Three-component sensor orthogonal correction algorithm based on conditional generative adversarial network

Abstract

The embodiment of the invention provides a three-component sensor orthogonal correction algorithm based on a conditional generative adversarial network, and the algorithm comprises the steps: measuring the orthogonal and non-orthogonal measurement values of a plurality of groups of three-axis sensors in different physical environments, and building a database; based on the database, a conditional generative adversarial network is used for training and learning orthogonal features of the three-component sensor, the conditional generative adversarial network comprises a generator and a discriminator, the generator is used for simulating three-component measured values corrected by the three-axis sensor, and the discriminator is used for discriminating the three-component measured values; the discriminator is used for judging whether the generated measurement value can be distinguished from a given true value or not; performing optimization through mutual game of the generator and the discriminator to obtain a trained conditional generative adversarial network; and based on the trained conditional generative adversarial network, correcting a sensor measurement value needing to be corrected, and outputting a result after orthogonal correction.

Description

Technical Field

[0001] The present invention relates to the technical field of three-component sensor application, and in particular to an orthogonal correction algorithm for a three-component sensor based on a conditional generative adversarial network. Background Art

[0002] A three-component sensor is a sensor that can measure physical quantities on three axes, such as a three-component accelerometer, a three-component gyroscope, and a three-component magnetic sensor. In theory, the three axes of a three-component sensor should be completely orthogonal to ensure the correctness of the output results. However, due to limitations in processing technology, sensor aging and loss, etc., the three axes in a three-axis sensor are not strictly orthogonal in actual applications.

[0003] The non-orthogonality of the three-component sensor may cause the following problems: 1. Non-orthogonality may cause deviations in the measurement results of certain axes; 2. Non-orthogonality may cause cross-coupling, that is, when measuring the physical quantity on one axis, the physical quantity on other axes may cause errors in the measurement axis; 3. In applications that require high precision, the error accumulation caused by non-orthogonality may cause serious deviations in the final results.

[0004] Therefore, in order to ensure the accuracy and stability of the measurement results, it is very necessary to ensure that the three axes of the three-axis sensor are strictly orthogonal. If the orthogonal correction of the sensor is only achieved by mechanical processing technology, it will not only be costly, but also may not be able to achieve the required accuracy due to the limitations of the existing process. Therefore, it is necessary to design an orthogonal correction algorithm for a three-component sensor. Summary of the invention

[0005] In view of the above problems existing in the prior art, an embodiment of the present invention provides an orthogonal correction algorithm for a three-component sensor based on a conditional generative adversarial network, which aims to solve the accuracy problem caused by the non-strict orthogonality of the three axes in the three-component sensor. First, the orthogonal and non-orthogonal measurement values ​​of multiple groups of three-axis sensors in different physical environments are measured to establish a database; then, on this basis, the orthogonal features of the three-component sensor are trained and learned using a conditional generative adversarial network; wherein the conditional generative adversarial network includes two modules, a generator and a discriminator; the generator simulates the three-component measurement value after correction of the three-axis sensor, and the discriminator determines whether the generated measurement value can be distinguished from the given true value; the generator and the discriminator compete with each other for optimization; finally, the stability and accuracy of the orthogonal value corrected by the generator are high, and the training result is good. The performance of the three-component sensor corrected by this method is significantly improved.

[0006] The embodiment of the present invention provides a three-component sensor orthogonal correction algorithm based on a conditional generative adversarial network, including:

[0007] Step 1: Measure the orthogonal and non-orthogonal measurement values ​​of multiple sets of three-axis sensors in different physical environments and establish a database;

[0008] Step 2: Based on the database, a conditional generative adversarial network is used to train and learn the orthogonal features of the three-component sensor, wherein the conditional generative adversarial network includes a generator and a discriminator, wherein the generator is used to simulate the corrected three-component measurement value of the three-axis sensor, and the discriminator is used to determine whether the generated measurement value can be distinguished from the given true value;

[0009] Step 3: The generator and the discriminator are optimized through mutual competition to obtain a trained conditional generative adversarial network;

[0010] Step 4: Generate an adversarial network based on the trained condition, correct the sensor measurements that need to be corrected, and output the orthogonal corrected results.

[0011] In some embodiments of the present invention, in step 1, the algorithm includes:

[0012] All sensors are measured, and the data from all sensors together form the data set Q, which is used to accumulate data for subsequent CGAN training and learning, including:

[0013] Step 11: Under the condition that the external physical environment remains unchanged, rotate N orthogonal sensors around x, y, and z respectively, 15° each time, to obtain 36 sets of data. Repeat the experiment 5-10 times to form data set Q1;

[0014] Step 12: Under the condition that the external physical environment remains unchanged, rotate the N non-orthogonal sensors around x, y, and z respectively, 15° each time, to obtain 36 sets of data. Repeat the experiment 5-10 times to form data set Q2.

[0015] Step 13: Set the dataset label in Q1 to "1" and the label in Q2 to 0, and construct a labeled dataset Q = [(x1, y1), (x2, y2), ..., (x m ,y m )], where x∈Q1,y=1 or x∈Q2,y=0.

[0016] In some embodiments of the present invention, in step 2, the algorithm includes:

[0017] Using the labeled dataset Q, a conditional generative adversarial network is constructed and trained to obtain a conditional generative adversarial network CGAN that can correct non-orthogonal sensors to orthogonal ones.

[0018] In some embodiments of the present invention, step 2 specifically includes:

[0019] Step 21: Initialize the conditional generative adversarial network;

[0020] Step 22: The generator uses the U-net encoder-decoder to perform encoding operations;

[0021] Step 23: The generator performs a decoding operation to obtain orthogonally corrected data;

[0022] Step 24: The discriminator uses a four-layer convolution structure to determine whether the input three-component sensor data is orthogonal data, and constructs a loss function of the conditional generative adversarial network;

[0023] Step 25: Construct the loss function G in the conditional generative adversarial network:

[0024] G = argm G inm D ax c GAN(G,D)+λL l1 (G)

[0025] Among them, λ is a constant, which represents the proportion of the generator's loss function in the entire conditional generative adversarial network;

[0026] L c GAN(G,D)=E x,y [logD(x,y)]+E x,z [log1-D(x,G(x,z))] is the objective function of the conditional generative adversarial network;

[0027] L l1 =E x,y,x [||yG(x,z)||] is the loss function of the generator;

[0028] G is the orthogonal correction result 0 generated by the generator, D is the discriminant result of the corrected data; (x, y) is the orthogonal correction structure, x is the data before orthogonal correction, y is the data after orthogonal correction; G(x, z) is the orthogonal data generated by the generator under the conditions of random noise z and non-orthogonal data x;

[0029] D(x,y) is the probability that the discriminator determines that the transformation deconstruction of x and y is true under the orthogonal correction structure (x,y).

[0030] 1-D(x,G(x,z)) is the probability that the discriminator judges the conversion deconstruction of x and G(x,z) to be false, and E is the expectation;

[0031] Step 26: Use the Adam optimizer to optimize the loss function of the conditional generative adversarial network.

[0032] In some embodiments of the present invention, step 21 specifically includes:

[0033] During initialization, for both the generator and the discriminator, the initial mean of the weight parameters is set to 0, the variance is set to a Gaussian distribution of 0.01, and the bias parameters of the generator and the discriminator are also set to 0.

[0034] In some embodiments of the present invention, step 22 specifically includes:

[0035] Using a 3×3 convolution kernel, a generative adversarial network with an 8-layer convolution structure is built to downsample the non-orthogonal sensors. After sampling, the number of feature channels is doubled to complete the encoding and obtain the corrected data.

[0036] In some embodiments of the present invention, step 23 specifically includes:

[0037] The generator uses an 8-layer deconvolution structure to upsample the encoded image, and symmetrically connects each layer of the deconvolution structure and the convolution structure to complete the decoding. The encoder and decoder are connected with a U-net structure.

[0038] In some embodiments of the present invention, step 4 specifically includes:

[0039] The sensor measurement value C that needs to be corrected is input into the trained conditional generative adversarial network, and the orthogonal corrected result is output.

[0040] Compared with the prior art, the three-component sensor orthogonal correction algorithm based on conditional generative adversarial network provided by the embodiment of the present invention has the following beneficial effects: traditional orthogonal correction algorithms are usually based on linear models and have difficulty in processing complex nonlinear relationships. In practical applications, there may be nonlinear associations between sensors, and traditional algorithms may not be able to accurately capture these complex relationships. Conditional generative adversarial network is a deep learning model that can learn and capture nonlinear relationships in data, which makes it more flexible in processing complex associations and nonlinear corrections of sensor data.

[0041] Traditional orthogonal correction algorithms have difficulty handling multimodal generation tasks, especially when the sensor data types are diverse or multiple conditions need to be considered. Conditional Generative Adversarial Networks support multimodal generation by introducing conditional information, and can generate different sensor corrections according to different conditions.

[0042] Traditional algorithms require multiple stages of processing, including data preprocessing, orthogonalization, etc. Some algorithms also have more stringent assumptions and even require manual adjustment of data. Conditional Generative Adversarial Networks use end-to-end learning to automatically learn and correct tasks by jointly training the generator and discriminator. This end-to-end learning method reduces manual intervention, making the model more adaptive and better able to adapt to different data characteristics and task requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 An operation flow chart of a three-component sensor orthogonal correction algorithm based on a conditional generative adversarial network provided by an embodiment of the present invention;

[0044] Figure 2 A flow chart of a conditional generative adversarial network training orthogonal algorithm in a three-component sensor orthogonal correction algorithm based on a conditional generative adversarial network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0046] Various aspects and features of the present application are described herein with reference to the accompanying drawings.

[0047] These and other characteristics of the present application will become apparent from the following description of a preferred form of embodiment given as a non-limiting example with reference to the accompanying drawings.

[0048] It should also be understood that, although the present application has been described with reference to some specific examples, those skilled in the art will be able to realize many other equivalent forms of the present application that have the features described in the claims and are therefore within the scope of protection defined thereby.

[0049] The above and other aspects, features and advantages of the present application will become more apparent in view of the following detailed description when taken in conjunction with the accompanying drawings.

[0050] Specific embodiments of the present application are described hereinafter with reference to the accompanying drawings; however, it should be understood that the embodiments applied for are merely examples of the present application, which may be implemented in a variety of ways. Well-known and / or repeated functions and structures are not described in detail to determine the true intent based on the user's historical operations and to avoid unnecessary or redundant details that make the present application unclear. Therefore, the specific structural and functional details applied for herein are not intended to be limiting, but are merely used as the basis and representative basis for the claims to teach those skilled in the art to use the present application in a variety of ways with substantially any suitable detailed structure.

[0051] This specification may use the phrases "in one embodiment," "in another embodiment," "in yet another embodiment," or "in other embodiments," all of which may refer to one or more of the same or different embodiments according to the present application.

[0052] The embodiment of the present invention provides a three-component sensor orthogonal correction algorithm based on a conditional generative adversarial network, such as Figure 1 and Figure 2 As shown, the algorithm includes:

[0053] Step 1: Measure the orthogonal and non-orthogonal measurement values ​​of multiple sets of three-axis sensors in different physical environments and establish a database;

[0054] Step 2: Based on the database, a conditional generative adversarial network is used to train and learn the orthogonal features of the three-component sensor, wherein the conditional generative adversarial network includes a generator and a discriminator, wherein the generator is used to simulate the corrected three-component measurement value of the three-axis sensor, and the discriminator is used to determine whether the generated measurement value can be distinguished from the given true value;

[0055] Step 3: The generator and the discriminator are optimized through mutual competition to obtain a trained conditional generative adversarial network;

[0056] Step 4: Generate an adversarial network based on the trained condition, correct the sensor measurements that need to be corrected, and output the orthogonal corrected results.

[0057] In order to facilitate a detailed understanding of the above technical solution, the following is a detailed description with reference to the accompanying drawings and specific examples, as follows:

[0058] Step 1: Measure all sensors, and the data from all sensors together form a data set Q, which is used to accumulate data for subsequent CGAN training and learning. Step 1 mainly includes three steps.

[0059] Step 11: While the external physical environment remains unchanged, rotate N orthogonal sensors around x, y, and z respectively, 15° each time, to obtain 36 sets of data. Repeat the experiment 5-10 times to form data set Q1.

[0060] Step 12: While keeping the external physical environment unchanged, rotate N non-orthogonal sensors around x, y, and z respectively, 15° each time, to obtain 36 sets of data. Repeat the experiment 5-10 times to form data set Q2.

[0061] Step 13: Set the dataset labels in Q1 to "1" and the labels in Q2 to "0". Construct a labeled dataset Q = [(x1, y1), (x2, y2), ..., (x m ,ym )], where x∈Q1,y=1 or x∈Q2,y=0.

[0062] Step 2: Use the labeled dataset Q to construct and train a conditional generative adversarial network to obtain a conditional generative adversarial network CGAN that can correct non-orthogonal sensors to orthogonal ones. Step 2 mainly includes three steps:

[0063] Step 21: Initialize the conditional generative adversarial network. Specifically, the adversarial network consists of two parts: the generator and the discriminator. During initialization, for both the generator and the discriminator, the initial mean of the weight parameters is set to 0, and the variance is also set to a Gaussian distribution of 0.01. In addition, the bias parameters of the generator and the discriminator are also set to 0.

[0064] Step 22: The generator uses the U-net encoder-decoder to perform encoding operations. Specifically, a 3×3 convolution kernel is used to build an 8-layer convolutional structure of the generative adversarial network to downsample the non-orthogonal sensor, and after sampling, the number of feature channels is doubled to complete the encoding and obtain the corrected data.

[0065] Step 23: The generator performs decoding operation to obtain orthogonally corrected data. Specifically, the generator uses an 8-layer deconvolution structure to upsample the encoded image, symmetrically connects each layer of the deconvolution structure and the convolution structure to complete the decoding. The encoder and decoder are connected using a U-net structure.

[0066] Step 24: The discriminator uses a four-layer convolutional structure to determine whether the input three-component sensor data is orthogonal data, and constructs the loss function of the conditional generative adversarial network.

[0067] Step 25: Construct the loss function G in the conditional generative adversarial network:

[0068]

[0069] Among them, λ is a constant, which represents the proportion of the generator's loss function in the entire conditional generative adversarial network; L c GAN(G,D)=E x,y [logD(x,y)]+E x,z [log1-D(x,G(x,z))] is the objective function of the conditional generative adversarial network; is the loss function of the generator; G is the orthogonal correction result 0 generated by the generator, D is the discriminant result of the corrected data; (x, y) is the orthogonal correction structure, x is the data before orthogonal correction, y is the data after orthogonal correction; G(x, z) is the orthogonal data generated by the generator under the conditions of random noise z and non-orthogonal data x; D(x, y) is the probability that the discriminator judges that the conversion deconstruction of x and y is true under the orthogonal correction structure (x, y), 1-D(x, G(x, z)) is the probability that the discriminator judges that the conversion deconstruction of x and G(x, z) is false, and E is the expectation.

[0070] Step 26: Use the Adam optimizer to optimize the loss function of the conditional generative adversarial network.

[0071] Step 3: Input the sensor measurement value C that needs to be corrected into the trained conditional generative adversarial network and output the orthogonal corrected result.

[0072] It can be seen from the above technical solution that the beneficial effect of the three-component sensor orthogonal correction algorithm based on the conditional generative adversarial network provided by the above embodiment of the present invention is that the traditional orthogonal correction algorithm is usually based on a linear model and is difficult to handle complex nonlinear relationships. In practical applications, there may be nonlinear associations between sensors, and traditional algorithms may not be able to accurately capture these complex relationships. The conditional generative adversarial network is a deep learning model that can learn and capture nonlinear relationships in data, which makes it more flexible in processing complex associations and nonlinear corrections of sensor data.

[0073] Traditional orthogonal correction algorithms have difficulty handling multimodal generation tasks, especially when the sensor data types are diverse or multiple conditions need to be considered. Conditional Generative Adversarial Networks support multimodal generation by introducing conditional information, and can generate different sensor corrections according to different conditions.

[0074] Traditional algorithms require multiple stages of processing, including data preprocessing, orthogonalization, etc. Some algorithms also have more stringent assumptions and even require manual adjustment of data. Conditional Generative Adversarial Networks use end-to-end learning to automatically learn and correct tasks by jointly training the generator and discriminator. This end-to-end learning method reduces manual intervention, making the model more adaptive and better able to adapt to different data characteristics and task requirements.

[0075] The above embodiments are only exemplary embodiments of the present invention and are not intended to limit the present invention. The protection scope of the present invention is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present invention within the essence and protection scope of the present invention, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present invention.

Claims

1. A three-component sensor orthogonal correction algorithm based on conditional generative adversarial network, characterized in that: include: Step 1: Measure the orthogonal and non-orthogonal measurement values ​​of multiple sets of three-axis sensors in different physical environments and establish a database; Step 2: Based on the database, a conditional generative adversarial network is used to train and learn the orthogonal features of the three-component sensor, wherein the conditional generative adversarial network includes a generator and a discriminator, wherein the generator is used to simulate the corrected three-component measurement value of the three-axis sensor, and the discriminator is used to determine whether the generated measurement value can be distinguished from the given true value; Step 3: The generator and the discriminator are optimized through mutual competition to obtain a trained conditional generative adversarial network; Step 4: Generate an adversarial network based on the trained condition, correct the sensor measurements that need to be corrected, and output the orthogonal corrected results.

2. The three-component sensor orthogonal correction algorithm based on conditional generative adversarial network according to claim 1 is characterized in that: In step 1, the algorithm includes: All sensors are measured, and the data from all sensors together form the data set Q, which is used to accumulate data for subsequent CGAN training and learning, including: Step 11: Under the condition that the external physical environment remains unchanged, rotate N orthogonal sensors around x, y, and z respectively, 15° each time, to obtain 36 sets of data. Repeat the experiment 5-10 times to form data set Q1; Step 12: Under the condition that the external physical environment remains unchanged, rotate the N non-orthogonal sensors around x, y, and z respectively, 15° each time, to obtain 36 sets of data. Repeat the experiment 5-10 times to form data set Q2. Step 13: Set the dataset label in Q1 to "1", the label in Q2 to 0, and construct a labeled dataset Q = [(x1, y1), (x2, y2), ..., (x m ,y m )], where x∈Q1,y=1 or x∈Q2,y=0.

3. The three-component sensor orthogonal correction algorithm based on conditional generative adversarial network according to claim 2 is characterized in that: In step 2, the algorithm includes: Using the labeled dataset Q, a conditional generative adversarial network is constructed and trained to obtain a conditional generative adversarial network CGAN that can correct non-orthogonal sensors to orthogonal ones.

4. The three-component sensor orthogonal correction algorithm based on conditional generative adversarial network according to claim 3 is characterized in that: The step 2 specifically includes: Step 21: Initialize the conditional generative adversarial network; Step 22: The generator uses the U-net encoder-decoder to perform encoding operations; Step 23: The generator performs a decoding operation to obtain orthogonally corrected data; Step 24: The discriminator uses a four-layer convolution structure to determine whether the input three-component sensor data is orthogonal data, and constructs a loss function of the conditional generative adversarial network; Step 25: Construct the loss function G in the conditional generative adversarial network: G=argm G inm D axL c GAN(G,D)+λL l1 (G) Among them, λ is a constant, which represents the proportion of the generator's loss function in the entire conditional generative adversarial network; L c GAN(G,D)=E x,y [logD(x,y)]+E x,z [log1-D(x,G(x,z))] is the objective function of the conditional generative adversarial network; L l1 =E x,y,x [||yG(x,z)||] is the loss function of the generator; G is the orthogonal correction result 0 generated by the generator, D is the discriminant result of the corrected data; (x, y) is the orthogonal correction structure, x is the data before orthogonal correction, y is the data after orthogonal correction; G(x, z) is the orthogonal data generated by the generator under the conditions of random noise z and non-orthogonal data x; D(x,y) is the probability that the discriminator determines that the transformation deconstruction of x and y is true under the orthogonal correction structure (x,y). 1-D(x,G(x,z)) is the probability that the discriminator judges the conversion deconstruction of x and G(x,z) to be false, and E is the expectation; Step 26: Use the Adam optimizer to optimize the loss function of the conditional generative adversarial network.

5. The three-component sensor orthogonal correction algorithm based on conditional generative adversarial network according to claim 4 is characterized in that: The step 21 specifically includes: During initialization, for both the generator and the discriminator, the initial mean of the weight parameters is set to 0, the variance is set to a Gaussian distribution of 0.01, and the bias parameters of the generator and the discriminator are also set to 0.

6. The three-component sensor orthogonal correction algorithm based on conditional generative adversarial network according to claim 5 is characterized in that: The step 22 specifically includes: Using a 3×3 convolution kernel, a generative adversarial network with an 8-layer convolution structure is built to downsample the non-orthogonal sensors. After sampling, the number of feature channels is doubled to complete the encoding and obtain the corrected data.

7. The three-component sensor orthogonal correction algorithm based on conditional generative adversarial network according to claim 6 is characterized in that: The step 23 specifically includes: The generator uses an 8-layer deconvolution structure to upsample the encoded image, and symmetrically connects each layer of the deconvolution structure and the convolution structure to complete the decoding. The encoder and decoder are connected with a U-net structure.

8. The three-component sensor orthogonal correction algorithm based on conditional generative adversarial network according to claim 7 is characterized in that: The step 4 specifically includes: The sensor measurement value C that needs to be corrected is input into the trained conditional generative adversarial network, and the orthogonal corrected result is output.

Images