方案中，测试机200将接收云服务器100发送的人工智能模型，并存储人工智能模型。当需要利用人工智能模型检测待测试的功率半导体器件时，测试机200将待测试的功率半导体器件的数据输入人工智能模型中，获得并输出诊断结果。
需要说明的是，本申请实施例中由于人工智能模型存储在测试机中，测试机可以直接利用人工智能模型，单机完成检测待测试功率半导体器件的任务，不需要与云端的云服务器进行通信，因此可以避免板端的测试机与云端的云服务器之间网络中断时，本申请实施例中的功率半导体器件的测试系统无法得到待测试的功率半导体器件的诊断结果。如此，本申请实施例所提供的功率半导体器件的测试系统，可以减少云服务器和测试机之间的网络依赖，提高了本申请实施例中功率半导体器件的测试系统获得诊断结果的实时性。
综上所述，本申请实施例提供的测试系统可以由云端的云服务器将待测试的功率半导体的数据输入预选训练的人工智能模型并获得诊断结果，也可以由板端的测试机将待测试的功率半导体的数据输入预选训练的人工智能模型并获得诊断结果。由云服务器获得诊断结果时，云服务器可以直接存储训练好的人工智能模型，不需要下发给板端的测试机。由板端的测试机获得诊断结果时，云服务器需要将预先训练的人工智能模型下发给测试机，测试机可以单机完成待测试功率半导体器件的测试工作。当由测试机获得诊断结果时，云服务器不需要下发诊断结果给测试机，因此，测试机获得诊断结果不受网络故障的影响，对网络的依赖性较低。
本申请实施例提供的测试系统可以包括一台测试机，也可以包括多台测试机。下面结合附图介绍本申请实施例提供的测试系统包括多台测试机的工作原理。
参见图8，该图为本申请实施例提供的一种包含多个测试机的功率半导体器件的测试系统示意图。
本申请实施例所提供的功率半导体器件的测试系统至少包括以下两台测试机：第一测试机201和第二测试机202；
云服务器100，具体用于根据第一测试机201发送的第一历史数据和第二测试机202发送的第二历史数据进行训练获得全局人工智能模型，利用全局人工智能智能模型对第一测试机201对应的待测试功率半导体器件进行测试获得第一诊断结果，利用全局人工智能模型对第二测试机202对应的待测试功率半导体器件进行测试获得第二诊断结果，将第一诊断结果发送给第一测试机201，将第二诊断结果发送给第二测试机202。
可以理解的是，本申请实施例提供的云服务器利用第一历史数据和第二历史数据训练全局人工智能模型，可以充分利用所有的历史数据，从而提高全局人工智能模型的泛化能力，利用精准性更高的全局人工智能模型对待测试功率半导体器件进行诊断，可以提高人工智能模型的普适性，并增加对新器件诊断的精准性。
本申请实施例提供的云服务器既可以向板端的测试机直接发送诊断结果，也可以向测试机发送训练好的全局人工智能模型由各个测试机来完成测试，本申请实施例在此不做限定。
参见图9，该图为本申请实施例提供的另一种包含多个测试机的功率半导体器件的测试系统示意图。
本申请实施例所提供的功率半导体器件的测试系统至少包括以下两台测试机：第一测试机201和第二测试机202。
其中，云服务器100，具体用于根据第一测试机201发送的第一历史数据和第二测试机202发送的第二历史数据进行训练获得全局人工智能模型，将全局人工智能模型发送给第一测试机201和第二测试机202。
第一测试机201，用于利用全局人工智能智能模型对对应的待测试功率半导体器件进行测试；第二测试机202，用于利用全局人工智能模型对对应的待测试功率半导体器件进行测试。
本申请实施例提供的测试系统，由于云服务器直接下发全局人工智能模型，测试机可以直接利用存储的全局人工智能模型，单机完成检测待测试功率半导体器件的任务，不需要与云端的云服务器进行通信。避免了板端的测试机与云端的云服务器之间网络中断时，本申请实施例中的功率半导体器件的测试系统无法得到待测试的功率半导体器件的诊断结果。如此，本申请实施例所提供的功率半导体器件的测试系统，可以减少云服务器和测试机之间对网络稳定性的依赖，提高了本申请实施例中功率半导体器件的测试系统获得诊断结果的实时性。
由上可知，本申请实施例所提供的功率半导体器件的测试系统可以通过所有的历史数据获得全局人工智能模型，从而利用该全局人工智能模型获得检测结果。但考虑到测试机可能存在一定的系统误差，不同的测试机之间测量得到的数据中的系统误差不同，且这种误差可能与测试机本身相关。因此，本申请实施例还提供了一种功率半导体器件的测试系统，该系统在获得了全局人工智能模型以后，根据不同的测试机获得的历史数据，对全局人工智能模型进行微调，得到不同的测试机对应的局部人工智能模型。
参见图10，该图为本申请实施例提供的再一种包含多个测试机的功率半导体器件的测试系统示意图。
如图10所示，本申请实施所提供的功率半导体器件的测试系统至少包括以下两台测试机：第一测试机201和第二测试机202。
其中，云服务器100，具体用于根据第一测试机201发送的第一历史数据和第二测试机202发送的第二历史数据进行训练获得全局人工智能模型，利用第一历史数据对全局人工智能模型进行调整获得第一人工智能模型，利用第二历史数据对全局人工智能模型进行调整获得第二人工智能模型；利用第一人工智能模型对第一测试机201对应的待测试功率半导体器件进行测试获得第一诊断结果，利用第二人工智能模型对第二测试机202对应的待测试功率半导体器件进行测试获得第二诊断结果，将第一诊断结果发送给第一测试机201，将第二诊断结果发送给第二测试机202。
需要说明的是，本申请实施例中第一测试机201发送的第一历史数据为第一测试机201对应的历史数据。即，第一历史数据包括第一测试机201直接或间接测量得到的数据。第一测试机202发送的第二历史数据为第二测试机202对应的历史数据。
本申请实施例利用第一历史数据和第二历史数据训练全局人工智能模型，可以充分的利用所有的历史数据，提高全局人工智能模型的测试精准性。在此基础上，考虑到不同的测试机之间的测试性能可能存在差异，本申请实施还利用第一历史数据对全局人工智能模型进行调整获得第一人工智能模型，利用第二历史数据对全局人工智能模型进行调整获得第二人工智能模型，从而可以得到针对第一测试机特性的局部人工智能模型(第一人工智能模型)和针对第二测试机特性的局部人工智能模型(第二人工智能模型)，进一步提高了实际检测时使用的人工智能模型的测试精准性。
例如，本申请实施例提供的功率半导体器件的测试系统中包含：测试机Q和测试机W，测试机Q发送测试机Q的历史数据至云服务器，测试机W发送测试机W的历史数据至云服务器。云服务器使用测试机Q的历史数据和测试机W的历史数据一起训练出全局人工智能模型。然后，云服务器使用测试机Q的历史数据对全局人工智能模型进行调整，获得测试机Q对应的局部人工智能模型，并利用该局部人工智能模型获得测试机Q对应的待测试的功率半导体器件进行测试。云服务器使用测试机W的历史数据对全局人工智能模型进行调整，获得测试机W对应的局部人工智能模型，并利用该局部人工智能模型获得测试机W对应的待测试的功率半导体器件进行测试。
本申请实施例提供的云服务器既可以向板端的测试机直接发送诊断结果，也可以向板端的测试机发送训练好的人工智能模型，由板端的测试机利用训练好的人工智能模型对功率半导体器件进行测试，获得诊断结果，下面结合附图进行详细介绍。
参见图11，该图为本申请实施例提供的又一种包含多个测试机的功率半导体器件的测试系统示意图。
本申请实施所提供的功率半导体器件的测试系统至少包括以下两台测试机：第一测试机201和第二测试机202；
云服务器100，具体用于根据第一测试机201发送的第一历史数据和第二测试机202 发送的第二历史数据进行训练获得全局人工智能模型，利用第一历史数据对全局人工智能模型进行调整获得第一人工智能模型发送给第一测试机201，利用第二历史数据对全局人工智能模型进行调整获得第二人工智能模型发送给第二测试机202。
第一测试机201，具体用于利用第一人工智能模型对对应的待测试功率半导体器件进行测试；第二测试机202，具体用于利用第二人工智能模型对对应的待测试功率半导体器件进行测试。
本申请实施例中，由于云服务器直接下发第一人工智能模型和第二人工智能模型分别给第一测试机和第二测试机，测试机可以直接利用对应的人工智能模型，单机完成检测待测试功率半导体器件的任务，不需要与云端的云服务器进行通信，避免了板端的测试机与云端的云服务器之间网络中断时，本申请实施例中的功率半导体器件的测试系统无法得到待测试的功率半导体器件的诊断结果。如此，本申请实施例所提供的功率半导体器件的测试系统，可以减少云服务器和测试机之间对网络稳定性的依赖，提高了本申请实施例中功率半导体器件的测试系统获得诊断结果的实时性。
由上可知，本申请实施例所提供的功率半导体器件的测试系统，既可以直接利用所有的历史数据获得全局人工智能模型并直接使用，也可以在获得全局人工智能模型以后利用不同的测试机对应的历史数据，对全局人工智能模型进行调整获得局部人工智能模型，并使用局部人工智能模型对其对应的测试机的功率半导体器件进行检测。
本申请实施例为了获得更高精准度的人工智能模型，作为一种可能的实施方式，还可以对诊断为失效的功率半导体器件进行后续测试，以验证该功率半导体器件是否失效，进而验证AI模型是否精准，如果验证功率半导体器件没有失效，则说明AI模型不准确，需要对AI模型进行调整。对判断为未失效的功率半导体器件，持续收集其在应用侧的运行数据，然后将以上两种功率半导体器件的数据上传至云服务器，云服务器使用这两种数据对AI模型进行调整。
为了使得本申请实施例的人工智能模型的精准度更高，作为一种可能的实施方式，本申请实施例中的测试机还用于结合测试机侧的数据对所述预先训练的模型进行微调，利用微调后的模型对所述待测试功率半导体器件进行诊断，输出诊断结果。需要说明的是，本申请实施例微调可以包括对预先训练的模型进行参数寻优。作为一个示例，可以利用迁移学习的方法对预先训练的AI模型进行微调。例如可以利用微调后的AI模型对晶圆是否失效进行测试。
考虑到本申请实施例的测试机中存储有其采集的少量历史数据，但这些数据并未被上传至云服务器，作为一种可能的实施方式，本申请实施例中的测试机侧的数据还可以包括未上传至云服务器中的数据。可以理解的是，本申请实施例中的测试机通常将其采集的数据定期上传至云服务器，本申请实施例中的测试机通常存储有一定量的历史数据未上传，因此，测试机可以利用这些未上传的数据对AI模型进行微调，然后基于微调后的AI模型对功率半导体器件进行检测。本申请实施例中的测试机对AI模型进行微调后，还可以将微调后的AI模型上传到云服务器，以便云服务器对AI模型进行统一管理。
综上所述，本申请实施例提供的功率半导体器件的测试系统，考虑到测试机的计算能 力和存储能力有限，利用具有较大计算能力和存储能力的云服务器训练人工智能模型，可以在云端的云服务器中利用人工智能模型直接获得诊断结果，也可以将人工智能模型发送至板端的测试机，由测试机利用人工智能模型获得诊断结果，还可以仅将未完全训练完成的人工智能模型发送至板端的测试机，板端的测试机根据其存储的数据完成人工智能模型的训练，并利用该训练完成的人工智能模型获得诊断结果。
服务器实施例
根据上述实施例提供的功率半导体器件的测试系统，本申请实施例还提供了一种云服务器。
参见图12，该图为本申请实施例提供的一种云服务器的结构示意图。
如图12所示，本申请实施例所提供的云服务器包括：第一收发设备1201和第一控制器1202。
其中，第一收发设备1201，用于接收测试机发送的待测试功率半导体器件的测试数据。
第一控制器1202，用于预先利用多个功率半导体器件的历史数据进行训练获得人工智能模型；还用于利用将待测试功率半导体器件的数据输入人工智能模型，预先训练的AI模型的输出为待测试功率半导体器件的诊断结果；第一收发设备，还用于将诊断结果发送给测试机；
或，第一控制器1202，用于将预先训练的人工智能模型发送给测试机，以使测试机利用预先训练的人工智能模型对待测试功率半导体器件进行诊断。
本申请实施例提供的云服务器的具体实现方式可以参见以上的测试系统实施例对云服务器的介绍，在此仅简要介绍。
本申请实施例的一种可能的实施方式，第一控制器具体用于利用无监督学习模型训练获得人工智能模型时，具体用于获得多个功率半导体器件的失效类型，利用专家知识在功率半导体器件的测试项中提取每个失效类型对应的所有测试项，每个失效类型对应的所有测试项形成测试项子集，对每个测试项子集利用无监督学习模型进行异常水平检测，得到异常器件和正常器件分别记第一分数和第二分数，获得每个功率半导体器件的所有测试项子集的第一分数和第二分数之和作为异常水平总分；异常水平总分大于预设分数阈值时，判断该功率半导体器件为异常器件。
本申请实施例的一种可能的实施方式，云服务器对应以下至少两台测试机：第一测试机和第二测试机。其中，第一控制器，具体用于根据第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，利用第一历史数据对全局人工智能模型进行调整获得第一人工智能模型，利用第二历史数据对全局人工智能模型进行调整获得第二人工智能模型；利用第一人工智能模型对第一测试机对应的待测试功率半导体器件进行测试获得第一诊断结果，利用第二人工智能模型对第二测试机对应的待测试功率半导体器件进行测试获得第二诊断结果；第一收发设备，具体用于将第一诊断结果发送给第一测试机，还用于将第二诊断结果发送给第二测试机。
本申请实施例的一种可能的实施方式，云服务器对应以下至少两台测试机：第一测试机和第二测试机。其中，第一控制器，具体用于根据第一测试机发送的第一历史数据和第 二测试机发送的第二历史数据进行训练获得全局人工智能模型，利用全局人工智能智能模型对第一测试机对应的待测试功率半导体器件进行测试获得第一诊断结果，利用全局人工智能模型对第二测试机对应的待测试功率半导体器件进行测试获得第二诊断结果；第一收发设备，具体用于将第一诊断结果发送给第一测试机，还用于将第二诊断结果发送给第二测试机。
本申请实施例的一种可能的实施方式，第一收发设备，还用于接收测试机发送的更新数据；第一控制器，还用于根据更新数据更新人工智能模型。
本申请实施例提供的云服务器，除了包括第一收发设备和第一控制器以外，还可以包括第一存储设备(图中未示出)，第一存储设备可以用于存储测试机上传的历史数据以及更新数据，便于第一控制器进行后续处理。第一控制器利用历史数据进行AI模型训练之前，为了训练的AI模型的参数更准确，可以剔除一些干扰数据，具体可以进行数据清洗。
综上所述，本申请实施例中的云服务器，可以利用其强大的计算能力和存储能力的云服务器对人工智能模型进行训练，从而获得诊断精准度较高的人工智能模型。在获得人工智能模型后，可以在云服务器中直接利用人工智能模型获得诊断结果，也可以仅为测试机提供人工智能模型，而不提供直接的诊断结果。
测试机实施例
基于上述实施例提供的功率半导体器件的测试系统和云服务器，本申请实施例还提供了一种测试机，下面结合附图进行详细介绍。
参见图13，该图为本申请实施例提供的一种测试机的结构示意图。
如图13所示，本申请实施例中的测试机，包括：第二收发设备1301和第二控制器1302；
第二收发设备1301，用于接收云服务器发送的人工智能模型，人工智能模型由云服务器预先利用多个功率半导体器件的历史数据进行训练获得；
第二控制器1302，用于获得待测试功率半导体器件的数据；将待测试功率半导体器件的数据输入预先训练的人工智能模型，预先训练的AI模型的输出为待测试功率半导体器件的诊断结果。
本申请实施例提供的测试机泛指任意一个测试机，一台云服务器可以对应多台本申请实施例提供的测试机，测试机的具体工作方式可以参见以上测试系统实施例针对测试机的介绍，在此仅简要说明。
本申请实施例的一种可能的实施方式，第二控制器还用于收集多个功率半导体器件的历史数据，并将收集的多个功率半导体器件的历史数据发送给云服务器；历史数据包括多个功率半导体器件的芯片测试数据或封装后的模组功能测试数据中的至少一项。
本申请实施例的一种可能的实施方式，第二控制器，还用于将待测试功率半导体器件的数据与预设的上下限数值进行比较，待测试功率半导体器件的数据超过上下限数值时，判断该待测试功率半导体器件为异常器件。且在本申请中，测试机可以直接获得来自云服务器发送的诊断结果，也可以获得云服务器发送的人工智能模型，并在测试机内通过人工智能模型获得诊断结果。
综上所述，本申请实施例提供的测试机的计算能力和存储能力有限，因此其通过将数 据发送至云服务器，利用云服务器强大计算能力和存储能力的云服务器对人工智能模型进行训练，从而获得诊断精准度较高的人工智能模型。
应当理解，在本申请中，“至少一个(项)”是指一个或者多个，“多个”是指两个或两个以上。“和/或”，用于描述关联对象的关联关系，表示可以存在三种关系，例如，“A和/或B”可以表示：只存在A，只存在B以及同时存在A和B三种情况，其中A，B可以是单数或者复数。字符“/”一般表示前后关联对象是一种“或”的关系。“以下至少一项(个)”或其类似表达，是指这些项中的任意组合，包括单项(个)或复数项(个)的任意组合。例如，a，b或c中的至少一项(个)，可以表示：a，b，c，“a和b”，“a和c”，“b和c”，或“a和b和c”，其中a，b，c可以是单个，也可以是多个。
以上所述，以上实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围。

Claims (22)

  1. 一种功率半导体器件的测试系统，其特征在于，包括：云服务器和至少一台测试机；
    所述云服务器，用于预先利用多个功率半导体器件的历史数据进行训练获得人工智能模型；
    所述测试机，用于获得待测试功率半导体器件的数据；
    所述云服务器或所述测试机，用于将所述待测试功率半导体器件的数据输入所述人工智能模型，所述人工智能模型的输出为所述待测试功率半导体器件的诊断结果。
  2. 根据权利要求1所述的系统，其特征在于，所述云服务器，用于将所述待测试功率半导体器件的数据输入所述人工智能模型，所述人工智能模型的输出为所述待测试功率半导体器件的诊断结果，将所述诊断结果发送给所述测试机。
  3. 根据权利要求1所述的系统，其特征在于，所述云服务器，用于将预先训练的人工智能模型发送给所述测试机；
    所述测试机，用于将所述待测试功率半导体器件的数据输入所述人工智能模型，所述人工智能模型的输出为所述待测试功率半导体器件的诊断结果。
  4. 根据权利要求1-3任一项所述的系统，其特征在于，所述测试机，还用于收集所述多个功率半导体器件的历史数据，并将收集的所述多个功率半导体器件的历史数据发送给所述云服务器；所述历史数据包括所述多个功率半导体器件的芯片测试数据、封装后的模组功能测试数据或实际应用异常时的相关数据中的至少一项；
    所述云服务器，还用于利用所述多个功率半导体器件的历史数据利用有监督学习模型或无监督学习模型中的至少一种进行训练获得所述人工智能模型。
  5. 根据权利要求1-4任一项所述的系统，其特征在于，所述云服务器利用所述无监督学习模型训练获得所述人工智能模型，具体用于获得所述多个功率半导体器件的失效类型，利用专家知识在功率半导体器件的测试项中提取每个失效类型对应的所有测试项，每个失效类型对应的所有测试项形成测试项子集，对每个测试项子集利用无监督学习模型进行异常子集和正常子集区分，得到所述异常子集和所述正常子集分别记第一分数和第二分数，获得每个功率半导体器件的所有测试项子集的所述第一分数和所述第二分数之和作为异常水平总分；所述异常水平总分大于或等于预设分数阈值时，判断该功率半导体器件为异常器件。
  6. 根据权利要求5所述的系统，其特征在于，所述云服务器，具体用于利用专家知识为不同的测试项子集中的异常器件记不同的第一分数。
  7. 根据权利要求1-4任一项所述的系统，其特征在于，所述云服务器利用所述有监督学习模型训练获得所述人工智能模型，具体用于获得所述功率半导体器件异常时的相关数据作为数据标签，利用有监督学习模型提取所述数据标签的数据特征，利用所述数据特征获得每个功率半导体器件的异常水平总分，所述异常水平总分大于预设分数阈值时，判断该功率半导体器件为异常器件。
  8. 根据权利要求1-7任一项所述的系统，其特征在于，所述测试机，还用于将所述待测试功率半导体器件的数据与预设的上下限数值进行比较，所述待测试功率半导体器件的数 据超过所述上下限数值时，判断该待测试功率半导体器件为异常器件。
  9. 根据权利要求2所述的系统，其特征在于，所述系统至少包括以下两台测试机：第一测试机和第二测试机；
    所述云服务器，具体用于根据所述第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，利用所述第一历史数据对所述全局人工智能模型进行调整获得第一人工智能模型，利用所述第二历史数据对所述全局人工智能模型进行调整获得第二人工智能模型；利用所述第一人工智能模型对所述第一测试机对应的待测试功率半导体器件进行测试获得第一诊断结果，利用所述第二人工智能模型对所述第二测试机对应的待测试功率半导体器件进行测试获得第二诊断结果，将所述第一诊断结果发送给所述第一测试机，将所述第二诊断结果发送给所述第二测试机。
  10. 根据权利要求2所述的系统，其特征在于，所述系统至少包括以下两台测试机：第一测试机和第二测试机；
    所述云服务器，具体用于根据所述第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，利用所述全局人工智能智能模型对所述第一测试机对应的待测试功率半导体器件进行测试获得第一诊断结果，利用所述全局人工智能模型对所述第二测试机对应的待测试功率半导体器件进行测试获得第二诊断结果，将所述第一诊断结果发送给所述第一测试机，将所述第二诊断结果发送给所述第二测试机。
  11. 根据权利要求3所述的系统，其特征在于，所述系统至少包括以下两台测试机：第一测试机和第二测试机；
    所述云服务器，具体用于根据所述第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，利用所述第一历史数据对所述全局人工智能模型进行调整获得第一人工智能模型发送给所述第一测试机，利用所述第二历史数据对所述全局人工智能模型进行调整获得第二人工智能模型发送给所述第二测试机；
    所述第一测试机，具体用于利用所述第一人工智能模型对对应的待测试功率半导体器件进行诊断；
    所述第二测试机，具体用于利用所述第二人工智能模型对对应的待测试功率半导体器件进行诊断。
  12. 根据权利要求3所述的系统，其特征在于，所述系统至少包括以下两台测试机：第一测试机和第二测试机；
    所述云服务器，具体用于根据所述第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，将所述全局人工智能模型发送给所述第一测试机和所述第二测试机；
    所述第一测试机，用于利用所述全局人工智能智能模型对对应的待测试功率半导体器件进行诊断；
    所述第二测试机，用于利用所述全局人工智能模型对对应的待测试功率半导体器件进行诊断。
  13. 根据权利要求1-12任一项所述的系统，其特征在于，所述测试机还用于向所述云服 务器发送更新数据；
    所述云服务器，还用于根据所述更新数据更新所述人工智能模型。
  14. 根据权利要求3-13任一项所述的系统，其特征在于，所述测试机还用于结合测试机侧的数据对所述预先训练的AI模型进行微调，利用微调后的模型对所述待测试功率半导体器件进行诊断，输出诊断结果。
  15. 一种云服务器，其特征在于，包括：第一收发设备和第一控制器；
    所述第一收发设备，用于接收测试机发送的待测试功率半导体器件的测试数据；
    所述第一控制器，用于预先利用多个功率半导体器件的历史数据进行训练获得人工智能模型；还用于利用将所述待测试功率半导体器件的数据输入所述人工智能模型，所述人工智能模型的输出为所述待测试功率半导体器件的诊断结果；所述第一收发设备，还用于将所述诊断结果发送给所述测试机；
    或，
    所述第一控制器，用于将所述人工智能模型发送给所述测试机，以使所述测试机利用所述人工智能模型对待测试功率半导体器件进行诊断。
  16. 根据权利要求15所述的云服务器，其特征在于，所述第一控制器，具体用于利用无监督学习模型训练获得所述人工智能模型时，具体用于获得所述多个功率半导体器件的失效类型，利用专家知识在功率半导体器件的测试项中提取每个失效类型对应的所有测试项，每个失效类型对应的所有测试项形成测试项子集，对每个测试项子集利用无监督学习模型进行异常水平检测，得到异常器件和正常器件分别记第一分数和第二分数，获得每个功率半导体器件的所有测试项子集的第一分数和第二分数之和作为异常水平总分；所述异常水平总分大于预设分数阈值时，判断该功率半导体器件为异常器件。
  17. 根据权利要求15或16所述的云服务器，其特征在于，所述云服务器对应以下至少两台测试机：第一测试机和第二测试机；
    所述第一控制器，具体用于根据所述第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，利用所述第一历史数据对所述全局人工智能模型进行调整获得第一人工智能模型，利用所述第二历史数据对所述全局人工智能模型进行调整获得第二人工智能模型；利用所述第一人工智能模型对所述第一测试机对应的待测试功率半导体器件进行测试获得第一诊断结果，利用所述第二人工智能模型对所述第二测试机对应的待测试功率半导体器件进行测试获得第二诊断结果；
    所述第一收发设备，具体用于将所述第一诊断结果发送给所述第一测试机，还用于将所述第二诊断结果发送给所述第二测试机。
  18. 根据权利要求15或16所述的云服务器，其特征在于，所述云服务器对应以下至少两台测试机：第一测试机和第二测试机；
    所述第一控制器，具体用于根据所述第一测试机发送的第一历史数据和第二测试机发送的第二历史数据进行训练获得全局人工智能模型，利用所述全局人工智能智能模型对所述第一测试机对应的待测试功率半导体器件进行测试获得第一诊断结果，利用所述全局人工智能模型对所述第二测试机对应的待测试功率半导体器件进行测试获得第二诊断结果；
    所述第一收发设备，具体用于将所述第一诊断结果发送给所述第一测试机，还用于将所述第二诊断结果发送给所述第二测试机。
  19. 根据权利要求14-17任一项所述的云服务器，其特征在于，所述第一收发设备，还用于接收所述测试机发送的更新数据；
    所述第一控制器，还用于根据所述更新数据更新所述人工智能模型。
  20. 一种测试机，其特征在于，包括：第二收发设备和第二控制器；
    所述第二收发设备，用于接收云服务器发送的人工智能模型，所述人工智能模型由所述云服务器预先利用多个功率半导体器件的历史数据进行训练获得；
    所述第二控制器，用于获得待测试功率半导体器件的数据；将所述待测试功率半导体器件的数据输入所述人工智能模型，所述人工智能模型的输出为所述待测试功率半导体器件的诊断结果。
  21. 根据权利要求20所述的测试机，其特征在于，所述第二控制器，还用于收集所述多个功率半导体器件的历史数据，并将收集的所述多个功率半导体器件的历史数据发送给所述云服务器；所述历史数据包括所述多个功率半导体器件的芯片测试数据或封装后的模组功能测试数据中的至少一项。
  22. 根据权利要求20或21所述的测试机，其特征在于，所述第二控制器，还用于将所述待测试功率半导体器件的数据与预设的上下限数值进行比较，所述待测试功率半导体器件的数据超过所述上下限数值时，判断该待测试功率半导体器件为异常器件。
PCT/CN2021/091550 2021-04-30 2021-04-30 一种功率半导体器件的测试系统、云服务器及测试机 WO2022227018A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202180007976.XA CN115398251A (zh) 2021-04-30 2021-04-30 一种功率半导体器件的测试系统、云服务器及测试机
PCT/CN2021/091550 WO2022227018A1 (zh) 2021-04-30 2021-04-30 一种功率半导体器件的测试系统、云服务器及测试机

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/091550 WO2022227018A1 (zh) 2021-04-30 2021-04-30 一种功率半导体器件的测试系统、云服务器及测试机

Publications (1)

Publication Number Publication Date
WO2022227018A1 true WO2022227018A1 (zh) 2022-11-03

Family

ID=83847720

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/091550 WO2022227018A1 (zh) 2021-04-30 2021-04-30 一种功率半导体器件的测试系统、云服务器及测试机

Country Status (2)

Country Link
CN (1) CN115398251A (zh)
WO (1) WO2022227018A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2024080148A1 (ja) * 2022-10-10 2024-04-18 株式会社デンソー 半導体装置の製造方法および製造装置

Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2006146459A (ja) * 2004-11-18 2006-06-08 Renesas Technology Corp 半導体デバイスの製造方法および製造システム
CN103310285A (zh) * 2013-06-17 2013-09-18 同济大学 可用于半导体生产线动态调度的性能预测方法
US20140197861A1 (en) * 2013-01-17 2014-07-17 Samsung Electronics Co., Ltd. Test Equipment for Testing Semiconductor Device and Methods of Testing Semiconductor Device Using the Same
CN104898036A (zh) * 2014-03-05 2015-09-09 优力勤股份有限公司 集成电路装置的云端测试及远程监视的系统及方法
CN108268899A (zh) * 2018-01-22 2018-07-10 清华大学 一种电子元件的检测方法、装置和设备
US20180356807A1 (en) * 2017-06-13 2018-12-13 Streammosaic, Inc. Generating robust machine learning predictions for semiconductor manufacturing processes
CN109801479A (zh) * 2018-12-12 2019-05-24 西安电子科技大学 一种半导体工艺过程监测与预警系统及其控制方法
CN110024097A (zh) * 2016-11-30 2019-07-16 Sk 株式会社 基于机器学习的半导体制造良率预测系统和方法
CN111539552A (zh) * 2020-03-19 2020-08-14 卡斯柯信号有限公司 基于半导体市场数据的电子元器件生命周期预测方法
WO2021014282A1 (en) * 2019-07-24 2021-01-28 Abb Schweiz Ag Overheating detection in an electric drive based on monitoring of thermal behaviour of semiconductor modules of the electric drive
CN112308298A (zh) * 2020-10-16 2021-02-02 同济大学 一种面向半导体生产线的多场景性能指标预测方法及系统
CN112380768A (zh) * 2020-11-11 2021-02-19 长沙理工大学 一种基于bp神经网络的led芯片寿命预测方法
CN112433472A (zh) * 2019-08-26 2021-03-02 长鑫存储技术有限公司 半导体生产控制方法及控制系统
CN112488182A (zh) * 2020-11-27 2021-03-12 上海华力集成电路制造有限公司 一种半导体器件的良率预测方法及装置

Patent Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2006146459A (ja) * 2004-11-18 2006-06-08 Renesas Technology Corp 半導体デバイスの製造方法および製造システム
US20140197861A1 (en) * 2013-01-17 2014-07-17 Samsung Electronics Co., Ltd. Test Equipment for Testing Semiconductor Device and Methods of Testing Semiconductor Device Using the Same
CN103310285A (zh) * 2013-06-17 2013-09-18 同济大学 可用于半导体生产线动态调度的性能预测方法
CN104898036A (zh) * 2014-03-05 2015-09-09 优力勤股份有限公司 集成电路装置的云端测试及远程监视的系统及方法
CN110024097A (zh) * 2016-11-30 2019-07-16 Sk 株式会社 基于机器学习的半导体制造良率预测系统和方法
US20180356807A1 (en) * 2017-06-13 2018-12-13 Streammosaic, Inc. Generating robust machine learning predictions for semiconductor manufacturing processes
CN108268899A (zh) * 2018-01-22 2018-07-10 清华大学 一种电子元件的检测方法、装置和设备
CN109801479A (zh) * 2018-12-12 2019-05-24 西安电子科技大学 一种半导体工艺过程监测与预警系统及其控制方法
WO2021014282A1 (en) * 2019-07-24 2021-01-28 Abb Schweiz Ag Overheating detection in an electric drive based on monitoring of thermal behaviour of semiconductor modules of the electric drive
CN112433472A (zh) * 2019-08-26 2021-03-02 长鑫存储技术有限公司 半导体生产控制方法及控制系统
CN111539552A (zh) * 2020-03-19 2020-08-14 卡斯柯信号有限公司 基于半导体市场数据的电子元器件生命周期预测方法
CN112308298A (zh) * 2020-10-16 2021-02-02 同济大学 一种面向半导体生产线的多场景性能指标预测方法及系统
CN112380768A (zh) * 2020-11-11 2021-02-19 长沙理工大学 一种基于bp神经网络的led芯片寿命预测方法
CN112488182A (zh) * 2020-11-27 2021-03-12 上海华力集成电路制造有限公司 一种半导体器件的良率预测方法及装置

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2024080148A1 (ja) * 2022-10-10 2024-04-18 株式会社デンソー 半導体装置の製造方法および製造装置

Also Published As

Publication number Publication date
CN115398251A (zh) 2022-11-25

Similar Documents

Publication Publication Date Title
US6792373B2 (en) Methods and apparatus for semiconductor testing
US11204824B1 (en) Intelligent network operation platform for network fault mitigation
US11714700B2 (en) Intelligent network operation platform for network fault mitigation
CN103199041B (zh) 晶圆允收测试程式的管理系统及其应用方法
TWI498573B (zh) 半導體晶片重測系統及其重測方法
WO2022227018A1 (zh) 一种功率半导体器件的测试系统、云服务器及测试机
US10782348B2 (en) Automatic device detection and connection verification
CN115576738B (zh) 一种基于芯片分析实现设备故障确定的方法及系统
CN103164309A (zh) Sol功能测试方法及系统
CN115840915A (zh) 机电设备故障的自动识别方法、系统、终端及存储介质
CN115965230A (zh) 一种晶圆制程良率分析方法、设备及系统
US10656204B2 (en) Failure detection for wire bonding in semiconductors
TWI698644B (zh) 自動測試設備的異常測試信號通道檢測方法
US20080004829A1 (en) Method and apparatus for automatic test equipment
US6792386B2 (en) Method and system for statistical comparison of a plurality of testers
CN107562561A (zh) 计算机硬件快速诊断测试系统
CN111190092B (zh) 一种fpga测试质量控制优化系统
CN109597392A (zh) 有助于故障诊断的方法、装置和设备以及机器可读介质
US20080206903A1 (en) Adaptive threshold wafer testing device and method thereof
US20120089879A1 (en) Method and system for identifying power defects using test pattern switching activity
WO2022257058A1 (zh) 基于晶圆测试的相关性分析方法及装置
KR20170127876A (ko) 로그 결함 분석 기반 장애 대응 시스템 및 방법
CN110991673A (zh) 用于复杂系统的故障隔离和定位方法
CN110532512A (zh) 基于大数据分析的设备失效模式诊断特征参量分析方法
Kang et al. A method to group reliability data by hierarchical clustering

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21938492

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 27/02/2024)