Joint calibration method for multiple micro inertial measurement components based on machine learning

Abstract

The invention discloses a joint calibration method for multiple micro inertial measurement components based on machine learning, and the method comprises the following steps: S1, obtaining original data of multiple inertial sensors, relates to the technical field of machine learning, and can effectively map high-dimensional data to multiple local subspaces by adopting principal component analysis and a clustering algorithm, so as to improve the accuracy of calibration of the multiple micro inertial measurement components; and the dimension of the local subspace is dynamically adjusted. In the process, through dynamic dimension reduction, the computing resource consumption of the data is effectively reduced. Particularly, on the aspect of high-dimension data processing, the problem of calculation bottleneck caused by too high data dimension in a traditional method is relieved by dynamically adjusting the dimension of the subspace. Each local subspace only pays attention to one specific error type. The calculation complexity caused by high-dimensional data is reduced; according to the method, the problem of high-dimension data is effectively relieved through subspace decomposition and dynamic subspace dimension adjustment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to a joint calibration method of a multi-micro inertial measurement assembly based on machine learning. BACKGROUND

[0002] The multi-micro inertial measurement assembly refers to an assembly composed of multiple micro inertial sensors, and the joint calibration is a process of simultaneously identifying and compensating for error parameters of each assembly and coupling errors between assemblies through a unified calibration framework and collaborative algorithm to improve the overall measurement accuracy of the system. At present, there is a method of using a deep learning model to predict the error of an inertial sensor and correct the calibration result of the inertial sensor.

[0003] The existing method of using a deep learning model to correct the calibration result of an inertial sensor has the following problems: Inertial sensors involve multi-dimensional time series data of accelerometers and gyroscopes, and have high feature dimension and strong coupling. When a traditional deep learning model directly processes original high-dimensional data, the computational load is too large to meet the real-time requirements of embedded platforms. Although the existing dimension reduction techniques PCA and t-SNE can compress the dimension, they do not consider the independence of physical characteristics such as bias, noise and scale factor of sensor errors, resulting in loss or confusion of key error features in the dimension reduction process, affecting the accuracy of the model. SUMMARY To solve the technical problems in the background art, the present application proposes a joint calibration method of a multi-micro inertial measurement assembly based on machine learning.

[0004] The joint calibration method of a multi-micro inertial measurement assembly based on machine learning proposed by the present application comprises the following steps: S1, obtaining original data of multiple inertial sensors to generate an original data set; obtaining a deep learning model, which is used to predict the error of the inertial sensor and correct the calibration result of the inertial sensor; As an illustration, the deep learning model used to predict the error of the inertial sensor and correct the calibration result of the inertial sensor is an existing deep learning model, so it can be directly obtained without construction. S2, decomposing the original data set into multiple local subspaces through a dynamic adaptive subspace adjustment mechanism; S3, constructing feature calibration parameters corresponding to error features in each local subspace; S4, inputting the feature calibration parameters corresponding to the error features in each local subspace into the deep learning model to generate a compensation value for each error feature through the deep learning model, correcting the corresponding error of the inertial sensor, and generating a correction result; S5, performing data fusion on the correction result to generate fused data as the final calibration result.

[0005] Preferably, in S2, the original data set is decomposed into multiple local subspaces by a dynamic adaptive subspace adjustment mechanism, as follows: Obtain the original data of multiple inertial sensors, generate an original data set, and perform statistical analysis on the original data set to obtain statistical features and generate a statistical feature set; The statistical features include mean, variance, and covariance; Statistical analysis is performed on the statistical feature set to extract error features in the statistical feature set; The error features include bias, noise, and scale factor error; The error features are mapped into multiple local subspaces using clustering algorithms or principal component analysis techniques, and each local subspace corresponds to a specific error feature.

[0006] Preferably, in S2, the dimension of the local subspace is also dynamically adjusted; Preferably, the dimension of the local subspace is dynamically adjusted as follows: Each local subspace obtains a real-time training error sequence and a historical training error sequence corresponding to a deep learning model; and time series analysis is performed on the real-time training error sequence and the historical training error sequence to obtain a real-time training error value and an error change rate of each local subspace; An absolute error threshold and an error change rate threshold are preset for each local subspace; When the real-time training error value of the local subspace is greater than the absolute error threshold, the dimension of the local subspace is increased by a preset proportion; When the real-time training error value of the local subspace is equal to the absolute error threshold: If the error change rate is greater than the error change rate threshold, the dimension of the local subspace is increased by a preset proportion; If the error change rate is less than or equal to the error change rate threshold, the dimension of the local subspace remains unchanged; When the real-time training error value of the local subspace is less than the absolute error threshold: If the error change rate is greater than the error change rate threshold, the dimension of the local subspace is increased by a preset proportion; If the error change rate is equal to the error change rate threshold, the dimension of the local subspace remains unchanged; If the error change rate is less than the error change rate threshold, the dimension of the local subspace is decreased by a preset proportion; The purpose of dynamically adjusting the dimension of the subspace is to adapt to the changes in data features; As an illustration, a deep learning model is generated by training a deep neural network according to historical data and calibration data of the inertial sensor, the deep learning model is used to predict errors of the inertial sensor and correct calibration results of the inertial sensor, an output result of the deep learning model is compared with actual inertial sensor data, so as to obtain a training error, the training error is used to represent a difference between the output result of the deep learning model and the actual inertial sensor data.

[0007] Preferably, in S3, in each local subspace, a feature calibration parameter corresponding to an error feature is constructed, as follows: For each local subspace, an original feature vector of the error feature in the local subspace is extracted; As an illustration, the error features include bias, noise, and scale factor error; regarding the original feature vector: The original feature vector extracted by the local subspace corresponding to the bias is a bias feature vector, the bias feature vector includes a sequence vector of zero bias values at each time point; The original feature vector extracted by the local subspace corresponding to the noise is a noise feature vector, the noise feature vector includes a statistical vector of noise amplitude and frequency distribution; The original feature vector extracted by the local subspace corresponding to the scale factor error is a scale factor feature vector, the scale factor feature vector includes a parameter vector of scale factor bias rate and non-linear coefficient; As a further illustration, other error features of the error features are not exhaustively listed here and can be directly extracted using existing technologies; The plurality of original feature vectors generate a set of to-be-fused features; The original feature vectors in the set of to-be-fused features are divided into a plurality of fixed-width intervals by equal-width discretization, and the probability of each fixed-width interval is counted; According to the probability of the fixed-width interval, mutual information values of any two original feature vectors in the set of to-be-fused features are obtained, the mutual information values are normalized, and a mutual information matrix is generated; Equal-width discretization is an existing discretization method; According to the mutual information matrix, a global correlation feature vector is generated by using a weighted average algorithm; The global correlation feature vector is mapped to a high-level subspace by linear transformation or a feature selection algorithm; The linear transformation can be selected as a matrix projection technology; The feature selection algorithm can be selected as a mutual information-based dimension screening; In each local subspace, for the error feature corresponding to the local subspace, a feature calibration parameter corresponding to the error feature is obtained; the feature calibration parameter corresponding to the error feature is independently adjusted according to a preset rule; The quality of the error features input into the deep learning model is improved by adjusting the feature calibration parameters corresponding to the error features. As an illustration: In the local subspace corresponding to the bias, the corresponding feature calibration parameter is the smoothing coefficient of the bias estimation. Specifically, by smoothing the zero-bias value sequence, such as smoothing the fluctuation sequence of 5 sampling points into a stable sequence of 10 sampling points, the bias feature input into the deep learning model is closer to the true error trend, helping the deep learning model to learn the bias correction rule more accurately. In the local subspace corresponding to the noise, the corresponding feature calibration parameter is the noise filtering window size. Specifically, by expanding the filtering window, such as expanding from 20ms to 50ms to filter out high-frequency impulse noise, the interference components in the noise feature vector are reduced, so that the deep learning model can learn the effective noise rule rather than noise interference. In the local subspace corresponding to the scale factor error, the corresponding feature calibration parameter is the non-linear correction segmentation interval. Specifically, by refining the segmentation, such as increasing from 10 segments to 20 segments, the non-linear representation accuracy of the scale factor feature vector is improved, providing the deep learning model with more delicate error gradient information and optimizing the model's fitting effect on the scale factor error. As a further illustration, for the local subspace corresponding to other error features not exhaustively listed, the feature calibration parameters corresponding to the error features can be implemented using the quantization adjustment rules disclosed in existing technologies in the field. In the high-level subspace, the global correlation feature vector is mined using correlation rules to generate correlation information data, and the feature calibration parameters corresponding to the error features in each local subspace are adjusted according to the correlation information data through a preset rule.

[0008] Preferably, in S5, it further includes: performing error correction on the fusion data actual value of the future unit time obtained in the future unit time to obtain a final fusion data actual value as a final calibration result.

[0009] Preferably, the fusion data actual value of the future unit time obtained in the future unit time is error-corrected to obtain a final fusion data actual value as a final calibration result, as follows: The fusion data of the historical correction result is obtained, and a long short-term memory network is used to generate a fusion data predicted value of the future unit time; When the time comes to the time of the future unit time, the fusion data actual value of the future unit time is obtained; The fusion data actual value of the future unit time is compared with the fusion data predicted value of the future unit time to obtain a residual error; As an illustration, the residual error is the difference between the actual value of the fused data of the future unit time and the predicted value of the fused data of the future unit time; The Kalman filtering algorithm is adopted, the residual error is taken as the observation input of the Kalman filtering, the Kalman gain is obtained through a preset residual noise model, the residual error is corrected through the Kalman gain, the product of the Kalman gain and the residual error is taken as the correction result, and the difference between the actual value of the fused data of the future unit time and the correction result is taken as the final fused data actual value; and the final fused data actual value is taken as the final calibration result of the actual value of the fused data of the future unit time.

[0010] As an illustration, the final fused data actual value is a vector; correspondingly, the actual value of the fused data of the future unit time, the residual error and the correction result are all vectors and have consistent dimensions. The residual noise model is a Gaussian white noise model. The residual error is corrected through the Kalman filtering algorithm, and the calibration accuracy is improved.

[0011] The joint calibration system of multiple micro inertial measurement assemblies based on machine learning comprises: A data acquisition module: acquiring original data of multiple inertial sensors to generate an original data set; and acquiring a deep learning model, which is used to predict errors of the inertial sensors and correct calibration results of the inertial sensors; A local subspace generation module: the original data set is decomposed into multiple local subspaces through a dynamic adaptive subspace adjustment mechanism; A feature calibration parameter generation module: in each local subspace, a feature calibration parameter corresponding to an error feature is constructed; A correction result generation module: the feature calibration parameter corresponding to the error feature in each local subspace is input into the deep learning model, a compensation value of each error feature is generated through the deep learning model, a corresponding error of the inertial sensor is corrected, and a correction result is generated; A calibration result generation module: the correction result is subjected to data fusion to generate fused data as a final calibration result; A future calibration result optimization module: an actual value of the fused data of the future unit time obtained in the future unit time is subjected to error correction to obtain a final fused data actual value as a final calibration result.

[0012] In the present application, the joint calibration method of multiple micro inertial measurement assemblies based on machine learning has the following beneficial technical effects: 1. By adopting principal component analysis and clustering algorithm, the application can effectively map high-dimensional data to multiple local subspaces and dynamically adjust the dimension of the local subspace. This process effectively reduces the computational resource consumption of data through dynamic dimension reduction. Especially in the processing of high-dimensional data, by dynamically adjusting the dimension of the subspace, the computational bottleneck problem caused by the high dimension of the data in the traditional method is alleviated. Each local subspace of the application only focuses on a specific error type. This approach can improve the processing accuracy of each error feature and avoid interference between different error types, reducing the computational complexity of high-dimensional data; the application effectively alleviates the problem of high-dimensional data by subspace decomposition and dynamic adjustment of subspace dimension, thereby improving the efficiency of deep learning model reasoning.

[0013] 2. The application can dynamically adjust the dimension of the subspace according to the real-time training error and historical error through a dynamic adaptive subspace adjustment mechanism, thereby effectively alleviating the overfitting problem of the deep learning model. Specifically, the dynamic adjustment mechanism enables each subspace to automatically adjust the dimension during the training process according to the error change, thereby alleviating the overfitting of the deep learning model to the training data. In the case of large error change, the system increases the subspace dimension to better fit the data features, while in the case of small error, it reduces the subspace dimension to reduce the model complexity and improve the generalization ability of the model. In addition, when using a deep learning model to correct errors, combining error feature calibration parameters further improves the quality of the features, ensuring that the error features input into the model can better match the true error law, effectively alleviating the overfitting phenomenon of using a deep learning model.

[0014] 3. By introducing the combination of long short-term memory network and Kalman filter algorithm, the error correction of future unit time fusion data is realized. By predicting future data and comparing it with the actual value, the residual error is calculated, and then the Kalman filter algorithm is used to accurately correct the error, thereby greatly improving the calibration accuracy. Especially in complex and variable sensor data environment, not only can it effectively deal with the error of data prediction, but also can dynamically correct the residual error through Kalman filter, to ensure that the final calibration result is more accurate. Through this error correction and real-time feedback, the stability and robustness of the calibration process are effectively improved, which adapts to the error characteristics and data fluctuations of different sensors. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 Flowchart of the joint calibration method of the multi-micro inertial measurement assembly based on machine learning of the application; Figure 2 Principle block diagram of the joint calibration system of the multi-micro inertial measurement assembly based on machine learning of the application. DETAILED DESCRIPTION

[0016] Embodiments of the present application are described below in the accompanying drawings, which are indicative of non-limiting examples of possible embodiments of the present application, wherein like or similar elements across the various figures are denoted by like or similar reference signs. The embodiments described below are examples only and are not intended to limit the present application in any way.

[0017] As Figure 1 indicated in the joint calibration method of the multi-micro inertial measurement assembly based on machine learning, comprising the following steps: S1, obtaining the original data of a plurality of inertial sensors to generate an original data set; obtaining a deep learning model, the deep learning model is used to predict the error of the inertial sensor and correct the calibration result of the inertial sensor; As an illustration, the deep learning model for predicting the error of the inertial sensor and correcting the calibration result of the inertial sensor is an existing deep learning model, so it can be directly obtained without construction; S2, decompose the original data set into a plurality of local subspaces through a dynamic adaptive subspace adjustment mechanism; S3, in each local subspace, construct a feature calibration parameter corresponding to an error feature; S4, input the feature calibration parameter corresponding to the error feature in each local subspace into the deep learning model, generate a compensation value for each error feature through the deep learning model, correct the corresponding error of the inertial sensor, and generate a correction result; S5, data fusion is performed on the correction result to generate fusion data as the final calibration result.

[0018] In an optional embodiment, in S2, the original data set is decomposed into a plurality of local subspaces through a dynamic adaptive subspace adjustment mechanism, as follows: Obtain the original data of a plurality of inertial sensors to generate an original data set, and perform statistical analysis on the original data set to obtain statistical features and generate a statistical feature set; The statistical features include mean, variance, and covariance; Statistical analysis is performed on the statistical feature set to extract error features in the statistical feature set; The error features include bias, noise, and scale factor error; The error features are mapped into a plurality of local subspaces using a clustering algorithm or principal component analysis technology, and each local subspace corresponds to a specific error feature.

[0019] In an optional embodiment, in S2, the dimension of the local subspace is also dynamically adjusted; The dimension of the local subspace is dynamically adjusted as follows: Each local subspace obtains a real-time training error sequence and a historical training error sequence corresponding to the deep learning model; the real-time training error sequence and the historical training error sequence are analyzed by using time series analysis to obtain a real-time training error value and an error change rate of each local subspace; An absolute error threshold and an error change rate threshold are preset for each local subspace; When the real-time training error value of the local subspace is greater than the absolute error threshold, the dimension of the local subspace is increased by a preset proportion; When the real-time training error value of the local subspace is equal to the absolute error threshold: If the error change rate is greater than the error change rate threshold, the dimension of the local subspace is increased by a preset proportion; If the error change rate is less than or equal to the error change rate threshold, the dimension of the local subspace is unchanged; When the real-time training error value of the local subspace is less than the absolute error threshold: If the error change rate is greater than the error change rate threshold, the dimension of the local subspace is increased by a preset proportion; If the error change rate is equal to the error change rate threshold, the dimension of the local subspace is unchanged; If the error change rate is less than the error change rate threshold, the dimension of the local subspace is decreased by a preset proportion; The purpose of dynamically adjusting the dimension of the subspace is to adapt to the change of data characteristics; As an illustration, according to historical data and calibration data of an inertial sensor, a deep neural network is trained to generate a deep learning model, which is used to predict the error of the inertial sensor and correct the calibration result of the inertial sensor. The output result of the deep learning model is compared with the actual inertial sensor data to obtain a training error, which is used to represent the difference between the output result of the deep learning model and the actual inertial sensor data.

[0020] By using principal component analysis and clustering algorithm, the application can effectively map high-dimensional data to multiple local subspaces and dynamically adjust the dimension of the local subspace. This process effectively reduces the computational resource consumption of data through dynamic dimension reduction. Especially in the processing of high-dimensional data, by dynamically adjusting the dimension of the subspace, the computational bottleneck problem caused by the high dimension of data in the traditional method is alleviated. Each local subspace of the application only focuses on a specific error type. This approach can improve the processing accuracy of each error feature and avoid interference between different error types, reducing the computational complexity of high-dimensional data; the application effectively alleviates the problem of high-dimensional data by subspace decomposition and dynamic adjustment of the dimension of the subspace, thereby improving the efficiency of the deep learning model inference.

[0021] The application can dynamically adjust the dimension of the subspace according to the real-time training error and historical error, thereby effectively alleviating the overfitting problem of the deep learning model. Specifically, the dynamic adjustment mechanism enables each subspace to automatically adjust the dimension according to the error change during the training process, thereby alleviating the overfitting of the deep learning model to the training data. In the case of large error change, the system increases the subspace dimension to better fit the data characteristics, while in the case of small error, it reduces the subspace dimension to reduce the model complexity and improve the generalization ability of the model. In addition, when using a deep learning model to correct errors, the error feature calibration parameters are combined to further improve the quality of the features, ensuring that the error features input into the model can better match the true error law, effectively alleviating the overfitting phenomenon of using a deep learning model.

[0022] In an optional embodiment, in S3, in each local subspace, the feature calibration parameters corresponding to the error features are constructed as follows: For each local subspace, the original feature vector of the error feature in the local subspace is extracted; For illustration, the error features include bias, noise, and scale factor error. Regarding the original feature vector: The original feature vector extracted by the local subspace corresponding to the bias is the bias feature vector, which includes a sequence vector of zero bias values at each time; The original feature vector extracted by the local subspace corresponding to the noise is the noise feature vector, which includes a statistical vector of noise amplitude and frequency distribution; The original feature vector extracted by the local subspace corresponding to the scale factor error is the scale factor feature vector, which includes a parameter vector of scale factor bias rate and nonlinear coefficient; For further illustration, other error features of the error features are not exhaustively listed here and can be directly extracted using existing technologies; The plurality of original feature vectors generate a set of features to be fused; The original feature vectors in the set of features to be fused are divided into a plurality of fixed-width intervals using equal-width discretization, and the probability of each fixed-width interval is counted; According to the probability of the fixed-width interval, the mutual information value of any two original feature vectors in the set of features to be fused is obtained, the mutual information value is normalized, and a mutual information matrix is generated; Equal-width discretization is an existing discretization method; According to the mutual information matrix, a weighted average algorithm is used to generate a global correlation feature vector; The global correlation feature vector is mapped to a high-level subspace through linear transformation or feature selection algorithm; Linear transformable matrix projection technology The feature selection algorithm can be a mutual information-based dimension screening. In each local subspace, a feature calibration parameter corresponding to an error feature is obtained for the error feature corresponding to the local subspace; and the feature calibration parameter corresponding to the error feature is independently adjusted through a preset rule. The quality of the error feature input into the deep learning model is improved by adjusting the feature calibration parameter corresponding to the error feature. As an illustration: In the local subspace corresponding to the bias, the corresponding feature calibration parameter is a smoothing coefficient of bias estimation. Specifically, the bias feature input into the deep learning model can be smoothed, for example, from a fluctuating sequence of 5 sampling points to a stable sequence of 10 sampling points, so as to help the deep learning model to learn the bias correction rule more accurately. In the local subspace corresponding to the noise, the corresponding feature calibration parameter is a noise filtering window size. Specifically, the noise feature vector can be filtered, for example, from a high-frequency impulse noise of 20 ms to a noise of 50 ms, to reduce the interference components in the noise feature vector, so that the deep learning model can learn the effective noise rule rather than noise interference. In the local subspace corresponding to the scale factor error, the corresponding feature calibration parameter is a nonlinear correction segmentation interval. Specifically, the nonlinear representation accuracy of the scale factor feature vector can be improved, for example, from 10 segments to 20 segments, to provide more delicate error gradient information for the deep learning model, and to optimize the fitting effect of the model on the scale factor error. As a further illustration, the local subspace corresponding to other error features not exhaustively listed, the feature calibration parameter corresponding to the error feature can be quantitatively adjusted according to the adjustment rules disclosed in the prior art. In the high-level subspace, the global correlation feature vector is mined using a correlation rule to generate correlation information data, and the feature calibration parameters corresponding to the error features in each local subspace are adjusted according to the correlation information data through a preset rule.

[0023] Each local subspace of the present application independently adjusts the feature calibration parameter corresponding to the error feature according to the corresponding error feature through a preset rule; in the high-level subspace, the global correlation feature vector is mined using a correlation rule to generate correlation information data, and the feature calibration parameters corresponding to the error features in each local subspace are adjusted according to the correlation information data through a preset rule, the quality of the error feature input into the deep learning model is improved by optimizing the feature input, which is conducive to the stability of the deep learning model and ensures the calibration accuracy.

[0024] In an optional embodiment, S5 further includes: performing error correction on the actual value of the fused data acquired in the future unit time to obtain the final actual value of the fused data as the final calibration result.

[0025] Error correction is applied to the actual values ​​of the fused data acquired at future unit times to obtain the final actual values ​​of the fused data as the final calibration result, as follows: The fused data of historical correction results is obtained, and the predicted value of the fused data for future unit time is generated through a long short-term memory network; When the time reaches the point where the future unit of time is located, obtain the actual value of the fused data for the future unit of time; The residual error is obtained by comparing the actual value of the fused data at a future unit time with the predicted value of the fused data at a future unit time. For clarification, residual error is the difference between the actual value of the fused data per unit time in the future and the predicted value of the fused data per unit time in the future; The Kalman filtering algorithm is adopted, with the residual error as the observation input of the Kalman filter. The Kalman gain is obtained through a preset residual noise model. The residual error is corrected by the Kalman gain. The product of the Kalman gain and the residual error is used as the correction result. The difference between the actual value of the fused data in future unit time and the correction result is used as the final actual value of the fused data. The final actual value of the fused data is used as the final calibration result of the actual value of the fused data in future unit time.

[0026] As an explanation, the final fused data value is a vector; correspondingly, the fused data value, residual error, and correction result per unit time in the future are all vectors with the same dimension. The residual noise model is a Gaussian white noise model; By combining a Long Short-Term Memory (LSTM) network with a Kalman filter algorithm, error correction for future unit-time fused data is achieved. By predicting future data and comparing it with actual values, the residual error is calculated, and then the Kalman filter algorithm is used to accurately correct this error, thus significantly improving calibration accuracy. Especially in complex and variable sensor data environments, this approach not only effectively addresses data prediction errors but also dynamically corrects residuals through Kalman filtering, ensuring more accurate final calibration results. This error correction and real-time feedback effectively improves the stability and robustness of the calibration process, adapting to the error characteristics and data fluctuations of different sensors.

[0027] like Figure 2 The machine learning-based joint calibration system for multiple micro inertial measurement units shown includes: The data acquisition module: acquires raw data of a plurality of inertial sensors, and generates a raw data set; acquires a deep learning model, and the deep learning model is used for predicting errors of the inertial sensors and correcting calibration results of the inertial sensors; The local subspace generation module: decomposes the raw data set into a plurality of local subspaces through a dynamic adaptive subspace adjustment mechanism; The feature calibration parameter generation module: in each local subspace, a feature calibration parameter corresponding to an error feature is constructed; The correction result generation module: inputs the feature calibration parameter corresponding to the error feature in each local subspace into the deep learning model, generates a compensation value of each error feature through the deep learning model, corrects a corresponding error of the inertial sensor, and generates a correction result; The calibration result generation module: performs data fusion on the correction result, and generates fused data as a final calibration result; The future calibration result optimization module: performs error correction on the fused data actual value of a future unit time to obtain a final fused data actual value as a final calibration result.

[0028] Meanwhile, the contents not described in detail in the specification all belong to the prior art known by those skilled in the art.

[0029] In the embodiments provided by the present application, it should be understood that the disclosed system or method can be implemented in other ways. For example, the above-described embodiments are merely illustrative, and the division of modules is merely a logical function division. In actual implementation, another division manner can be used.

[0030] The modules illustrated as separate components can or can not be physically separate, and the components illustrated as modules can or can not be physical modules, and can be located in one place, or can be distributed on multiple network modules. According to actual needs, part or all of the modules can be selected to achieve the purpose of the embodiments.

[0031] In addition, each functional module in each embodiment of the present application can be integrated in one processing module, or each module can exist physically, or two or more modules can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of hardware plus software functional module.

[0032] For those skilled in the art, it is obvious that the present application is not limited to the details of the above exemplary embodiments, and the present application can be realized in other specific forms without departing from the essential characteristics of the present application.

[0033] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A joint calibration method for multiple micro-inertial measurement units based on machine learning, characterized in that, Includes the following steps: S1. Obtain raw data from multiple inertial sensors and generate a raw dataset; obtain a deep learning model, which is used to predict the errors of the inertial sensors and correct the calibration results of the inertial sensors. S2. The original dataset is decomposed into multiple local subspaces through a dynamic adaptive subspace adjustment mechanism; S3. In each local subspace, construct the feature calibration parameters corresponding to the error features; S4. Input the feature calibration parameters corresponding to the error features in each local subspace into the deep learning model, generate the compensation value for each error feature through the deep learning model, correct the corresponding error of the inertial sensor, and generate the correction result. S5. Perform data fusion on the calibration results to generate fused data as the final calibration result.

2. The joint calibration method for multiple micro-inertial measurement components based on machine learning according to claim 1, characterized in that, In S2, the original dataset is decomposed into multiple local subspaces through a dynamic adaptive subspace adjustment mechanism, as follows: Acquire raw data from multiple inertial sensors to generate a raw dataset, perform statistical analysis on the raw dataset to obtain statistical features, and generate a statistical feature set; Statistical characteristics include mean, variance, and covariance; Perform statistical analysis on the statistical feature set to extract the error features from the statistical feature set; Error characteristics include bias, noise, and scaling factor error; Clustering algorithms or principal component analysis techniques are used to map error features to multiple local subspaces, with each local subspace corresponding to a specific error feature.

3. The joint calibration method for multiple micro-inertial measurement components based on machine learning according to claim 1 or 2, characterized in that, S2 also includes dynamically adjusting the dimensions of the local subspace.

4. The joint calibration method for multiple micro-inertial measurement components based on machine learning according to claim 3, characterized in that, The dimensions of the local subspace are dynamically adjusted as follows: For each local subspace, the real-time training error sequence and historical training error sequence corresponding to the deep learning model are obtained; time series analysis is performed on the real-time training error sequence and historical training error sequence to obtain the real-time training error value and error change rate of each local subspace; Preset absolute error threshold and error rate of change threshold for each local subspace; When the real-time training error value of the local subspace is greater than the absolute error threshold, the dimension of the local subspace is increased according to a preset ratio. When the real-time training error value of the local subspace equals the absolute error threshold: If the error rate of change is greater than the error rate of change threshold, the dimension of the local subspace is increased according to a preset ratio; If the rate of change of error is less than or equal to the threshold of the rate of change of error, the dimension of the local subspace remains unchanged; When the real-time training error value of the local subspace is less than the absolute error threshold: If the error rate of change is greater than the error rate of change threshold, the dimension of the local subspace is increased according to a preset ratio; If the rate of change of error equals the threshold of the rate of change of error, the dimension of the local subspace remains unchanged; If the error rate of change is less than the error rate of change threshold, the dimension of the local subspace is reduced according to a preset ratio.

5. The joint calibration method for multiple micro-inertial measurement components based on machine learning according to claim 1 or 2, characterized in that, In S3, the feature calibration parameters corresponding to the constructed error features in each local subspace are as follows: For each local subspace, extract the original feature vector of the error features in the local subspace; multiple original feature vectors generate a feature set to be fused; The original feature vectors in the feature set to be fused are discretized into several fixed-width intervals using equal-width discretization, and the probability of each fixed-width interval is calculated. Based on the probability of a fixed width interval, for any two original feature vectors in the feature set to be fused, obtain the mutual information value of any two original feature vectors, normalize the mutual information value, and generate a mutual information matrix. Based on the mutual information matrix, a weighted average algorithm is used to generate a global correlation feature vector; Map the globally correlated feature vectors to a higher-level subspace using linear transformations or feature selection algorithms; In each local subspace, for the error features corresponding to the local subspace, the feature calibration parameters corresponding to the error features are obtained; and the feature calibration parameters corresponding to the error features are independently adjusted according to preset rules. In the high-level subspace, association rule mining is used on the global association feature vector to generate association information data. Based on the association information data, the feature calibration parameters corresponding to the error features in each local subspace are adjusted according to preset rules.

6. The joint calibration method for multiple micro-inertial measurement components based on machine learning according to claim 1, characterized in that, S5 also includes: performing error correction on the actual value of the fused data acquired in the future unit time, and obtaining the final actual value of the fused data as the final calibration result.

7. The joint calibration method for multiple micro-inertial measurement components based on machine learning according to claim 6, characterized in that, Error correction is applied to the actual values ​​of the fused data acquired at future unit times to obtain the final actual values ​​of the fused data as the final calibration result, as follows: The fused data of historical correction results is obtained, and the predicted value of the fused data for future unit time is generated through a long short-term memory network; When the time reaches the point where the future unit of time is located, obtain the actual value of the fused data for the future unit of time; The residual error is obtained by comparing the actual value of the fused data at a future unit time with the predicted value of the fused data at a future unit time. The Kalman filtering algorithm is adopted, with the residual error as the observation input of the Kalman filter. The Kalman gain is obtained through a preset residual noise model, and the product of the Kalman gain and the residual error is used as the correction result. The difference between the actual value of the fused data and the correction result in a future unit of time will be used as the final actual value of the fused data. The final actual value of the fused data will be used as the final calibration result for the actual value of the fused data per unit time in the future.

8. A joint calibration system for multiple micro-inertial measurement units based on machine learning, used in accordance with the joint calibration method for multiple micro-inertial measurement units based on machine learning as described in any one of claims 1 to 7, characterized in that, include: Data acquisition module: acquires raw data from multiple inertial sensors and generates raw datasets; acquires deep learning models, which are used to predict the errors of inertial sensors and correct the calibration results of inertial sensors; Local subspace generation module: Decomposes the original dataset into multiple local subspaces through a dynamic adaptive subspace adjustment mechanism; Feature calibration parameter generation module: Constructs feature calibration parameters corresponding to error features in each local subspace; The calibration result generation module inputs the feature calibration parameters corresponding to the error features in each local subspace into the deep learning model, generates compensation values ​​for each error feature through the deep learning model, corrects the corresponding errors of the inertial sensor, and generates calibration results. Calibration result generation module: performs data fusion on the calibration results and generates fused data as the final calibration result; Future calibration result optimization module: Corrects the error of the actual value of the fused data acquired in the future unit time, and obtains the final actual value of the fused data as the final calibration result.

Images