MEMS gyroscope electrical performance test method based on deep learning
By deep learning -based MEMS gyruotics electrical performance testing methods, the problem of insufficient existing test methods was solved, and a comprehensive evaluation and applicability screening of the electro -performance of MEMS gyroscope was realized, and the reliability and accuracy of product delivery were improved. Essence
Patent Information
- Application Number
- CN202411365635.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-29
- Publication Date
- 2025-01-24
AI Technical Summary
The existing MEMS gyroscope testing methods are not perfect enough, resulting in large deviations in product quality status during delivery, and traditional testing methods are difficult to comprehensively detect complex behaviors and potential problems caused by the interaction of multiple performance indicators.
The MEMS gyroscope electrical performance testing method based on deep learning is adopted to achieve a comprehensive assessment of the electrical performance of the MEMS gyroscope by building a test database, establishing a deep learning model, measured electrical performance data, scoring and categorizing, and introducing a loss function to repair deviations. Assessment and suitability screening.
Improve the accuracy and comprehensiveness of the electronic performance testing of the MEMS gyroscope, reduce product quality deviation, can better meet the needs of different users, and enhance the reliability of product delivery.
Smart Images
Figure BDA0005066140690000021 
Figure BDA0005066140690000025 
Figure BDA0005066140690000026
Abstract
Description
Technical Field
[0001] The present invention relates to the field of MEMS gyroscope testing, and in particular to a MEMS gyroscope electrical performance testing method based on deep learning. Background Art
[0002] In the existing MEMS gyroscope testing process, people often only focus on the functional verification of a single electrical performance indicator, and easily ignore the complex behaviors and potential problems caused by the interaction of multiple performance indicators. In this case, even if each performance indicator performs normally in an isolated test, there may still be a risk of insufficient detection at the system level. In addition, MEMS gyroscopes are widely used in a wide range of fields, and users involve aerospace, weapons, industry, medical and other fields. In the actual delivery process of products, products of the same model and the same batch, or the same model and different batches, often receive different feedback and evaluations from different users, or even completely opposite results. Therefore, if the existing process production process cannot be further improved, it is still necessary to test and screen out products suitable for different users. Summary of the invention
[0003] The purpose of the present invention is: to address the deficiencies in the prior art, the present invention proposes a MEMS gyroscope electrical performance testing method based on deep learning, which is used to solve the problem that traditional testing methods are not perfect and the product quality status has large deviations during the delivery process.
[0004] The technical solution of the present invention is:
[0005] A method for testing electrical performance of a MEMS gyroscope based on deep learning, comprising the steps of:
[0006] S1. Build a MEMS gyroscope test database and construct a data model for historical test results, including data models of products with high user satisfaction and product data models of user returns, as training data for deep learning;
[0007] S2. Establish a deep learning model, train it on historical typical feature data, and automatically extract data features on this basis. In the uninterrupted self-learning process, find the most suitable features, and then obtain a model with a higher score, and verify the accuracy of the model;
[0008] S3. According to the actual production test tasks, measure a batch of MEMS gyroscope electrical performance data; according to the standardized test content, use the MEMS gyroscope automatic acquisition test system to obtain XX amount of script data;
[0009] S4. Use the deep learning model to score and classify the measured electrical performance data of MEMS gyroscopes to obtain products suitable for different users;
[0010] S5. Feedback information from users after using the product is used to introduce a loss function into the model, correct the deviations in the screening process, and obtain more accurate data as much as possible.
[0011] Preferably, the deep learning model in step S2 introduces a linear relationship score function f(x, W); the parameters of the score function f(x, W) include the category of the data, the weight parameter, and the bias parameter for fine-tuning, as follows:
[0012] f(x,W)=W*x+b
[0013] The parameter x is expressed as the category matrix of gyroscope data. The MEMS gyroscope collects electrical performance data scripts, including zero bias, zero bias stability, scaling factor, and zero bias scaling temperature coefficient. Each indicator has a different impact on product performance. The number of corresponding weight parameters W should be the same as the category parameter. The introduction of parameter b is a fine-tuning parameter matrix, which plays a role in fine-tuning the overall result.
[0014] Preferably, the loss function in step S5 is expressed as:
[0015]
[0016] The regularization penalty term is:
[0017] Preferably, in step S5, after obtaining certain data results, the results are evaluated; in the decision-making process, the model is introduced to prevent the weight value of the W matrix of the optimization iteration from being too large at certain points, resulting in an erroneous evaluation of the results;
[0018] For the above loss function, if If It means that the result has loss; the regularization penalty term is only related to the weight parameter matrix W, and the coefficient λ is used to adjust the overfitting model.
[0019] Preferably, in order to better solve the classification problem, the evaluation result score value is converted into a probability value;
[0020]
[0021] Calculate the loss value: L i = -logP(Y = k | X = x i );
[0022] Optimize loss and reduce loss:
[0023]
[0024] The method behind gradient descent is to randomly select a combination of parameters (θ0, θ1, ..., θ n ), calculate the cost function, and then look for the next parameter combination that can reduce the cost function value the most, until a local minimum is found.
[0025] The advantages of the present invention are:
[0026] The MEMS gyroscope electrical performance testing method based on deep learning proposed in the present invention is used to solve the problem that traditional testing methods are not perfect and the product quality status has large deviations during the delivery process. DETAILED DESCRIPTION
[0027] The electrical performance testing method of a MEMS gyroscope based on deep learning proposed by the present invention comprises the following steps:
[0028] S1. Build a complete MEMS gyroscope test database and construct a data model based on the traditional test results, including product data models with high user satisfaction and product data models with user returns as training data for deep learning.
[0029] S2. Establish a deep learning model by training on a large amount of historical typical feature data, and automatically extract data features on this basis. In the uninterrupted self-learning process, find the most suitable features, and then obtain a model with a higher score and verify the accuracy of the model.
[0030] The deep learning model in S2 introduces a linear score function. The function parameters include the data category, weight parameters, and bias parameters for fine-tuning, as follows:
[0031] f(x,W)=W*x+b;
[0032] The parameter x is expressed as the category matrix of gyroscope data. The MEMS gyroscope collects electrical performance data scripts, including multiple parameters such as zero bias, zero bias stability, scale factor, zero bias scale temperature coefficient, etc. Each indicator has a different impact on product performance. The number of corresponding weight parameters W should be the same as the category parameter. The introduction of parameter b is a fine-tuning parameter matrix, which plays a role in fine-tuning the overall result.
[0033] S3. According to the actual production test tasks, a batch of MEMS gyroscope electrical performance data is measured. According to the detailed and standardized test content, the MEMS gyroscope automatic acquisition test system is used to obtain XX amount of script data.
[0034] S4. Use deep learning models to score and classify the measured electrical performance data of MEMS gyroscopes to obtain products suitable for different users.
[0035] S5. Feedback information from users after using the product is used to introduce a loss function into the model, correct the deviations in the screening process, and obtain more accurate data as much as possible.
[0036]
[0037] The regularization penalty term is:
[0038] After obtaining certain data results, the results need to be evaluated. In the decision-making process, this model is introduced to prevent the weight value of the W matrix of the optimization iteration from being too large at some points, resulting in an incorrect evaluation of the results. For the above function, if If It means that the result has loss; the regularization penalty term is only related to the weight parameter matrix W, and the coefficient λ is used to adjust the overfitting model.
[0039] In order to better solve the classification problem, the evaluation result score is converted into a probability value.
[0040]
[0041] Calculate the loss value: L i = -logP(Y = k | X = x i );
[0042] Optimize loss, reduce loss (gradient descent)
[0043]
[0044] The idea behind gradient descent is to randomly select a combination of parameters (θ0, θ1, ..., θ n ), calculate the cost function, and then find the next parameter combination that can reduce the cost function value the most. Continue doing this until you find a local minimum.
[0045] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any modifications made according to the spirit of the main technical solution of the present invention should be included in the protection scope of the present invention.
Claims
1. A MEMS gyroscope electrical performance testing method based on deep learning, characterized in that: Includes steps: S1. Build a MEMS gyroscope test database and construct a data model for historical test results, including data models of products with high user satisfaction and product returns, as training data for deep learning; S2. Establish a deep learning model, train it on historical typical feature data, and automatically extract data features on this basis. In the uninterrupted self-learning process, find the most suitable features, and then obtain a model with a higher score, and verify the accuracy of the model; S3. According to the actual production test tasks, measure a batch of MEMS gyroscope electrical performance data; according to the standardized test content, use the MEMS gyroscope automatic acquisition test system to obtain XX amount of script data; S4. Use the deep learning model to score and classify the measured electrical performance data of MEMS gyroscopes to obtain products suitable for different users; S5. Feedback information from users after using the product is used to introduce a loss function into the model, correct the deviations in the screening process, and obtain more accurate data as much as possible.
2. The MEMS gyroscope electrical performance testing method based on deep learning according to claim 1, characterized in that: The deep learning model in step S2 introduces a linear relationship score function f(x, W); the parameters of the score function f(x, W) include the data category, weight parameters, and bias parameters for fine-tuning, as follows: f(x,W)=W*x+b The parameter x is expressed as the category matrix of gyroscope data. The MEMS gyroscope collects electrical performance data scripts, including zero bias, zero bias stability, scaling factor, and zero bias scaling temperature coefficient. Each indicator has a different impact on product performance. The number of corresponding weight parameters W should be the same as the category parameter. The introduction of parameter b is a fine-tuning parameter matrix, which plays a role in fine-tuning the overall result.
3. The MEMS gyroscope electrical performance testing method based on deep learning according to claim 2, characterized in that: The loss function in step S5 is expressed as: The regularization penalty term is:
4. The MEMS gyroscope electrical performance testing method based on deep learning according to claim 3, characterized in that: In step S5, after obtaining certain data results, the results are evaluated; In the decision-making process, this model is introduced to prevent the weight value of the W matrix of the optimization iteration from being too large at some points, which may lead to an incorrect evaluation of the results; For the above loss function, if If It means that the result is lost; the regularization penalty term is only related to the weight parameter matrix W, and the coefficient λ is used to adjust the overfitting model.
5. The MEMS gyroscope electrical performance testing method based on deep learning according to claim 4, characterized in that: In order to better solve the classification problem, the evaluation result score is converted into a probability value; Calculate the loss value: L i = -logP(Y = k | X = x i ); Optimize loss and reduce loss: The method behind gradient descent is to randomly select a combination of parameters (θ0, θ1, ..., θ n ), calculate the cost function, and then look for the next parameter combination that can reduce the cost function value the most, until a local minimum is found.