Method for bipolar transistor manufacturing process identification based on machine learning physical feature enhancement

Abstract

The application relates to a bipolar transistor manufacturing process identification method based on machine learning physical feature enhancement. The device used in the method comprises a BC3193 semiconductor discrete device test system, a transistor test seat, a computer display and a bipolar transistor sample. By testing the electrical parameters of the bipolar transistor under four electrical stresses, physical feature enhancement is carried out, difference characteristics between the electrical stress and the electrical parameters are constructed, a multi-stress amplification coefficient feature is constructed, a manufacturer is identified by using feature matching, a batch classifier is constructed by using a random forest algorithm, and the identification of the bipolar transistor manufacturing process is realized. The application uses the electrical parameter test data under four electrical stresses to complete the identification of the bipolar transistor manufacturer and the batch under the non-destructive condition, and effectively improves the identification accuracy and reliability.

Description

Technical Field

[0001] This invention relates to fields such as semiconductor device reliability, and specifically to a method for identifying bipolar transistor manufacturing processes based on machine learning-enhanced physical features. Background Technology

[0002] Bipolar transistors (BJTs) are a crucial component of discrete semiconductor devices, widely used in aerospace, industrial control, and other fields with extremely high reliability requirements. Their performance is closely related to the manufacturing process, and accurately identifying the process origin is essential for component selection, supply chain management, and system reliability assessment. Existing BJT process identification methods have significant drawbacks: First, identification based on appearance markings and packaging is susceptible to refurbishment and tampering, resulting in poor reliability. Second, methods based on statistical comparison of electrical parameters under conventional electrical stress only reflect the basic performance of the device and cannot fully capture subtle differences caused by process variations. Furthermore, as processes mature, the overlap of parameters between different manufacturers and batches is high, leading to low identification accuracy. Third, while destructive physical analysis is accurate, it damages devices, is costly, and time-consuming, making it unsuitable for scenarios requiring the preservation of device integrity. In summary, conventional electrical stress parameters are insufficient to reflect process variations, and existing methods cannot meet the needs for accurate identification. There is an urgent need in this field for a non-destructive method that uses physical stress to enhance features and capture process differences, enabling accurate identification of manufacturers and batches. Summary of the Invention

[0003] The purpose of this invention is to provide a method for identifying bipolar transistor (BPT) manufacturing processes based on machine learning-enhanced physical features. The method involves a device comprising a BC3193 semiconductor discrete device testing system, a transistor test socket, a computer monitor, and BPT samples. By testing the electrical parameters of BPTs under normal conditions, and after data cleaning, the differences between electrical stress and electrical parameters, as well as the multi-stress amplification factor features, are extracted. All acquired data constitutes a dataset. Specifically, the method utilizes parameter matching to identify manufacturers and employs a random forest algorithm to train a batch classifier, outputting the identification results and confidence levels.

[0004] The present invention discloses a method for identifying bipolar transistor manufacturing processes based on machine learning-enhanced physical features. The apparatus involved in this method comprises a BC3193 semiconductor discrete device testing system (1), a transistor test socket (2), a computer monitor (3), and a bipolar transistor sample (4). The transistor test socket (2) is installed onto the test socket interface of the BC3193 semiconductor discrete device testing system (1), the computer monitor (3) is connected to the BC3193 semiconductor discrete device testing system (1), and the bipolar transistor sample (4) is inserted into the transistor test socket (2). The specific operation is performed according to the following steps: a. Based on the device datasheet of the manufacturer of the bipolar transistor sample (4), design a test procedure for the bipolar transistor sample (4) using the BC3193 semiconductor discrete device test system (1); b. Insert the bipolar transistor sample (4) into the transistor test socket (2) and install it into the test socket interface of the BC3193 semiconductor discrete device test system (1). At the same time, connect the computer monitor (3) to the BC3193 semiconductor discrete device test system (1). c. Based on the test device setup in step b, conduct tests on the current amplification factor, collector-emitter saturation voltage, base-emitter saturation voltage, collector-emitter reverse breakdown voltage, collector-base reverse cutoff current, and emitter-base reverse cutoff current of the bipolar transistor sample (4) under conventional electrical stress. After the test is completed, collect the test data. d. After completing the test in step c, specify one type of unconventional electrical stress condition in the base region broadening effect operating region and two types in the linear amplification region, and complete the tests of current amplification factor, collector-emitter saturation voltage, and base-emitter saturation voltage under the three electrical stress conditions. Collect the test data after the test is completed. e. Perform comprehensive numerical cleaning and format unification processing on the test data collected in steps c and d, automatically identify and convert non-standard data formats such as scientific notation and string format values, and clean up interfering characters such as spaces and newlines in the values; f. Based on the test data processed in step e, calculate the difference characteristics between electrical stress and electrical parameters. That is, use the difference between the current amplification factor value under conventional electrical stress and the current amplification factor value under the base region broadening effect, and then divide by the current amplification factor value under conventional electrical stress. Similarly, calculate the relative difference characteristics between collector-emitter saturation voltage and the difference characteristics between electrical stress and electrical parameters between base-emitter saturation voltage, and use the current amplification factor under two unconventional electrical stress conditions in the linear amplification region as the multi-stress amplification factor characteristics. g. Construct a dataset from the data processed in steps e and f, and divide it into a training set, a validation set, and a test set with a bipolar transistor sample ratio of 8:1:1. h. Establish a manufacturer parameter feature library based on the training set, group all electrical parameters by manufacturer and statistically analyze the quartile range, calculate the normal value range of each parameter as the typical parameter distribution of that manufacturer, and record the mean, median and standard deviation of each manufacturer in terms of relative difference characteristics. i. Traverse the parameter range databases of all manufacturers and calculate the matching degree between each parameter of the new bipolar transistor sample and the corresponding parameter range of each manufacturer; j. Convert the matching scores of all manufacturers into a probability distribution by probability normalization, select the manufacturer with the highest probability as the prediction result, use this probability value as the confidence level of manufacturer identification, and verify the model performance on the test set. k. Random forest is used as the core algorithm for batch identification, and the final batch identification result is determined by the voting results of multiple decision trees; 1. Select the corresponding batch classifier based on the manufacturer identification result, input the feature data of the new bipolar transistor sample into the classifier, obtain the probability distribution of each batch, use the label encoder to convert the predicted batch index into the actual batch number, take the batch with the highest probability as the batch identification result, take the highest probability as the confidence level of batch identification, and verify the model performance on the test set. m. By setting a Boolean switch in the model to control whether or not the characteristics of the difference between electrical stress and electrical parameters and the characteristics of multiple stress amplification factors are included. Under the premise of keeping the model structure unchanged, for the two states of the Boolean switch, steps h, i, j, k, and l are executed sequentially to quantitatively compare the differences in manufacturing process identification accuracy with and without the characteristics of the difference between electrical stress and electrical parameters and the characteristics of multiple stress amplification factors.

[0005] The bipolar transistor manufacturing process identification method based on machine learning-enhanced physical features described in this invention does not rely on physical device marking or destructive analysis; it can achieve process identification using only electrical parameters obtained from conventional semiconductor testing systems. It is applicable not only to distinguishing different batches from the same manufacturer but also to cross-identifying alternative models from different manufacturers. Furthermore, this method maintains stable identification accuracy even under complex operating conditions such as normal fluctuations in device parameters and insignificant changes in high-load characteristics, providing an effective technical means for electronic component supply chain traceability, quality reliability assessment, and failure analysis. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of the bipolar transistor sample test of the present invention; Figure 2 This is a graph evaluating the manufacturer identification accuracy of the physical feature enhancement machine learning method of the present invention. Figure 3 This is a batch recognition accuracy evaluation diagram for the physical feature enhancement machine learning method of the present invention. Figure 4 The diagram is a composite image. The left side shows a comparison and evaluation chart of the manufacturer identification accuracy using the Boolean switch to control the presence or absence of features, while the right side shows a comparison and evaluation chart of the batch identification accuracy using the Boolean switch to control the presence or absence of features. Detailed Implementation

[0007] The present invention will now be described in further detail with reference to the accompanying drawings. Example 1

[0008] This invention discloses a bipolar transistor (BPT) manufacturing process identification method based on machine learning-enhanced physical features. The method involves an apparatus comprising a BC3193 semiconductor discrete device testing system 1, a transistor test socket 2, a computer monitor 3, and a BPT sample 4. The transistor test socket 2 is installed onto the test socket interface of the BC3193 semiconductor discrete device testing system 1. The computer monitor 3 is then connected to the BC3193 semiconductor discrete device testing system 1. Finally, the BPT sample 4 is inserted into the transistor test socket 2. (Refer to...) Figure 1 The specific operation is carried out according to the following steps: a. Select bipolar transistor samples 4 from three manufacturers: Multicomp, Jinan Hengjing, and Shijiazhuang Radio Factory No. 2. For each manufacturer, select three independent production batches of the 2N2222A model and its alternative models 3DK2222 and F2N2222A. Select 50 to 70 bipolar transistor samples from each batch. Based on the device datasheet of the manufacturer to which the bipolar transistor sample 4 belongs, design a test program for the bipolar transistor sample 4 using the BC3193 semiconductor discrete device test system 1. b. Insert the bipolar transistor sample 4 into the transistor test socket 2 and install it onto the test socket interface of the BC3193 semiconductor discrete device test system 1. Simultaneously, connect the computer monitor 3 to the BC3193 semiconductor discrete device test system 1. Figure 1 ; c. Based on the test setup in step b, conduct tests on the current amplification factor, collector-emitter saturation voltage, base-emitter saturation voltage, collector-emitter reverse breakdown voltage, collector-base reverse cutoff current, and emitter-base reverse cutoff current of bipolar transistor sample 4 under normal electrical stress. After the tests are completed, collect the test data. d. After completing the tests in step b, specify one operating condition within the base-widening effect region (90% of the maximum collector current the bipolar transistor sample can withstand) and two unconventional electrical stress conditions within the linear amplification region. Refer to the device datasheet for the maximum rated values. Under these three electrical stress conditions, test the current amplification factor, collector-emitter saturation voltage, and base-emitter saturation voltage. After testing, collect the data and compare the current amplification factor distribution at the conventional and base-widening effect operating points to verify the effectiveness of the electrical stress change. Figure 2 , Figure 3 ; e. Perform comprehensive numerical cleaning and format unification processing on the test data collected in steps c and d, automatically identify and convert non-standard data formats such as scientific notation and string format values, and clean up interfering characters such as spaces and newlines in the values; f. Based on the test data processed in step e, calculate the difference characteristics between electrical stress and electrical parameters. This is done by subtracting the current amplification factor under base broadening effect from the current amplification factor under conventional electrical stress, and then dividing by the current amplification factor under conventional electrical stress. Similarly, calculate the relative difference characteristics between collector-emitter saturation voltage and the difference characteristics between electrical stress and electrical parameters for base-emitter saturation voltage. The current amplification factors under two unconventional electrical stress conditions in the linear amplification region are then used as the multi-stress amplification factor characteristics. g. Construct a dataset from the data processed in steps e and f, and divide it into a training set, a validation set, and a test set with a bipolar transistor sample ratio of 8:1:1. h. Build a manufacturer parameter feature library from the training set in step g, group all electrical parameters by manufacturer and calculate the quartile range of all electrical parameters, calculate the normal range of each parameter as the typical parameter distribution of the manufacturer, and record the mean, median and standard deviation of each manufacturer in terms of relative difference characteristics. i. Traverse the parameter range databases of all manufacturers and calculate the matching degree between each parameter of the new bipolar transistor sample and the corresponding parameter range of each manufacturer; j. Convert the matching scores of all manufacturers into a probability distribution by probability normalization, select the manufacturer with the highest probability as the prediction result, use this probability value as the confidence level of manufacturer identification, and verify the model performance on the test set. k. Random forest is used as the core algorithm for batch recognition. 100 decision trees are used, with a maximum depth of 10 and a minimum number of split samples per node of 2. The final batch recognition result is determined by the voting results of multiple decision trees. 1. Select the corresponding batch classifier based on the manufacturer identification result, input the feature data of the new bipolar transistor sample into the classifier, obtain the probability distribution of each batch, use the label encoder to convert the predicted batch index into the actual batch number, take the batch with the highest probability as the batch identification result, take the highest probability as the confidence level of batch identification, and verify the model performance on the test set. m. By setting a Boolean switch in the model to control whether or not the characteristics of the difference between electrical stress and electrical parameters and the characteristics of multiple stress amplification factors are included, while keeping the model structure unchanged, steps h, i, j, k, and l are executed sequentially for the two states of the Boolean switch, respectively. The differences in manufacturing process identification accuracy with and without the characteristics of the difference between electrical stress and electrical parameters and the characteristics of multiple stress amplification factors are quantitatively compared. Figure 3 .

[0009] The above description is merely a specific implementation of a bipolar transistor manufacturing process identification method based on machine learning physical feature enhancement according to the present invention. However, the scope of protection of the present invention is not limited thereto. Any substitutions or additions that can be understood by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of the present invention.

Claims

1. A method for identifying bipolar transistor manufacturing processes based on machine learning-enhanced physical features. The device involved in this method consists of a BC3193 semiconductor discrete device testing system (1), a transistor test socket (2), a computer monitor (3), and a bipolar transistor sample (4). The transistor test socket (2) is installed on the test socket interface of the BC3193 semiconductor discrete device testing system (1), the computer monitor (3) is connected to the BC3193 semiconductor discrete device testing system (1), and the bipolar transistor sample (4) is inserted into the transistor test socket (2). The specific operation is carried out according to the following steps: a. Based on the device datasheet of the manufacturer of the bipolar transistor sample (4), design a test procedure for the bipolar transistor sample (4) using the BC3193 semiconductor discrete device test system (1); b. Insert the bipolar transistor sample (4) into the transistor test socket (2) and install it into the test socket interface of the BC3193 semiconductor discrete device test system (1). At the same time, connect the computer monitor (3) to the BC3193 semiconductor discrete device test system (1). c. Based on the test device setup in step b, conduct tests on the current amplification factor, collector-emitter saturation voltage, base-emitter saturation voltage, collector-emitter reverse breakdown voltage, collector-base reverse cutoff current, and emitter-base reverse cutoff current of the bipolar transistor sample (4) under conventional electrical stress. After the test is completed, collect the test data. d. After completing the test in step c, specify one type of unconventional electrical stress condition in the base region broadening effect operating region and two types in the linear amplification region, and complete the tests of current amplification factor, collector-emitter saturation voltage, and base-emitter saturation voltage under the three electrical stress conditions. Collect the test data after the test is completed. e. Perform comprehensive numerical cleaning and format unification processing on the test data collected in steps c and d, automatically identify and convert non-standard data formats such as scientific notation and string format values, and clean up interfering characters such as spaces and newlines in the values; f. Based on the test data processed in step e, calculate the difference characteristics between electrical stress and electrical parameters. That is, use the difference between the current amplification factor value under conventional electrical stress and the current amplification factor value under the base region broadening effect, and then divide by the current amplification factor value under conventional electrical stress. Similarly, calculate the relative difference characteristics between collector-emitter saturation voltage and the difference characteristics between electrical stress and electrical parameters between base-emitter saturation voltage, and use the current amplification factor under two unconventional electrical stress conditions in the linear amplification region as the multi-stress amplification factor characteristics. g. Construct a dataset from the data processed in steps e and f, and divide it into a training set, a validation set, and a test set with a bipolar transistor sample ratio of 8:1:

1. h. Establish a manufacturer parameter feature library based on the training set, group all electrical parameters by manufacturer and statistically analyze the quartile range, calculate the normal value range of each parameter as the typical parameter distribution of that manufacturer, and record the mean, median and standard deviation of each manufacturer in terms of relative difference characteristics. i. Traverse the parameter range databases of all manufacturers and calculate the matching degree between each parameter of the new bipolar transistor sample and the corresponding parameter range of each manufacturer; j. Convert the matching scores of all manufacturers into a probability distribution by probability normalization, select the manufacturer with the highest probability as the prediction result, use this probability value as the confidence level of manufacturer identification, and verify the model performance on the test set. k. Random forest is used as the core algorithm for batch identification, and the final batch identification result is determined by the voting results of multiple decision trees; 1. Select the corresponding batch classifier based on the manufacturer identification result, input the feature data of the new bipolar transistor sample into the classifier, obtain the probability distribution of each batch, use the label encoder to convert the predicted batch index into the actual batch number, take the batch with the highest probability as the batch identification result, take the highest probability as the confidence level of batch identification, and verify the model performance on the test set. m. By setting a Boolean switch in the model to control whether or not the characteristics of the difference between electrical stress and electrical parameters and the characteristics of multiple stress amplification factors are included. While keeping the model structure unchanged, steps h, i, j, k, and l are executed sequentially for the two states of the Boolean switch, respectively, to quantitatively compare the differences in manufacturing process identification accuracy with and without the characteristics of the difference between electrical stress and electrical parameters and the characteristics of multiple stress amplification factors.

Images