Method for rapidly measuring INL and DNL parameters of ADC product by ATE

Abstract

The invention discloses a method for rapidly measuring INL and DNL parameters of an ADC product applied to ATE, and relates to the field of parameter measurement, ADC product data and environmental parameters are obtained and preprocessed, a three-layer neural network model is configured based on data obtained in S100, error prediction is performed based on an operation result of the three-layer neural network model in S200, and the three-layer neural network model is optimized. S300, complementing ADC product data and adjustment based on the error prediction value obtained in the S300, obtaining complemented data parameters based on the complemented ADC product data, and performing feedback processing based on the complemented data parameters obtained in the S500, performing error prediction and complementation through a neural network model by using a small amount of collected data, significantly reducing the test data collection amount, greatly reducing the test time and cost, and improving the test efficiency. Meanwhile, by complementing and correcting the data of the uncollected part, the precision of the test result is ensured, the method is suitable for ADC products with high integration level and high frequency, and the consistency of the test efficiency and the product performance is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of parameter measurement, and in particular to a method for quickly measuring INL and DNL parameters of ADC products by ATE. Background Art

[0002] ATE is a widely used equipment for testing integrated circuits to evaluate the electrical performance of chips during production. ADC is an interface circuit that converts analog signals into digital signals. INL and DNL are two important static parameters used to measure the conversion accuracy and performance of ADC. Traditionally, before the chip leaves the factory, the ADC needs to be fully tested by ATE to ensure that it meets the design requirements and application specifications. The test usually involves applying a specific input signal to the chip and using instruments such as an oscilloscope to observe the accuracy of the output signal. However, traditional ADC testing relies on large amounts of data collection and complex testing processes, resulting in high testing costs and long testing times. Especially in high-frequency ADC products, the hardware resources required for testing are extremely large, and during the data collection process, the sampling accuracy and testing efficiency are insufficient. In addition, traditional testing methods cannot effectively supplement the data that has not been collected, resulting in limited accuracy of the test results, making it difficult to meet the requirements of modern high-integration, high-frequency products. Summary of the invention

[0003] The purpose of the present invention is to provide a method for fast measurement of INL and DNL parameters of ADC products by ATE in view of the shortcomings of the prior art, aiming to use a small amount of collected data to predict and complete errors through a neural network model, significantly reducing the amount of test data collected, and greatly reducing test time and cost. At the same time, the method ensures the accuracy of the test results by completing and correcting the data of the uncollected part, is suitable for ADC products with high integration and high frequency, and effectively improves the test efficiency and consistency of product performance.

[0004] To this end, the present application provides a method for quickly measuring INL and DNL parameters of ADC products by ATE, comprising the following steps: Step S100: Acquire ADC product data and environmental parameters and perform preprocessing.

[0005] Step S200: Based on the data obtained in S100, a three-layer neural network model is configured.

[0006] Step S300: Based on the calculation result of the three-layer neural network model in S200, error prediction and optimization of the three-layer neural network model are performed.

[0007] Step S400: based on the error prediction value obtained in S300, complete the ADC product data and make adjustments.

[0008] Step S500: Obtaining supplemented data parameters based on the supplemented ADC product data.

[0009] Step S600: performing feedback processing based on the completed data parameters obtained in S500.

[0010] In some specific implementations, obtaining ADC product data and environmental parameters and preprocessing specifically include: Step S100.1: Acquire historical data and current partial data of ADC products based on ATE equipment.

[0011] Among them, the historical data include: INL and DNL data.

[0012] Among them, the current partial data are: partial INL and DNL data.

[0013] Step S100.2: Obtain test environment parameters.

[0014] The environmental parameters include: temperature and voltage fluctuations.

[0015] The temperature is obtained based on a temperature sensor.

[0016] The voltage fluctuation is obtained based on a voltage monitoring device.

[0017] Step S100.3: pre-processing is performed based on the obtained historical data, current partial data and environmental parameters.

[0018] It should be noted that the preprocessing includes: data clarification and standardization processing, wherein the preprocessing is an existing technology well known to those skilled in the art and will not be elaborated here.

[0019] In some specific implementations, a three-layer neural network model is used to perform preliminary operations on the data obtained in S100, specifically including: Steps Step S200.1. Based on the acquired historical data, divide the historical data into: a training set, a validation set and a test set.

[0020] Step S200.1.1, construct a neural network model.

[0021] The three-layer neural network comprises an input layer, a hidden layer and an output layer.

[0022] The input layer is: the collected ADC product input signal characteristics.

[0023] Among them, the signal characteristics include: input voltage, output value, temperature, and frequency.

[0024] The hidden layer is used to capture complex nonlinear relationships.

[0025] Among them, the hidden layer adopts ReLU activation function and contains 20 neurons.

[0026] The output layer is used for linear activation function.

[0027] Among them, the output layer contains 2 neurons, corresponding to the predicted values ​​of INL and DNL respectively.

[0028] Based on the configured three-layer neural network model, historical data is input, and the characteristics are , the output of each layer is:

[0029] In the formula, is the weight matrix, is the bias term, X is the input feature vector, X includes voltage, output, and temperature, and the ReLU activation function processes nonlinear features.

[0030] Step S200.2: Train the neural network model based on the training set and obtain the predicted values ​​of INL and DNL.

[0031] Step S200.2.1, using a loss function, comparing the predicted value obtained by the neural network model with the current partial data to minimize the difference between the two, the loss function is specifically:

[0032] In the formula, is the loss function, n is the number of samples in the dataset, is the predicted INL value of the ith data point, is the true INL value of the ith data point, is the predicted DNL value of the ith data point, is the true DNL value of the ith data point, is the square of the INL error of the ith data point, is the square of the DNL error of the ith data point.

[0033] Step S200.2.2, use the stochastic gradient descent algorithm to update the weights in the neural network model, specifically:

[0034] In the formula, is the weight matrix during the tth round of training, is the updated weight after the t+1th round of training, is the learning rate, which controls the step size of each weight update. is the partial derivative of the loss function with respect to the current weight and measures the impact of the current weight on the loss.

[0035] Step S200.2.3, use the validation set data to evaluate the trained neural network model.

[0036] In the evaluation, if the loss value of the neural network model on the validation set does not decrease by more than 5%, the number of training rounds is increased, and the number of increased training rounds is: 50 to 100 rounds.

[0037] Step S200.3, based on the test set and the trained neural network model, perform an error prediction test on the current part of the data described in step S100.2, and test the INL and DNL data of the unsampled part completed by the neural network model.

[0038] The error prediction is specifically:

[0039] In the formula, is the predicted INL and DNL value corresponding to the i-th input data point, is the output function processed by the neural network model, and the input is the data point , neural network model weights W and bias b, is the input feature of the i-th sampling point, including input voltage, output signal, and temperature.

[0040] The unsampled part is completed, and the result is predicted by using a neural network model to infer the error of the unsampled data point.

[0041] Step S200.4: Based on the prediction results of the validation set and the neural network model, further correct the INL and DNL errors.

[0042] Step S200.4.1, using a weighted correction algorithm, the INL and DNL predicted values ​​output by the neural network model are compared with a small amount of real data actually collected, specifically:

[0043] In the formula, is the final INL and DNL error after correction, is the correction coefficient, set to 0.8, which is used to control the weight of the predicted value and the true value in the final result. are the INL and DNL error values ​​predicted by the neural network model, It is the INL and DNL error value actually collected.

[0044] Step S200.4.2, verify and compare the prediction results based on the trained neural network model.

[0045] The verification result is: if the error is too large, the correction coefficient is adjusted, and the error is within the expected range of 2%.

[0046] In some specific implementations, based on the calculation results of the three-layer neural network model in S200, error prediction and optimization of the three-layer neural network model are performed, specifically including: Step S300.1: Based on the current partial data, historical data and environmental parameters obtained in step S100, input them into the trained neural network model for error prediction.

[0047] Step S300.1.1, obtain the predicted value based on the three-layer neural network operation.

[0048] The three-layer neural network operation is based on the overall content described in step S200.

[0049] Step S300.1.2: If the prediction error of the neural network model exceeds the preset threshold of 2%, a feedback mechanism is used to optimize the neural network model parameters.

[0050] The optimization is performed using a loss function and a stochastic gradient descent algorithm.

[0051] Based on the error prediction value obtained by S300, complete the ADC product data and make adjustments, including: Step S400.1: compare the error prediction value with the current partial data and complete the currently unsampled part.

[0052] The unsampled data is completed based on the neural network model, the current partial data and the historical data, and the INL and DNL error values ​​of the uncollected data points in the current partial data are inferred, and a complete data set is generated, specifically: In the formula, and is the predicted value of INL and DNL after completion, and are the INL and DNL values ​​in the historical data, and It is the actual INL and DNL data currently collected. It is a correction coefficient used to balance the relationship between historical data and real data. It is adjusted based on the actual difference. The initial value can be set to 0.8.

[0053] The INL and DNL data of the entire ADC that are not currently sampled are inferred based on the error prediction value, and the INL and DNL data are completed.

[0054] Neural Network The neural network model learns from historical data and identifies the INL and DNL trends under similar environmental parameters to complete the INL and DNL data.

[0055] In some specific implementations, based on the completed ADC product data, the completed data parameters are obtained, specifically including: Step S500.1: Calculate the final parameters of the ADC product based on the INL and DNL data completed in step S400 and the current partial data.

[0056] Among them, based on the completed INL data, the final INL parameter is obtained by integral calculation, and the final INL parameter is obtained, specifically:

[0057] Where: is the final INL value of input point i, which represents the calculation result of integral nonlinearity error. is the ideal output voltage value of input point i, indicating the output without error. is the actual measured output voltage value at input point i, indicating the actual output when there is an error. is the minimum input voltage of the ADC, indicating the lower limit of the input voltage, is the maximum input voltage of the ADC, indicating the upper limit of the input voltage. dV is a small increment of the input voltage, which is a variable used for integral calculation.

[0058] Among them, based on the completed DNL data, a differential calculation is performed to obtain the final DNL parameters, and the final DNL parameters are obtained, specifically:

[0059] Where: is the final DNL value of code value i, which represents the calculation result of differential nonlinearity error. is the actual output voltage of the adjacent code value i+1, indicating the actual measured output of the adjacent point, is the actual output voltage of code value i, indicating the actual measured output at the current point, It is the voltage difference between adjacent code values ​​under ideal conditions, indicating the ideal voltage interval between adjacent code values ​​when there is no error.

[0060] Step S500.2: Based on the final INL and DNL parameters obtained, verify whether the INL and DNL parameters exceed 2% by comparing with the standard values, and then dynamically adjust and correct them. and The coefficients are:

[0061] Where: It is the correction factor for INL, the correction factor after dynamic adjustment, is the initial correction factor of INL, indicating the initial value without dynamic adjustment, usually 0.8. is the temperature influence coefficient, which is used to adjust the correction coefficient to reflect the influence of temperature on the error. is the current ambient temperature, is the reference temperature, usually the standard operating temperature, It is the correction factor for DNL, ​​the correction factor after dynamic adjustment, is the initial correction factor of DNL, ​​which indicates the initial value without dynamic adjustment, usually 0.8.

[0062] In some specific implementations, feedback processing is performed based on the completed data parameters obtained in S500, specifically including: Step S600.1: Generate a relevant report based on the final INL and DNL parameter data.

[0063] The relevant report content includes: INL parameter data, DNL parameter data and calculated values, corrected values ​​and error ranges in the calculation process.

[0064] Step S600.2: Sort out the related reports obtained, set the INL parameter data and DNL parameter data as high priority, and layout them at the top.

[0065] Step S600.3: back up and store the obtained relevant reports in a database, and output them through an API interface.

[0066] In summary, the present application provides a method for fast measurement of INL and DNL parameters of ADC products by ATE, which obtains ADC product data and environmental parameters, constructs a three-layer neural network based on the data obtained in S100, performs error prediction and optimization based on the three-layer neural network model constructed in S200, completes the current unsampled data and adjusts it based on the error prediction value obtained in S300, obtains the completed data parameters based on the completed data, performs feedback processing based on the completed data parameters obtained in S500, uses a small amount of collected data, and performs error prediction and completion through a neural network model, which significantly reduces the amount of test data collected, greatly reduces test time and cost, and at the same time, the method ensures the accuracy of the test results by completing and correcting the data of the uncollected part, is suitable for ADC products with high integration and high frequency, and effectively improves the test efficiency and consistency of product performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is an overall flow chart of a method for quickly measuring INL and DNL parameters of ADC products by ATE, provided in an embodiment of the present application. DETAILED DESCRIPTION

[0068] Please refer to Figure 1 , which shows a process of an embodiment of a method for fast measuring INL and DNL parameters of ADC products by ATE according to the present disclosure like Figure 1 As shown, a method for quickly measuring INL and DNL parameters of ADC products by ATE includes the following steps: Step S100, obtaining ADC product data and environmental parameters and preprocessing; Step S200, using a three-layer neural network model to perform preliminary calculations on the data obtained in S100; Step S300: Based on the calculation result of the three-layer neural network model in S200, error prediction and optimization of the three-layer neural network model are performed; Step S400, completing ADC product data and making adjustments based on the error prediction value obtained in S300; Step S500, obtaining a completed data parameter based on the completed ADC product data; Step S600: performing feedback processing based on the completed data parameters obtained in S500.

[0069] In some specific implementations, obtaining ADC product data and environmental parameters and preprocessing specifically include: Step S100.1: Acquire historical data and current partial data of ADC products based on ATE equipment.

[0070] Among them, the historical data include: INL and DNL data.

[0071] Among them, the current partial data are: partial INL and DNL data.

[0072] Step S100.2: Obtain test environment parameters.

[0073] The environmental parameters include: temperature and voltage fluctuations.

[0074] The temperature is obtained based on a temperature sensor.

[0075] The voltage fluctuation is obtained based on a voltage monitoring device.

[0076] Step S100.3: pre-processing is performed based on the obtained historical data, current partial data and environmental parameters.

[0077] It should be noted that the preprocessing includes: data clarification and standardization processing, wherein the preprocessing is an existing technology well known to those skilled in the art and will not be elaborated here.

[0078] In some specific implementations, a three-layer neural network model is used to perform preliminary operations on the data obtained in S100, specifically including: Steps Step S200.1. Based on the acquired historical data, divide the historical data into: a training set, a validation set and a test set.

[0079] Step S200.1.1, construct a neural network model.

[0080] The three-layer neural network comprises an input layer, a hidden layer and an output layer.

[0081] The input layer is: the collected ADC product input signal characteristics.

[0082] Among them, the signal characteristics include: input voltage, output value, temperature, and frequency.

[0083] The hidden layer is used to capture complex nonlinear relationships.

[0084] Among them, the hidden layer adopts ReLU activation function and contains 20 neurons.

[0085] The output layer is used for linear activation function.

[0086] Among them, the output layer contains 2 neurons, corresponding to the predicted values ​​of INL and DNL respectively.

[0087] Based on the configured three-layer neural network model, historical data is input, and the characteristics are , the output of each layer is:

[0088]

[0089] In the formula, is the weight matrix, is the bias term, X is the input feature vector, X includes voltage, output, and temperature, and the ReLU activation function processes nonlinear features.

[0090] Step S200.2: Train the neural network model based on the training set and obtain the predicted values ​​of INL and DNL.

[0091] Step S200.2.1, using a loss function, comparing the predicted value obtained by the neural network model with the current partial data to minimize the difference between the two, the loss function is specifically:

[0092] In the formula, is the loss function, n is the number of samples in the dataset, is the predicted INL value of the ith data point, is the true INL value of the ith data point, is the predicted DNL value of the ith data point, is the true DNL value of the ith data point, is the square of the INL error of the ith data point, is the square of the DNL error of the ith data point.

[0093] Step S200.2.2, use the stochastic gradient descent algorithm to update the weights in the neural network model, specifically:

[0094] In the formula, is the weight matrix during the tth round of training, is the updated weight after the t+1th round of training, is the learning rate, which controls the step size of each weight update. is the partial derivative of the loss function with respect to the current weight and measures the impact of the current weight on the loss.

[0095] Step S200.2.3, use the validation set data to evaluate the trained neural network model.

[0096] In the evaluation, if the loss value of the neural network model on the validation set does not decrease by more than 5%, the number of training rounds is increased, and the number of increased training rounds is: 50 to 100 rounds.

[0097] Step S200.3, based on the test set and the trained neural network model, perform an error prediction test on the current part of the data described in step S100.2, and test the INL and DNL data of the unsampled part completed by the neural network model.

[0098] The error prediction is specifically:

[0099] In the formula, is the predicted INL and DNL value corresponding to the i-th input data point, is the output function processed by the neural network model, and the input is the data point , neural network model weights W and bias b, is the input feature of the i-th sampling point, including input voltage, output signal, and temperature.

[0100] The unsampled part is completed, and the result is predicted by using a neural network model to infer the error of the unsampled data point.

[0101] Step S200.4: Based on the prediction results of the validation set and the neural network model, further correct the INL and DNL errors.

[0102] Step S200.4.1, using a weighted correction algorithm, the INL and DNL predicted values ​​output by the neural network model are compared with a small amount of real data actually collected, specifically:

[0103] In the formula, is the final INL and DNL error after correction, is the correction coefficient, set to 0.8, which is used to control the weight of the predicted value and the true value in the final result. The neural network model predicts the INL and DNL error values. It is the INL and DNL error value actually collected.

[0104] Step S200.4.2, verify and compare the prediction results based on the trained neural network model.

[0105] The verification result is: if the error is too large, the correction coefficient is adjusted, and the error is within the expected range of 2%.

[0106] In some specific implementations, based on the calculation results of the three-layer neural network model in S200, error prediction and optimization of the three-layer neural network model are performed, specifically including: Step S300.1: Based on the current partial data, historical data and environmental parameters obtained in step S100, input them into the trained neural network model for error prediction.

[0107] Step S300.1.1, obtain the predicted value based on the three-layer neural network operation.

[0108] The three-layer neural network operation is based on the overall content described in step S200.

[0109] Step S300.1.2: If the prediction error of the neural network model exceeds the preset threshold of 2%, a feedback mechanism is used to optimize the neural network model parameters.

[0110] The optimization is performed using a loss function and a stochastic gradient descent algorithm.

[0111] Based on the error prediction value obtained in step S300, the ADC product data is supplemented and adjusted. Based on the error prediction value obtained in step S300, the current unsampled data is supplemented and adjusted, specifically including: Step S400.1: compare the error prediction value with the current partial data and complete the currently unsampled part.

[0112] The unsampled data is completed based on the neural network model, the current partial data and the historical data, and the INL and DNL error values ​​of the uncollected data points in the current partial data are inferred, and a complete data set is generated, specifically:

[0113] In the formula, and is the predicted value of INL and DNL after completion, and are the INL and DNL values ​​in the historical data, and It is the actual INL and DNL data currently collected. It is a correction coefficient used to balance the relationship between historical data and real data. It is adjusted based on the actual difference. The initial value can be set to 0.8.

[0114] The INL and DNL data of the entire ADC that are not currently sampled are inferred based on the error prediction value, and the INL and DNL data are completed.

[0115] Neural Network The neural network model learns from historical data and identifies the INL and DNL trends under similar environmental parameters to complete the INL and DNL data.

[0116] In some specific implementations, based on the completed ADC product data, the completed data parameters are obtained, specifically including: Step S500.1: Calculate the final parameters of the ADC product based on the INL and DNL data completed in step S400 and the current partial data.

[0117] Among them, based on the completed INL data, the final INL parameter is obtained by integral calculation, and the final INL parameter is obtained, specifically:

[0118] Where: is the final INL value of input point i, indicating the calculation result of integral nonlinear error. is the ideal output voltage value of input point i, indicating the output without error. is the actual measured output voltage value at input point i, indicating the actual output when there is an error. is the minimum input voltage of the ADC, indicating the lower limit of the input voltage, is the maximum input voltage of the ADC, indicating the upper limit of the input voltage. dV is a small increment of the input voltage, which is a variable used for integral calculation.

[0119] Among them, based on the completed DNL data, a differential calculation is performed to obtain the final DNL parameters, and the final DNL parameters are obtained, specifically:

[0120] Where: is the final DNL value of code value i, which represents the calculation result of differential nonlinearity error. is the actual output voltage of the adjacent code value i+1, indicating the actual measured output of the adjacent point, is the actual output voltage of code value i, indicating the actual measured output at the current point, It is the voltage difference between adjacent code values ​​under ideal conditions, indicating the ideal voltage interval between adjacent code values ​​when there is no error.

[0121] Step S500.2: Based on the final INL and DNL parameters obtained, verify whether the INL and DNL parameters exceed 2% by comparing with the standard values, and then dynamically adjust and correct them. and The coefficients are:

[0122] Where: It is the correction factor for INL, the correction factor after dynamic adjustment, is the initial correction factor of INL, indicating the initial value without dynamic adjustment, usually 0.8. is the temperature influence coefficient, which is used to adjust the correction coefficient to reflect the influence of temperature on the error. is the current ambient temperature, is the reference temperature, usually the standard operating temperature, It is the correction factor for DNL, ​​the correction factor after dynamic adjustment, is the initial correction factor of DNL, ​​which indicates the initial value without dynamic adjustment, usually 0.8.

[0123] In some specific implementations, feedback processing is performed based on the completed data parameters obtained in S500, specifically including: Step S600.1: Generate a relevant report based on the final INL and DNL parameter data.

[0124] The relevant report content includes: INL parameter data, DNL parameter data and calculated values, corrected values ​​and error ranges in the calculation process.

[0125] Step S600.2: Sort out the related reports obtained, set the INL parameter data and DNL parameter data as high priority, and layout them at the top.

[0126] Step S600.3: back up and store the obtained relevant reports in a database, and output them through an API interface.

[0127] In the above content, in actual application, first, the historical INL and DNL data of ADC products are obtained based on ATE. These historical data reflect the performance of ADC under different environmental conditions, including operating voltage, temperature and load, etc. Then, a small amount of test data of current ADC products is collected through ATE as input data of the neural network model.

[0128] Next, a three-layer neural network model was constructed to predict the INL and DNL parameters. The input layer of the model receives feature data, including voltage, temperature, environmental parameters, etc. The hidden layer learns complex nonlinear error patterns through nonlinear activation functions, and the output layer generates predicted values ​​of INL and DNL.

[0129] During the model training phase, historical data is used to train the neural network model, and the weights and biases of the model are optimized by the stochastic gradient descent method to minimize the error between the predicted value and the actual measured value. During the training process, cross-validation technology is used to evaluate the generalization ability of the model to ensure that the model can maintain good prediction accuracy on unseen data.

[0130] After completing the model training, the neural network model is used to predict the INL and DNL errors of the small amount of test data currently collected. The model can not only predict the collected data, but also infer the uncollected data to complete the INL and DNL parameter completion within the entire input voltage range.

[0131] In order to improve the accuracy of the prediction, a weighted correction algorithm is used to combine the prediction results of the neural network model with the actual measured values. Specifically, the correction coefficient is set , the prediction results are adjusted by the following formula:

[0132] This formula ensures that the predicted value is closer to the actual value during the completion process, reducing the deviation caused by insufficient data.

[0133] After the data is completed, the INL and DNL parameters are further calculated. The total INL error within the input voltage range is obtained by integrating the completed INL data set. The DNL parameter is calculated by differential calculation to evaluate the local nonlinear error of each adjacent code value point.

[0134] If the errors of some test points are found to exceed the design tolerance (for example, ±2LSB) during the calculation process, these data points will be re-sampled or corrected by adjusting the neural network model to ensure that all INL and DNL parameters meet the product design requirements.

[0135] Finally, the system generates a test report containing INL and DNL parameters. The report records various test data, model prediction results, correction process, and the final calculated INL and DNL parameters. The report also includes information such as environmental parameters and test conditions to facilitate subsequent analysis and verification. The report will be stored in the database for future product quality tracking and further optimization.

Claims

1. A method for fast measurement of INL and DNL parameters of ADC products by ATE, characterized in that: The steps include: S100, obtaining ADC product data and environmental parameters and preprocessing; S200, using a three-layer neural network model to perform preliminary operations on the data obtained in S100; S300, based on the calculation result of the three-layer neural network model in S200, performing error prediction and optimizing the three-layer neural network model; S400, based on the error prediction value obtained in S300, completing ADC product data and making adjustments; S500, obtaining a completed data parameter based on the completed ADC product data; S600 , performing feedback processing based on the completed data parameters obtained in S500 .

2. The method for quickly measuring INL and DNL parameters of ADC products by ATE according to claim 1, characterized in that: Obtain ADC product data and environmental parameters and preprocessing, including: S100.

1. Obtain historical data and current partial data of ADC products based on ATE equipment; Among them, the historical data are: INL and DNL data; Among them, the current partial data are: partial INL and DNL data; S100.

2. Obtain test environment parameters; The environmental parameters include: temperature and voltage fluctuations; The temperature is obtained based on a temperature sensor; The voltage fluctuation is obtained based on a voltage monitoring device; S100.

3. Preprocessing is performed based on the obtained historical data, current partial data and environmental parameters.

3. The method for rapid measurement of INL and DNL parameters of ADC products by ATE according to claim 1, characterized in that: A three-layer neural network model is used to perform preliminary operations on the data obtained in S100, including: S200.

1. Based on the acquired historical data, divide the historical data into: a training set, a validation set and a test set; S200.1.

1. Configure a three-layer neural network model; Based on the configured three-layer neural network model, historical data is input, and the characteristics are , the output of each layer is: ； ； ； ； In the formula, is the weight matrix, is the bias term, is the input feature vector, Including voltage, output, temperature, the ReLU activation function is used to process nonlinear features; S200.

2. Training the three-layer neural network model based on the training set, and obtaining the predicted values ​​of INL and DNL; S200.2.

1. Use a loss function to minimize the difference between the predicted value obtained by the neural network model and the current partial data; S200.2.

2. Use the stochastic gradient descent algorithm to update the weights in the neural network model; S200.2.

3. Use validation set data to evaluate the trained neural network model; S200.

3. Based on the test set and the trained neural network model, perform an error prediction test on the current partial data described in S100.2, and test the completion of unsampled INL and DNL data by the neural network model; The error prediction is specifically: ； In the formula, It is The predicted INL and DNL values ​​for the input data points are: is the output function processed by the neural network model, and the input is the data point , model weight and bias , It is The input characteristics of the sampling points include input voltage, output signal, and temperature; The unsampled portion is completed, and the prediction result is made using a neural network model to infer the error of the unsampled INL and DNL data points; S200.4, further correcting the INL and DNL errors based on the prediction results of the validation set and the neural network model; S200.4.

1. Use a weighted correction algorithm to compare the INL and DNL prediction values ​​output by the neural network model with the current partial data, specifically: ； ； In the formula, is the final INL and DNL error after correction, is the correction coefficient, set to 0.8, which is used to control the weight of the predicted value and the true value in the final result. The model predicts and DNL error value, is the actual collected INL and DNL error value; S200.4.

2. Verify and compare the prediction results based on the trained neural network model.

4. The method for quickly measuring INL and DNL parameters of ADC products by ATE according to claim 1, characterized in that: Based on the calculation results of the three-layer neural network model in S200, error prediction and optimization of the three-layer neural network model are carried out, including: S300.

1. Based on the current partial data, historical data and environmental parameters obtained in S100, the trained neural network model is input to perform error prediction; S300.1.1, obtain the error prediction value based on the three-layer neural network operation; S300.1.

2. If the model prediction error exceeds the preset threshold of 2%, a feedback mechanism is used to optimize the neural network model parameters; The optimization is performed using a loss function and a stochastic gradient descent algorithm.

5. The method for rapid measurement of INL and DNL parameters of ADC products by ATE according to claim 1, characterized in that: Based on the error prediction value obtained by S300, complete the ADC product data and make adjustments, including: S400.

1. Compare the error prediction value with the current partial data and complete the currently unsampled part; The unsampled data is completed, based on the neural network model, the current partial data and the historical data, to infer the INL and DNL error prediction values ​​of the uncollected data points in the current partial data, and generate a complete data set; The INL and DNL data of the entire ADC that are not currently sampled are inferred based on the error prediction value, and the INL and DNL data are completed.

6. The method for rapid measurement of INL and DNL parameters of ADC products by ATE according to claim 1, characterized in that: Based on the completed ADC product data, the completed data parameters are obtained, including: S500.

1. Calculate the final parameters of the ADC product based on the completed INL and DNL data and the current partial data; Among them, based on the completed INL data, the final INL parameter is obtained by integral calculation; Among them, based on the completed DNL data, differential calculation is performed to obtain the final DNL parameters; S500.

2. Based on the final INL and DNL parameters obtained, verify whether the INL and DNL parameters exceed 2% by comparing with the standard values, and then dynamically adjust and correct them. and coefficient.

7. The method for rapid measurement of INL and DNL parameters of ADC products by ATE according to claim 1, characterized in that: Feedback processing is performed based on the completed data parameters obtained in S500, specifically including: S600.

1. Generate relevant reports based on the final INL and DNL parameter data; The relevant report content includes: INL parameter data, DNL parameter data and calculated values, corrected values, and error ranges during the calculation process; S600.

2. Arrange the obtained relevant reports, set the INL parameter data and DNL parameter data as high priority, and layout them at the top; S600.

3. The obtained relevant reports are backed up and stored in a database, and output through an API interface.

Images