A method and platform for aviation gravity vector assessment based on deep learning

Abstract

This invention discloses a deep learning-based method and platform for airborne gravity vector assessment, relating to the field of airborne gravity measurement technology. The method includes: acquiring and preprocessing airborne gravity measurement data through an airborne gravity vector measurement system to generate a standard airborne gravity measurement dataset, including INS and GNSS data. Features related to the gravity vector, such as carrier attitude, velocity, and acceleration, are extracted from this dataset to form a gravity vector feature vector set. A deep learning network is selected based on the data characteristics, and the network is trained to construct a gravity vector feature recognition model. A Kalman filter is used to extract and correct the gravity vector features, thereby obtaining accurate airborne gravity vector assessment results. This invention solves the technical problems of poor accuracy and stability in existing airborne gravity vector measurements. By combining deep learning and Kalman filtering, it achieves high accuracy and high stability even in complex environments.

Description

Technical Field

[0001] The present application relates to the technical field of aerial gravity measurement, and in particular to an aerial gravity vector evaluation method and platform based on deep learning. Background Art

[0002] As an important means of geophysical exploration, aerial gravity vector measurement is widely used in resource exploration, geological structure research and modern national defense construction. However, there are many error sources in the measurement process, such as carrier motion error, instrument noise, GNSS measurement noise and atmospheric disturbance, which will significantly affect the measurement accuracy. Although the traditional Kalman filtering method can suppress errors to a certain extent, its performance is limited in complex environments and cannot meet high-precision requirements. Therefore, how to effectively reduce the impact of error sources and improve measurement accuracy has become a technical problem that needs to be solved urgently. With the advancement of deep learning technology, this scheme combines deep learning and Kalman filtering, uses the feature extraction capability of deep learning networks to optimize the data processing process, and significantly improves the accuracy and stability of aerial gravity vector measurement.

[0003] At present, there are technical problems in the relevant technologies such as poor accuracy and stability in aerial gravity vector measurement. Summary of the invention

[0004] This application provides an aerial gravity vector evaluation method and platform based on deep learning. The aerial gravity measurement data is acquired and preprocessed through an aerial gravity vector measurement system to generate a standard aerial gravity measurement data set, including INS and GNSS data. Features related to the gravity vector, such as carrier attitude, velocity and acceleration, are extracted from the data set to form a gravity vector feature vector set. A deep learning network is selected based on the data characteristics, and the network is trained to construct a gravity vector feature recognition model. The gravity vector features are extracted and corrected using a Kalman filter to obtain accurate aerial gravity vector evaluation results, achieving a technical effect of maintaining high precision and high stability in complex environments by combining deep learning and Kalman filtering.

[0005] This application provides an aviation gravity vector evaluation method based on deep learning, including:

[0006] An aerial gravity measurement data set is acquired through an aerial gravity vector measurement platform, and the aerial gravity measurement data set is preprocessed to obtain a standard aerial gravity measurement data set, wherein the standard aerial gravity measurement data set includes INS data and GNSS data; associated features are extracted from the standard aerial gravity measurement data set to obtain a gravity vector associated feature vector set, wherein the gravity vector associated feature vector set includes carrier attitude, velocity and acceleration; a target deep learning network is selected and designed according to the characteristic information of the gravity vector data, and the target deep learning network is used to train and optimize the standard aerial gravity measurement data set and the gravity vector associated feature vector set to construct a gravity vector feature recognition model; a Kalman filter is initialized to obtain a feature of the current standard gravity measurement data, and the gravity vector feature recognition model is used to extract features from the current standard gravity measurement data to obtain a current gravity vector feature vector; the current gravity vector feature vector is used as an observation input, and the Kalman filter is executed to perform gravity vector estimation correction to obtain an aerial gravity vector evaluation result.

[0007] This application also provides an aviation gravity vector evaluation platform based on deep learning, including:

[0008] A measurement data set acquisition module, the measurement data set acquisition module is used to acquire an aerial gravity measurement data set through an aerial gravity vector measurement system, pre-process the aerial gravity measurement data set, and obtain a standard aerial gravity measurement data set, the standard aerial gravity measurement data set includes INS data and GNSS data; an associated feature extraction module, the associated feature extraction module is used to extract associated features from the standard aerial gravity measurement data set, and obtain a gravity vector associated feature vector set, the gravity vector associated feature vector set includes carrier attitude, velocity and acceleration; a training optimization module, the training optimization module is used to select a design target depth according to the characteristic information of the gravity vector data A learning network, using the target deep learning network to train and optimize the standard aerial gravity measurement data set and the gravity vector associated feature vector set, and construct a gravity vector feature recognition model; a gravity vector feature vector acquisition module, the gravity vector feature vector acquisition module is used to initialize and obtain a Kalman filter, and use the gravity vector feature recognition model to extract features from the current standard gravity measurement data to obtain a current gravity vector feature vector; a vector evaluation result acquisition module, the vector evaluation result acquisition module is used to use the current gravity vector feature vector as an observation input, execute the Kalman filter to perform gravity vector estimation correction, and obtain an aerial gravity vector evaluation result.

[0009] The present application proposes a deep learning-based aerial gravity vector assessment method and platform. First, the aerial gravity measurement data is acquired and preprocessed through the aerial gravity vector measurement system to generate a standard aerial gravity measurement data set, including INS and GNSS data. Features related to the gravity vector, such as carrier attitude, velocity, and acceleration, are extracted from the data set to form a gravity vector feature vector set. A deep learning network is selected based on the data characteristics, and the network is trained to construct a gravity vector feature recognition model. The Kalman filter is used to extract and correct the gravity vector features to obtain accurate aerial gravity vector assessment results. By combining deep learning and Kalman filtering, the technical effect of maintaining high precision and high stability in complex environments is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solution of the embodiment of the present invention, the accompanying drawings of the embodiment of the present invention will be briefly introduced below. A flow chart is used in this application to illustrate the operations performed by the platform according to the embodiment of the present application. It should be understood that the previous or following operations are not necessarily performed precisely in order. On the contrary, various steps can be processed in reverse order or simultaneously as needed. At the same time, other operations can also be added to these processes, or one or more operations can be removed from these processes.

[0011] Figure 1 A schematic diagram of a flow chart of an aerial gravity vector evaluation method based on deep learning provided in an embodiment of the present application;

[0012] Figure 2 A schematic diagram of the structure of an aerial gravity vector assessment platform based on deep learning provided in an embodiment of the present application.

[0013] Explanation of reference numerals: measurement data set acquisition module 10 , association feature extraction module 20 , training optimization module 30 , gravity vector feature vector acquisition module 40 , vector evaluation result acquisition module 50 . DETAILED DESCRIPTION

[0014] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0015] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0016] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict, and the terms "first\second" involved are merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "including" and "having" and any variations are intended to cover non-exclusive inclusions, for example, a process, method, platform, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by technicians in the technical field of this application. The terms used herein are for the purpose of describing the embodiments of the present application only.

[0017] The present application embodiment provides an aerial gravity vector evaluation method based on deep learning, such as Figure 1 As shown, the method includes:

[0018] Step S100, obtain an aviation gravity measurement data set through an aviation gravity vector measurement system, pre-process the aviation gravity measurement data set, and obtain a standard aviation gravity measurement data set, wherein the standard aviation gravity measurement data set includes INS data and GNSS data. Specifically, the aviation gravity vector assessment first obtains an aviation gravity measurement data set through an aviation gravity vector measurement system, and this system uses a variety of sensors to collect data, including accelerometers to sense gravity acceleration, INS and GNSS to obtain aircraft attitude, position and speed information. The data set is then preprocessed. First, abnormal data is identified, data integrity is checked, outliers are detected, and the consistency of INS and GNSS data is analyzed. Then, the preprocessing steps are determined according to the abnormal situation. Then, data cleaning, noise filtering, and INS and GNSS data fusion pre-processing are performed. Finally, a suitable algorithm is used to fuse and transform the data to a unified coordinate system to obtain a standard aerial gravity measurement data set containing INS and GNSS data. Subsequently, associated feature extraction, target deep learning network design and training optimization will be carried out based on this, and a feature recognition model will be constructed. The Kalman filter will be initialized and the feature vector extracted by the model will be input into the filter correction estimate. Finally, the aerial gravity vector evaluation result will be obtained. The results will also be evaluated and fed back to continuously optimize the entire process.

[0019] In a possible implementation, an aerial gravity measurement data set is acquired through an aerial gravity vector measurement system, and the aerial gravity measurement data set is preprocessed to obtain a standard aerial gravity measurement data set, wherein the standard aerial gravity measurement data set includes INS data and GNSS data. Step S100 further includes step S110, in which abnormal data is identified on the aerial gravity measurement data set to obtain abnormal gravity measurement data, wherein the abnormal gravity measurement data includes inconsistent data and data outliers. Specifically, the aerial gravity measurement data set is comprehensively scanned to check the integrity of the data and whether there is a time period with missing data. For example, in a certain flight phase, the data of the accelerometer or other sensors is suddenly interrupted or part of the records are lost. The record timestamps and data formats of the INS data and GNSS data are compared to ensure the consistency of the two in the time series. If it is found that there are obvious differences in the time stamps of the INS data and the GNSS data, or the data formats do not match, these may be manifestations of data inconsistency. They are marked as inconsistent data, and statistical analysis methods are used to detect data outliers. The mean and standard deviation of each measurement data in the entire data set (such as gravity acceleration, aircraft attitude angle, speed, etc.) are calculated. For each data point, it is determined whether it deviates from the mean by more than a certain multiple of the standard deviation, which is usually set to 3 times the standard deviation or a suitable threshold is determined according to the actual data distribution characteristics. For example, in the gravity acceleration data, if a measurement value is far beyond the reasonable range of gravity acceleration under normal flight conditions and is obviously inconsistent with the change trend of the surrounding data points, it is determined as a data outlier. Accurately identify the inconsistent data and data outliers in the data set, and filter out the abnormal gravity measurement data for subsequent targeted processing.

[0020] Step S120, pre-process and analyze the abnormal gravity measurement data to determine the data pre-processing steps. Specifically, analyze the abnormal gravity measurement data in depth, especially the causes of inconsistent data. If the timestamps of INS data and GNSS data are inconsistent, it may be due to the fact that the data collection frequencies of the two are not synchronized, or there is a delay in the data transmission process. For data wild values, check the time period in which they appear and the corresponding flight environment. It is speculated that they may be affected by factors such as strong electromagnetic interference, short-term sensor failure or extreme flight attitude changes. For example, if a large number of data wild values ​​appear when flying over a strong magnetic field area, it is likely that electromagnetic interference causes sensor measurement abnormalities. According to the causes of the abnormal data, determine the corresponding data pre-processing steps. For the inconsistent timestamps of INS and GNSS data, To solve the problem, we plan to use data interpolation or resampling methods to align the two in time. If the data wild value is caused by sensor failure, if the failure duration is short and the amount of data missing is not large, we can consider using adjacent normal data points for interpolation repair; if the failure seriously affects the data quality, we may need to discard the INS or GNSS data of this time period. For the data wild values ​​caused by external factors such as electromagnetic interference, we use filtering algorithms to smooth them, remove abnormal fluctuations, and restore the authenticity of the data. According to the overall quality of the data and subsequent analysis requirements, we determine whether the data needs to be normalized or standardized so that data from different sources and magnitudes can be fused and analyzed on the same scale.

[0021] Step S130, preprocessing the abnormal gravity measurement data based on the data preprocessing step to obtain a usable aviation gravity measurement data set. Specifically, the abnormal gravity measurement data is processed according to the determined data preprocessing steps. For the case of missing data, a suitable interpolation method is selected according to the time series correlation of the data and the continuity of the flight status. For example, a linear interpolation method is used to calculate the interpolation result based on the known data points before and after the missing data point, and fill in the missing position to ensure the integrity of the data sequence. For data wild values, if filtering is used, a suitable filter is selected according to the noise characteristics of the data. For example, for data wild values ​​caused by high-frequency noise, a low-pass filter is used for smoothing to retain the low-frequency trend of the data. During the processing, the data quality is monitored in real time to ensure the data after cleaning and repair. In accordance with the requirements of subsequent analysis, after completing data cleaning and repair, the reliability of the processed abnormal gravity measurement data is verified, the integrity of the data is checked again to ensure that no new problems arise, and the statistical characteristics of the data before and after processing, such as the mean, standard deviation, data distribution, etc., are compared to see whether the data has become more reasonable and stable. At the same time, some key data points are manually checked or compared with other reliable data sources to ensure the accuracy and reliability of the data. After verification, a usable aeronautical gravity measurement data set is obtained, which has significantly improved in data quality and provides a reliable data basis for subsequent operations such as INS and GNSS data fusion.

[0022] Step S140, time synchronization and coordinate conversion fusion are performed on the INS data and GNSS data in the available aerial gravity measurement data set to obtain the standard aerial gravity measurement data set. Specifically, time synchronization operation is performed on the INS data and GNSS data in the available aerial gravity measurement data set. Since there may be differences in the acquisition frequency and time reference of INS data and GNSS data, it is necessary to align the two on the time axis through an accurate time synchronization algorithm. For example, based on the time of GNSS data, according to the acquisition frequency and time mark of INS data, the time of INS data is adjusted to be consistent with that of GNSS data by interpolation or extrapolation. In this process, the dynamic change characteristics of the data are fully considered to ensure that the INS data after time synchronization can accurately reflect the flight status and gravity measurement information at the corresponding moment, and coordinate conversion is performed to unify the INS and GNSS data into the same coordinate system. INS data is usually measured in the carrier coordinate system, while GNSS data is the position and velocity information determined in the earth coordinate system or geographic coordinate system. Through the coordinate conversion matrix, The INS data is converted to the same coordinate system as the GNSS data for fusion calculation. During the coordinate conversion process, the influence of factors such as the curvature and rotation of the earth on the coordinate conversion is considered to ensure the accuracy of the conversion. A suitable data fusion algorithm, such as weighted fusion or Kalman filter fusion algorithm, is used to assign reasonable weights to the INS and GNSS data according to their accuracy, reliability, and performance under different flight conditions. For example, in open areas with good satellite signals, the weight of GNSS data can be relatively high; in areas with strong satellite signal obstruction or interference, INS data has higher reliability and its weight increases accordingly. Through fusion calculation, the advantages of INS and GNSS data are fully utilized, and the information of the two is organically combined to obtain a standard aviation gravity measurement data set. This data set combines the advantages of INS and GNSS data and provides high-quality data support for subsequent gravity vector evaluation and other work.

[0023] Step S200, extracting the associated features of the standard aerial gravity measurement data set to obtain a gravity vector associated feature vector set, wherein the gravity vector associated feature vector set includes carrier attitude, velocity and acceleration. Specifically, when extracting the associated features of the standard aerial gravity measurement data set, the original measurement data is obtained from the INS data, the carrier pitch angle and roll angle are calculated by using the accelerometer in combination with the earth's gravity field model and a specific algorithm, the heading angle is obtained based on the angular velocity integration measured by the gyroscope and the drift error is corrected, and the carrier attitude feature vector is combined; for the velocity feature, the initial velocity estimate is first obtained by integrating the INS data, but due to its drift, the GNSS high-precision velocity data needs to be integrated, and the Kalman filter is used to dynamically weight the fusion according to the error characteristics and measurement conditions; the acceleration feature is first obtained from the INS data, and the high-frequency noise is removed by filtering, and then it is converted to a suitable coordinate system through the coordinate transformation matrix according to the carrier attitude information. The processed acceleration feature, the carrier attitude, and the velocity feature together constitute the gravity vector associated feature vector set.

[0024] Step S300, select and design a target deep learning network according to the characteristic information of the gravity vector data, use the target deep learning network to train and optimize the standard aerial gravity measurement data set and the gravity vector associated feature vector set, and build a gravity vector feature recognition model. Specifically, when building a deep learning network based on gravity vector data, first deeply analyze the characteristics of the gravity vector data, including spatiotemporal distribution, noise and correlation, and determine the appropriate network structure accordingly, such as selecting a convolutional neural network when the data has strong spatial correlation, using a recurrent neural network or a long short-term memory network when there are time series features, and considering a Transformer structure or a hybrid structure when the spatiotemporal features are complex. Then, according to the data complexity and processing requirements, the network layers and the number of neurons are designed, the depth is weighed to prevent overfitting, the number of neurons in the input, intermediate and output layers is determined, and a reasonable learning rate is set to control the training speed and convergence, and the regularization parameter is selected to prevent overfitting. The standard aerial gravity measurement data set and the gravity vector associated feature vector set are preprocessed and divided into training, verification and test sets in proportion. The target deep learning network is used for training, the error and gradient are calculated and the parameters are updated. The optimization algorithm is used for acceleration, the verification set is used for regular evaluation, and the optimization is adjusted according to the situation. After multiple iterations of training and optimization, the test set is used to evaluate the generalization ability, and finally a gravity vector feature recognition model is constructed.

[0025] In one possible implementation, a target deep learning network is selected and designed according to the characteristic information of the gravity vector data, and the standard aerial gravity measurement data set and the gravity vector associated feature vector set are trained and optimized using the target deep learning network to construct a gravity vector feature recognition model. Step S300 further includes step S310, performing multi-dimensional feature extraction on the characteristic information of the gravity vector data to determine data noise characteristics, spatial correlation, and temporal correlation.Specifically, with regard to the characteristic information of gravity vector data, we first focus on determining the characteristics of data noise. Through statistical analysis of a large amount of historical gravity vector measurement data, we study the distribution law of noise. For example, we calculate the statistical quantities such as the mean, variance, skewness and kurtosis of the noise, and determine whether the noise conforms to common distribution models, such as Gaussian distribution and salt and pepper noise distribution. We observe the changes of noise under different measurement conditions (such as different flight altitudes, different geographical regions, different meteorological conditions, etc.), analyze whether it has time-varying characteristics, and study the relationship between noise and measurement equipment parameters (such as sensor accuracy, sampling frequency, etc.) to determine the source and impact of noise. For example, if the noise is found to increase significantly at certain flight altitudes or areas, it may be related to the local electromagnetic environment or atmospheric turbulence, and then targeted measures can be taken, such as adjusting the anti-interference ability of the measurement equipment or optimizing the data collection strategy, deeply exploring the spatial correlation of gravity vector data, analyzing the changing trend of gravity vectors in different geographical locations, using spatial interpolation methods (such as Kriging interpolation, inverse distance weighted interpolation, etc.) to spatially interpolate gravity vector data, observing the difference between the interpolation results and the actual measured data, so as to evaluate the spatial continuity and correlation of the data, and studying the relationship between gravity vectors and geographical features (such as mountains). , ocean, plains, etc.), determine the spatial distribution pattern of gravity vector under different terrain and landforms. For example, in mountainous areas, the gravity vector may show a large change gradient due to the undulating terrain, and there is a strong correlation between adjacent areas; while in plain areas, the gravity vector is relatively stable and the spatial correlation is weak. Through analysis, the spatial correlation characteristics of gravity vector data can be accurately grasped to provide a basis for the subsequent network structure selection, and the change law of gravity vector data over time, that is, time correlation, is studied. The time series data of gravity vector in a flight mission is analyzed, the autocorrelation function is calculated, and the gravity vector under different lag times is observed. The correlation between quantities is studied. For example, in different flight phases such as takeoff, level flight, and landing, the gravity vector may show different change trends and correlations. In the takeoff phase, due to the large acceleration and attitude changes of the aircraft, the gravity vector changes dramatically and the correlation is low; in the level flight phase, the gravity vector is relatively stable and the time correlation is strong. At the same time, the seasonal and periodic change laws of the gravity vector in the long-term measurement process, as well as the relationship between the gravity vector and other time-related factors (such as the rotation of the earth, tidal changes, etc.) are studied to fully understand the time correlation of the gravity vector data in order to select a deep learning network structure suitable for processing time series data.

[0026] Step S320, select a deep learning network structure according to the data noise characteristics, spatial correlation and temporal correlation. Specifically, the deep learning network structure is selected according to the determined data noise characteristics, spatial correlation and temporal correlation. If the data noise presents a Gaussian distribution and the intensity is relatively small, it has little effect on the overall data. When selecting the network structure, relatively more attention can be paid to the ability to extract spatiotemporal features of the data. For data with strong spatial correlation, convolutional neural network (CNN) is an ideal choice. The convolution layer of CNN can effectively extract local spatial features in the data through convolution kernels. The pooling layer can reduce the data dimension and reduce the amount of calculation while retaining important spatial information. For example, when processing gravity vector data related to geographical areas, CNN can automatically learn the spatial pattern characteristics of gravity vectors under different terrains. If the data has obvious time series characteristics, such as the continuous change of gravity vectors with flight time, recurrent neural network (RNN) or long short-term memory network (LSTM) is more suitable. RNN can handle long-term dependencies in sequence data, but there is a problem of gradient disappearance or explosion. LSTM effectively solves this problem through a special gating structure and can better capture the changing trend and correlation of gravity vectors in the time dimension. If the spatiotemporal characteristics of the data are relatively complex, such as flight measurement data under complex terrain and changeable meteorological conditions, the Transformer structure may be the best choice. The self-attention mechanism of Transformer can simultaneously focus on different spatial locations and time points in the data, fully explore the spatiotemporal correlation information in gravity vector data, and effectively extract complex feature representations.

[0027] Step S330, the processing requirements of the characteristic information of the gravity vector data are analyzed to obtain the processing requirements of the gravity vector data. Specifically, a comprehensive processing requirements analysis is performed on the characteristic information of the gravity vector data, and the requirements of gravity vector data in practical applications are considered. For example, in resource exploration, high-precision gravity vector anomaly identification is required to accurately determine the location of potential resources; in geological structure research, not only high precision is required, but also a clear analysis capability of gravity vector changes under different geological structures is required; in modern national defense construction, more emphasis may be placed on real-time and stability to ensure that gravity vector information can be obtained quickly and accurately in complex environments. According to the requirements of different application scenarios, the processing requirements of gravity vector data are determined. For example, for applications with high precision requirements, the processing goal may be to control the gravity vector measurement error within a very small range and improve the resolution and accuracy of the data; for scenarios with high real-time requirements, the processing goal is to shorten the data processing time as much as possible while ensuring a certain accuracy, to achieve fast gravity vector estimation and feedback, and to consider the compatibility and scalability of the system so as to adapt to the changes in different measurement equipment and application requirements in the future.

[0028] Step S340, based on the gravity vector data processing demand target, the parameters of the deep learning network structure are designed to obtain the target deep learning network. Specifically, based on the gravity vector data processing demand target, the parameters of the selected deep learning network structure are designed. First, the complexity of the network is determined according to the accuracy requirements in the processing target. If high-precision gravity vector feature recognition is required, it may be necessary to increase the number of network layers and neurons to improve the expressive power of the model. For example, for high-precision gravity vector anomaly recognition tasks in resource exploration applications, the number of convolutional layers or fully connected layers can be increased so that the network can learn more subtle changes in gravity vector features. Considering real-time requirements, the network parameters are reasonably adjusted to optimize computing efficiency. If the real-time requirements are high, the number of network layers can be appropriately reduced or a more efficient computing structure can be adopted. For example, using 1D convolution instead of 2D convolution can reduce the amount of calculation in some cases without significantly affecting the model performance. According to the stability requirements in the processing target, set appropriate regularization parameters. If the system needs to run stably in a complex environment, prevent The occurrence of overfitting phenomenon increases the weight of the regularization term, such as using L2 regularization and appropriately increasing the value of the regularization parameter to make the model more robust. It is necessary to select appropriate activation functions, optimization algorithms and other parameters according to the characteristics of the data and processing requirements. For example, for gravity vector data with nonlinear characteristics, selecting activation functions such as ReLU or LeakyReLU can enhance the nonlinear expression ability of the network; selecting optimization algorithms such as Adam or Adagrad can accelerate the convergence speed of the network and improve the training efficiency. Through parameter design, a target deep learning network that meets the requirements of gravity vector data processing is obtained, so that it can effectively process gravity vector data in practical applications and realize accurate, fast and stable gravity vector feature recognition.

[0029] In a possible implementation, a target deep learning network is selected and designed according to the characteristic information of the gravity vector data, and the target deep learning network is used to train and optimize the standard aerial gravity measurement data set and the gravity vector associated feature vector set to construct a gravity vector feature recognition model. Step S300 further includes step S350, using the standard aerial gravity measurement data set and the gravity vector associated feature vector set as a gravity vector sample data set. Specifically, first, the standard aerial gravity measurement data set and the gravity vector associated feature vector set are integrated and used together as a gravity vector sample data set to ensure that the information of the two data sources accurately corresponds. For example, each measurement point in the standard aerial gravity measurement data set is accurately matched with the corresponding gravity vector associated feature vector (carrier attitude, velocity and acceleration, etc.). For each sample in the data set, it is labeled according to its corresponding true value of the gravity vector or a known classification label (if it is a classification task). For example, if it is known that the gravity vector characteristics of certain areas belong to a specific geological structure type, the corresponding category labels are labeled for these samples to provide an accurate reference basis for subsequent training and verification.

[0030] Step S360, divide the gravity vector sample data set into data training set, data verification set and data test set according to a certain ratio. Specifically, divide the gravity vector sample data set into data training set, data verification set and data test set according to a certain ratio. The common division ratio is 70% of the data as the training set, 20% of the data as the verification set, and 10% of the data as the test set. However, the specific ratio can be adjusted according to the scale and characteristics of the data set. In the division process, a random sampling method is adopted to ensure that each subset can better represent the characteristic distribution of the entire data set. For example, for a sample data set containing a large number of measurement data of different flight conditions and geographical areas, the data order is randomly disrupted, and then samples are extracted according to a predetermined ratio to form a training set, a verification set and a test set respectively, so as to avoid deviations in model training and evaluation due to uneven data distribution. The division method can make the training set used for model learning and parameter optimization, the verification set used to monitor the performance of the model and adjust the hyperparameters during the training process, and the test set used to finally evaluate the generalization ability and accuracy of the model.

[0031] Step S370: Use the target deep learning network to perform feature recognition training on the data training set to generate a basic vector feature recognition model. Specifically, the target deep learning network is used to perform feature recognition training on the data training set, and the samples in the training set are input into the target deep learning network one by one, and the prediction results are obtained through the forward propagation calculation of the network. For example, for a target deep learning network based on a convolutional neural network (CNN) and a long short-term memory network (LSTM), the standard aerial gravity measurement data and gravity vector associated feature vectors in the training samples are first extracted through the CNN layer to extract spatial features, and then the LSTM layer processes the time series features. Finally, the prediction of the gravity vector features is obtained through the fully connected layer. According to the difference between the prediction result and the true value of the sample annotation, the loss function is calculated (such as the mean square error loss function for regression tasks and the cross entropy loss function for classification tasks), and the gradient of the loss function for each parameter in the network is calculated using the back propagation algorithm. According to the gradient, the weight and bias of the network are adjusted to make the model gradually fit the training data. During the training process, multiple training rounds are usually set, and each round traverses the entire training set once. As the number of training rounds increases, the model continuously learns the feature patterns in the data, the loss function value gradually decreases, and the prediction ability of the model gradually improves, thereby generating a basic vector feature recognition model.

[0032] Step S380, use the data validation set and the data test set to verify and optimize the basic vector feature recognition model, and construct the gravity vector feature recognition model. Specifically, use the data validation set to verify the basic vector feature recognition model, input the samples in the validation set into the basic model, calculate the performance indicators of the model on the validation set, such as accuracy (for classification tasks), root mean square error (for regression tasks), etc., observe the changing trend of the performance indicators, if it is found that the performance of the model on the validation set no longer improves or begins to decline, it means that overfitting may have occurred, adjust the hyperparameters of the model according to the verification results, such as learning rate, regularization parameter, number of network layers or number of neurons, for example, if overfitting is found, you can appropriately reduce the learning rate, increase the weight of the regularization term, or reduce the number of network layers, by continuously adjusting the hyperparameters and evaluating the model performance on the validation set, optimize the parameter settings of the model, improve the generalization ability of the model, after completing the optimization based on the validation set, use the data test set to optimize the optimization. The final evaluation of the optimized model is carried out. The samples in the test set are input into the model, and the performance indicators of the model on the test set are calculated. If the performance of the model on the test set meets the expected requirements, it means that the model has good generalization ability and can be used for actual gravity vector feature recognition tasks. At this time, the model is determined to be a gravity vector feature recognition model. If the performance of the model on the test set is not ideal, it may be necessary to further analyze the reasons, such as whether there are problems with the data set, whether the model structure needs to be improved, etc., and then re-train, verify and optimize the process until a gravity vector feature recognition model that meets the requirements is constructed. Through the verification and optimization process, it is ensured that the constructed model not only performs well on the training data, but also can accurately identify gravity vector features on unseen test data, providing reliable technical support for aerial gravity vector measurement and related applications.

[0033] In a possible implementation, the data validation set and the data test set are used to verify and optimize the basic vector feature recognition model to construct the gravity vector feature recognition model. Step S380 further includes step S381, in which the basic vector feature recognition model is tested and verified using the data validation set and the data test set to obtain model performance parameter information. Specifically, first, the basic vector feature recognition model is tested using the data validation set, and the samples in the validation set are input into the basic model one by one. The model will make predictions based on the input standard aerial gravity measurement data and gravity vector-related feature vectors, and output the recognition results of the gravity vector features. For each sample, the prediction result is compared with the known true value (pre-labeled or obtained by other precise measurement methods), and multiple performance indicators are calculated to evaluate the performance of the model on the validation set. For example, for regression tasks, the root mean square error (RMSE) is calculated, which measures the average degree of deviation between the predicted value and the true value. The formula is: Where n is the number of samples in the validation set, yi is the true value of the i-th sample, y i , is the model's prediction value for the Ith sample. At the same time, the mean absolute error (MAE) is calculated, and its formula is It reflects the average absolute value of the prediction error. For classification tasks, the accuracy is calculated, that is, the proportion of correctly classified samples to the total number of samples. The formula is: Among them, TP is the number of true positive examples, TN is the number of true negative examples, FP is the number of false positive examples, and FN is the number of false negative examples. By calculating these performance indicators, we can obtain the model performance parameter information and fully understand the accuracy, stability and error distribution of the model on the validation set.

[0034] Step S382: Select a model optimizer based on the model performance parameter information. Specifically, according to the obtained model performance parameter information, the performance characteristics of the model are analyzed to select a suitable model optimizer. If the model converges slowly in the early stage of training, but can gradually approach better performance in the later stage, it may indicate that the current optimization algorithm is inefficient in finding the optimal solution. For example, if this happens when using a stochastic gradient descent optimizer, you can consider selecting an optimizer with an Adam adaptive learning rate; Adam combines the advantages of the momentum method and the adaptive learning rate, and usually achieves good results in practice. If the model overfits, that is, the performance index of the validation set first increases and then decreases during the training process, you may need to use an optimizer with weight decay, such as adding a weight decay term to the Adam optimizer to prevent overfitting by penalizing the complexity of the model. If the model has a gradient vanishing or exploding problem during training, you may need to select an optimizer that can handle gradient problems, such as an optimizer that uses gradient clipping technology to prevent excessive or small gradients from causing adverse effects on model training. According to the specific problems reflected by the model performance parameters, select the optimizer that best suits the current model from a variety of optimizers to improve the training effect and generalization ability of the model.

[0035] Step S383, using the model optimizer to optimize and update the model parameter information of the basic vector feature recognition model to obtain the gravity vector feature recognition model. Specifically, after selecting the model optimizer, the optimizer is used to optimize and update the model parameter information of the basic vector feature recognition model. The optimizer calculates the update amount of the model parameters according to the set optimization algorithm and learning rate and other parameters. In each iteration of the Adam optimizer, the learning rate of each parameter is adjusted according to the first-order moment estimate (mean) and second-order moment estimate (variance) of the gradient. For each parameter θ in the model, the update formula is: where θ t is the current parameter value, μ is the learning rate, m t , is the first-order moment estimate of the gradient, v t, is the second-order moment estimate of the gradient, and ε is a very small constant used to prevent the denominator from being zero. By repeating this process, in multiple training rounds, the optimizer gradually adjusts the model parameters based on the performance feedback obtained from the validation set test, so that the model can more accurately identify the gravity vector features when processing standard aerial gravity measurement data and gravity vector associated feature vectors. After multiple iterative optimizations, when the performance indicators of the model on the validation set and the test set achieve satisfactory results, or meet the preset stopping conditions (such as reaching the maximum training rounds, the performance indicators no longer improve, etc.), the optimized model is determined to be the gravity vector feature recognition model, which can better adapt to the characteristics of aerial gravity vector measurement data and provide more accurate feature recognition capabilities for subsequent gravity vector estimation correction and other tasks.

[0036] Step S400, initialize and obtain the Kalman filter, use the gravity vector feature recognition model to extract features from the current standard gravity measurement data, and obtain the current gravity vector feature vector. Specifically, first, the state equation of the Kalman filter is established according to the physical model of the aerial gravity vector measurement, taking into account factors such as the carrier motion state and the dynamic change of the gravity field, and at the same time, the observation equation is established according to the characteristics of the gravity vector data and the characteristics of the measurement system, and then the initial state estimation is determined based on prior knowledge or the initial state, and the error covariance matrix is ​​calculated and set according to the error characteristics of the measurement system to complete the initialization. Then, the constructed gravity vector feature recognition model is used to extract features from the current standard gravity measurement data. The model uses internal structures such as convolutional neural networks and long short-term memory networks, first extracts spatial features through the convolution layer, then processes time series features through the long short-term memory layer, and finally generates the current gravity vector feature vector containing key information such as carrier attitude, velocity, acceleration, etc. through comprehensive processing by the fully connected layer, providing input for the subsequent gravity vector estimation correction of the Kalman filter.

[0037] In one possible implementation, a Kalman filter is initialized and obtained, and features of the current standard gravity measurement data are extracted using the gravity vector feature recognition model to obtain a current gravity vector feature vector. Step S400 further includes step S410, in which a state equation and an observation equation are established based on a physical model of the aerial gravity vector measurement and the characteristic information of the gravity vector data. Specifically, the physical model of aerial gravity vector measurement is studied, which covers various physical relationships of the gravity vector in the actual measurement environment. It takes into account that the gravity vector is affected by the motion state of the carrier (such as the aircraft's flight attitude, speed change, acceleration, etc.) and the characteristics of the earth's gravity field itself (such as the spatial distribution unevenness of the gravity field, slight changes over time, etc.). For example, when the aircraft accelerates to rise or descend, the acceleration of the carrier will interfere with the measurement of the gravity vector; and in different geographical regions, due to differences in the distribution of underground materials, the size and direction of the gravity vector will also be different. Based on the above factors, the changes of the gravity vector in different dimensions (such as the three directions in three-dimensional space) are associated with the carrier motion parameters and the gravity field characteristics, and a state equation that describes the evolution of the gravity vector state over time is established. The equation can accurately reflect the inherent laws of the gravity vector in the dynamic measurement process, providing a basis for subsequent filtering calculations. Dynamic model, detailed analysis of the characteristics of gravity vector data and the characteristics of the measurement system to establish the observation equation. The measurement system includes inertial navigation system (INS) and global navigation satellite system (GNSS), each providing different types of measurement data and having different error characteristics. INS can continuously provide information such as the carrier's attitude and speed, but there is a certain drift error; GNSS can accurately obtain position and speed information, but in areas with strong signal shielding or interference, data interruption or increased error may occur. The role of the observation equation is to establish a connection between these measurement data and the true state of the gravity vector, clarify how to estimate the actual state of the gravity vector from the measurement value, and determine the mathematical relationship between the measurement value and the state of the gravity vector through analysis of the measurement data, thereby constructing an observation equation that can accurately reflect the measurement process, providing a basis for the Kalman filter to convert the measurement data into an estimate of the gravity vector state.

[0038] Step S420, setting the initial state estimate and error covariance matrix of the Kalman filter. Specifically, the initial state estimate of the Kalman filter is set, which is a preliminary guess of the initial state of the gravity vector. The guess needs to take into account a variety of factors, such as the known state of the aircraft at the beginning of the measurement (such as the aircraft is in a horizontal attitude when taking off, the initial velocity is zero, etc.), the approximate value range of the gravity vector at a specific geographical location (based on the local gravity field model and terrain characteristics) and the empirical data of similar measurement tasks in the past. For example, if the measurement starts in a relatively flat area with a relatively stable gravity field, and the aircraft takes off smoothly, then the initial state of the gravity vector in the horizontal direction can be estimated to be close to zero, and the initial state in the vertical direction can be estimated to be close to the local gravity acceleration value. Through the initial state estimation, a relatively accurate starting point is provided for the iterative calculation of the Kalman filter, so that it can converge to the true gravity vector state more quickly. According to the error characteristics of the measurement system and the uncertainty of the initial state estimate, the calculation Calculate and set the error covariance matrix. The error covariance matrix reflects the uncertainty of the initial state estimation and the error correlation between different state variables. For INS and GNSS in the measurement system, analyze their respective error sources (such as sensor accuracy, signal transmission error, etc.) and error size distribution. For example, the attitude measurement error of INS may fluctuate within a certain angle range, and the speed measurement error is related to the measurement frequency and sensor stability; the position error of GNSS may be affected by the satellite signal quality and multipath effect. Based on the error characteristics, calculate the error variance of each variable in the initial state estimation, and determine the covariance relationship between different variables, so as to construct the error covariance matrix. The matrix plays an important role in the calculation of the Kalman filter. It determines how to correct the initial state estimate according to the measurement data in the subsequent iteration process, and how to weigh the credibility of different measurement information.

[0039] Step S430, based on the state equation and observation equation, as well as the Kalman filter initial state estimation and error covariance matrix, the Kalman filter is initialized to obtain the Kalman filter. Specifically, based on the established state equation and observation equation, as well as the set Kalman filter initial state estimation and error covariance matrix, the Kalman filter initialization work is completed, and the Kalman filter has the basic framework and initial parameters required for gravity vector estimation and correction. The state equation provides a dynamic evolution model of the gravity vector state, the observation equation establishes the connection between the measurement data and the gravity vector state, the initial state estimation provides the filter with the initial state guess, and the error covariance matrix determines the initial uncertainty and error relationship. By integrating the elements together, the Kalman filter can start to receive real-time measurement data, and iterate according to its internal algorithm to gradually correct the estimation of the gravity vector state, improve the accuracy and stability of the gravity vector measurement, and in the subsequent measurement process, as new measurement data is continuously input, the Kalman filter will continue to run and continuously optimize the estimation result of the gravity vector to adapt to the changing measurement environment and data conditions.

[0040] Step S500, using the current gravity vector feature vector as observation input, executing the Kalman filter to perform gravity vector estimation correction, and obtaining the aviation gravity vector evaluation result. Specifically, the current gravity vector feature vector extracted by the gravity vector feature recognition model is provided as observation input to the initialized Kalman filter, and the filter processes the input according to the internal observation equation in combination with the carrier attitude, speed and other information in the feature vector and the relationship between the measured value and the gravity vector state, and at the same time, according to the state equation, uses the state estimation of the previous moment to predict the gravity vector state at the current moment, and then corrects the estimation in the update step based on the difference between the predicted value and the observed value and the error covariance matrix. After multiple iterations, the aviation gravity vector evaluation result containing accurate components and uncertainty information is obtained, which can be used for related research applications, and can also be compared and verified with other measurements or reference data to optimize and improve the entire system, and improve the accuracy and stability of aviation gravity vector measurement.

[0041] In a possible implementation, the current gravity vector feature vector is used as an observation input, the Kalman filter is executed to perform gravity vector estimation correction, and an aviation gravity vector evaluation result is obtained. Step S500 further includes step S510, in which verification and evaluation are performed on the aviation gravity vector evaluation result to obtain gravity vector evaluation performance parameters, and the gravity vector evaluation performance parameters include accuracy and stability. Specifically, a comprehensive verification and evaluation of the aerial gravity vector evaluation results is carried out to obtain the accuracy indicators in the gravity vector evaluation performance parameters, and the evaluation results are compared with high-precision reference data. The reference data can come from the precise measurement results of ground gravity measurement stations, the calculated values ​​of strictly calibrated physical models, or other recognized high-precision gravity vector data sources. For the gravity vector evaluation results of each measurement point or time period, the difference between it and the reference data in each component (such as the x, y, and z direction components in three-dimensional space) is calculated, and a variety of statistical indicators are used to quantify the accuracy, such as the root mean square error (RMSE), which is calculated by first summing the squares of the errors on each component, dividing it by the number of samples and then taking the square root. The smaller the RMSE value, the higher the accuracy; the mean absolute error (MAE), that is, averaging the absolute values ​​of the errors on each component, which can intuitively reflect the average size of the errors; the standard deviation of the errors can also be calculated to understand the degree of discreteness of the errors. Through the calculation of indicators, the accuracy level of the aerial gravity vector evaluation results in each direction can be accurately evaluated, and its difference with the real gravity vector can be judged. The degree of proximity of the values. When evaluating stability, the fluctuation of the aviation gravity vector evaluation results under different measurement conditions is analyzed. Various factors in the flight process are considered, such as different flight altitudes, speeds, attitude changes, and different geographical regions (mountainous areas, plains, oceans, etc.) and meteorological conditions (sunny, cloudy, strong winds, etc.) on the gravity vector measurement. The entire measurement process is divided into multiple sub-intervals, such as according to the flight stage (take-off, cruising, landing) or geographical area. The statistical characteristics of the gravity vector evaluation results in each sub-interval, such as the mean, variance, etc., are calculated, and the changes in statistical characteristics between different sub-intervals are observed. If the variance is small and the mean is relatively stable, it means that the evaluation results have good stability; on the contrary, if the evaluation results fluctuate greatly under different conditions, such as the variance of the gravity vector evaluation results when flying in mountainous areas is significantly greater than that in plains, or the evaluation results have a large deviation when the flight attitude changes drastically, it indicates that the stability is poor. Understanding the stability performance of the aviation gravity vector evaluation results in various complex environments provides a basis for the formulation of subsequent adjustment strategies.

[0042] Step S520, according to the gravity vector evaluation performance parameters, set the parameter closed-loop feedback adjustment strategy. Specifically, according to the results of the accuracy evaluation, if it is found that the accuracy does not meet the requirements, the possible reasons for the decrease in accuracy are analyzed in depth, and then the corresponding parameter closed-loop feedback adjustment strategy is set. If the accuracy problem mainly comes from the gravity vector feature recognition model, for example, the model does not extract certain features accurately or there is overfitting, it may be necessary to adjust the structure or parameters of the model. For the model structure, you can consider increasing or decreasing the number of network layers, adjusting the convolution kernel size or the number of neurons, etc.; for parameters, adjust the learning rate, regularization parameters, etc. If it is a Kalman filter problem, such as inaccurate state equations or observation equations, it may be necessary to review the physical model assumptions and correct them, or adjust the initial state estimate of the filter and parameters such as the error covariance matrix. For example, if it is found that the state equation or observation equation is inaccurate, it may be necessary to review the physical model assumptions and correct them, or adjust the parameters such as the initial state estimate and error covariance matrix of the filter. The estimation error of the gravity vector in a certain direction is large, and it is related to the unreasonable setting of the state equation parameters in the Kalman filter in this direction. Therefore, the state transfer matrix elements in this direction are adjusted in a targeted manner to improve the estimation accuracy of the gravity vector in this direction. Combined with the stability evaluation results, the parameter closed-loop feedback adjustment strategy is further optimized. If the stability is poor, it may be due to the insufficient adaptability of the model to different measurement conditions, or the Kalman filter is not robust enough when dealing with dynamic changes. For the model, the diversity of training data can be increased, including data from different flight conditions and geographical areas, so that the model can better learn the gravity vector characteristics under various conditions; at the same time, the regularization strategy of the model is adjusted to enhance its generalization ability and prevent large fluctuations under different conditions. For the Kalman filter, its adaptive mechanism is optimized, such as dynamically adjusting the observation noise and process noise covariance matrix according to the changes in the measurement environment, so that the filter can maintain good performance under different conditions. If the stability is significantly reduced under certain specific conditions (such as strong wind interference), it may be necessary to adjust the filter parameters specifically for these situations, or increase the recognition and processing capabilities of strong wind interference features in the feature recognition model, thereby improving the stability of the entire system in complex environments.

[0043] Step S530, based on the parameter closed-loop feedback adjustment strategy, the parameters of the gravity vector feature recognition model and the Kalman filter are adjusted and optimized. Specifically, based on the set parameter closed-loop feedback adjustment strategy, the parameters of the gravity vector feature recognition model are adjusted and optimized. If it is determined that the model structure needs to be adjusted, such as increasing the number of network layers, the convolutional layers or fully connected layers are gradually added while ensuring that the computing resources allow, and the model is retrained. During the training process, the performance of the model on the validation set is closely monitored to ensure that the new structure can improve the accuracy and stability of the model. For parameter adjustment, such as reducing the learning rate, if the previous learning rate is too large and the model convergence is unstable, the value of the learning rate is gradually reduced so that the model can adjust the parameters more finely during the training process to avoid missing the optimal solution. According to the adjustment strategy of the regularization parameters, the weight of the regularization term is increased to prevent the model from overfitting and improve its generalization ability on different data. Through multiple iterative training and parameter adjustment, the gravity vector feature recognition model can better adapt to the characteristics of the aerial gravity vector measurement data and improve its recognition accuracy and stability of the gravity vector features. According to the parameter The closed-loop feedback adjustment strategy is used to optimize the parameters of the Kalman filter. If the state equation or observation equation needs to be corrected, the physical process and data relationship of the aerial gravity vector measurement are reanalyzed, and the parameters or terms in the equation are modified based on the new understanding. For example, if it is found that the influence of a certain factor on the gravity vector measurement has been ignored before, it is incorporated into the equation and the appropriate parameter value is determined. For the adjustment of the initial state estimation and the error covariance matrix, more accurate initial values ​​are recalculated based on the new measurement data and the accuracy and stability evaluation results. During the flight, if it is found that the measurement data in certain areas or flight stages have special statistical characteristics, the parameters of the filter can be adjusted in real time based on these characteristics, such as dynamically adjusting the observation noise covariance matrix to improve the performance of the filter in different situations. By continuously optimizing the parameters of the Kalman filter, it can estimate the aerial gravity vector more accurately and stably when working in conjunction with the gravity vector feature recognition model, thereby achieving continuous improvement in the performance of the entire system.

[0044] The embodiment of the present application uses an aerial gravity vector measurement system to acquire and preprocess aerial gravity measurement data to generate a standard aerial gravity measurement data set, including INS and GNSS data. Features related to the gravity vector, such as carrier attitude, velocity and acceleration, are extracted from the data set to form a gravity vector feature vector set. A deep learning network is selected based on the data characteristics, and the network is trained to construct a gravity vector feature recognition model. The gravity vector features are extracted and corrected using a Kalman filter to obtain an accurate aerial gravity vector evaluation result. By combining deep learning and Kalman filtering, a technical effect of maintaining high precision and high stability in a complex environment is achieved.

[0045] In the above, refer to Figure 1A method for evaluating an aerial gravity vector based on deep learning according to an embodiment of the present invention is described in detail. Figure 2 An aerial gravity vector assessment platform based on deep learning according to an embodiment of the present invention is described.

[0046] According to an embodiment of the present invention, an aerial gravity vector evaluation platform based on deep learning is used to solve the technical problems of poor accuracy and stability in existing aerial gravity vector measurements. By combining deep learning and Kalman filtering, a technical effect of maintaining high accuracy and high stability in complex environments is achieved. An aerial gravity vector evaluation platform based on deep learning includes: a measurement data set acquisition module 10, an associated feature extraction module 20, a training optimization module 30, a gravity vector feature vector acquisition module 40, and a vector evaluation result acquisition module 50.

[0047] The measurement data set acquisition module 10 is used to acquire an aviation gravity measurement data set through an aviation gravity vector measurement system, pre-process the aviation gravity measurement data set, and obtain a standard aviation gravity measurement data set, wherein the standard aviation gravity measurement data set includes INS data and GNSS data.

[0048] The associated feature extraction module 20 is used to extract associated features from the standard aerial gravity measurement data set to obtain a gravity vector associated feature vector set, wherein the gravity vector associated feature vector set includes carrier attitude, velocity and acceleration.

[0049] The training optimization module 30 is used to select and design a target deep learning network according to the characteristic information of the gravity vector data, and use the target deep learning network to train and optimize the standard aerial gravity measurement data set and the gravity vector associated feature vector set to construct a gravity vector feature recognition model.

[0050] The gravity vector feature vector acquisition module 40 is used to initialize and obtain a Kalman filter, and to extract features from the current standard gravity measurement data using the gravity vector feature recognition model to obtain the current gravity vector feature vector.

[0051] The vector evaluation result acquisition module 50 is used to use the current gravity vector feature vector as observation input, execute the Kalman filter to perform gravity vector estimation correction, and obtain the aviation gravity vector evaluation result.

[0052] The specific configuration of the measurement data set acquisition module 10 will be described in detail below. As described above, an aviation gravity measurement data set is acquired through an aviation gravity vector measurement system, and the aviation gravity measurement data set is preprocessed to obtain a standard aviation gravity measurement data set, and the standard aviation gravity measurement data set includes INS data and GNSS data. The measurement data set acquisition module 10 further includes: an abnormal data identification unit, the abnormal data identification unit is used to identify abnormal data of the aviation gravity measurement data set, and obtain abnormal gravity measurement data, and the abnormal gravity measurement data includes inconsistent data and data wild values; a preprocessing analysis unit, the preprocessing analysis unit is used to perform preprocessing analysis on the abnormal gravity measurement data and determine the data preprocessing steps; an aviation gravity measurement data set acquisition unit, the aviation gravity measurement data set acquisition unit is used to preprocess the abnormal gravity measurement data based on the data preprocessing steps to obtain a usable aviation gravity measurement data set; a synchronous fusion unit, the synchronous fusion unit is used to perform time synchronization and coordinate transformation fusion on the INS data and GNSS data in the usable aviation gravity measurement data set to obtain the standard aviation gravity measurement data set.

[0053] The specific configuration of the training optimization module 30 will be described in detail below. As described above, a target deep learning network is selected and designed according to the characteristic information of the gravity vector data, and the target deep learning network is used to train and optimize the standard aerial gravity measurement data set and the gravity vector associated feature vector set to construct a gravity vector feature recognition model. The training optimization module 30 further includes: a multidimensional feature extraction unit, which is used to extract multidimensional features of the gravity vector data characteristic information and determine data noise characteristics, spatial correlation and time correlation; a network structure selection unit, which is used to select a deep learning network structure according to the data noise characteristics, spatial correlation and time correlation; a processing requirement analysis unit, which is used to perform processing requirement analysis on the gravity vector data characteristic information to obtain the gravity vector data processing requirement target; a parameter design unit, which is used to perform parameter design on the deep learning network structure based on the gravity vector data processing requirement target to obtain the target deep learning network.

[0054] Among them, the training optimization module 30 further includes: an associated feature unit, which is used to use the standard aerial gravity measurement data set and the gravity vector associated feature vector set as a gravity vector sample data set; a data ratio division unit, which is used to perform data ratio division on the gravity vector sample data set to obtain a data training set, a data verification set and a data test set; a feature recognition training unit, which is used to use the target deep learning network to perform feature recognition training on the data training set to generate a basic vector feature recognition model; a verification optimization unit, which is used to use the data verification set and the data test set to verify and optimize the basic vector feature recognition model to construct the gravity vector feature recognition model.

[0055] Among them, the basic vector feature recognition model is verified and optimized by using the data verification set and the data test set to construct the gravity vector feature recognition model, and the verification and optimization unit further includes: a model performance parameter information acquisition subunit, the model performance parameter information acquisition subunit is used to test and verify the basic vector feature recognition model by using the data verification set and the data test set to obtain model performance parameter information; a model optimizer selection subunit, the model optimizer selection subunit is used to select a model optimizer according to the model performance parameter information; an information optimization update subunit, the information optimization update subunit is used to optimize and update the model parameter information of the basic vector feature recognition model by using the model optimizer to obtain the gravity vector feature recognition model.

[0056] The specific configuration of the gravity vector feature vector acquisition module 40 will be described in detail below. As described above, the Kalman filter is initialized and obtained, and the gravity vector feature recognition model is used to extract features from the current standard gravity measurement data to obtain the current gravity vector feature vector. The gravity vector feature vector acquisition module 40 further includes: an equation establishment unit, which is used to establish a state equation and an observation equation according to the physical model of the aerial gravity vector measurement and the characteristic information of the gravity vector data; a matrix setting unit, which is used to set the initial state estimation and error covariance matrix of the Kalman filter; and a Kalman filter initialization unit, which is used to initialize and obtain the Kalman filter based on the state equation and observation equation, as well as the initial state estimation and error covariance matrix of the Kalman filter.

[0057] The specific configuration of the vector evaluation result acquisition module 50 will be described in detail below. As described above, the current gravity vector feature vector is used as the observation input, the Kalman filter is executed to perform gravity vector estimation correction, and the aviation gravity vector evaluation result is obtained. The vector evaluation result acquisition module 50 further includes: a gravity vector evaluation performance parameter acquisition unit, the gravity vector evaluation performance parameter acquisition unit is used to verify and evaluate the aviation gravity vector evaluation result, and obtain gravity vector evaluation performance parameters, and the gravity vector evaluation performance parameters include accuracy and stability; an adjustment strategy setting unit, the adjustment strategy setting unit is used to set a parameter closed-loop feedback adjustment strategy according to the gravity vector evaluation performance parameters; a parameter adjustment optimization unit, the parameter adjustment optimization unit is used to adjust and optimize the gravity vector feature recognition model and the Kalman filter based on the parameter closed-loop feedback adjustment strategy.

[0058] An aerial gravity vector assessment platform based on deep learning provided by an embodiment of the present invention can execute an aerial gravity vector assessment method based on deep learning provided by any embodiment of the present invention, and has functional modules and beneficial effects corresponding to the execution method.

[0059] Although the present application makes various references to certain modules in the platform according to the embodiments of the present application, any number of different modules may be used and run on the user terminal and / or server, and the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be implemented; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

[0060] The above specific implementation manner does not constitute a limitation to the protection scope of the present application. It should be understood by those skilled in the art that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application. In some cases, the actions or steps recorded in the present application can be performed in an order different from that in the embodiment and can still achieve the desired results. In addition, the process depicted in the accompanying drawings does not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for evaluating aerial gravity vector based on deep learning, characterized in that: The method comprises: Acquiring an aerial gravity measurement data set through an aerial gravity vector measurement system, preprocessing the aerial gravity measurement data set to obtain a standard aerial gravity measurement data set, wherein the standard aerial gravity measurement data set includes INS data and GNSS data; Extracting associated features from the standard aerial gravity measurement data set to obtain a gravity vector associated feature vector set, wherein the gravity vector associated feature vector set includes carrier attitude, velocity and acceleration; Selecting and designing a target deep learning network according to the characteristic information of the gravity vector data, using the target deep learning network to train and optimize the standard aerial gravity measurement data set and the gravity vector associated feature vector set, and constructing a gravity vector feature recognition model; Initialize and obtain a Kalman filter, use the gravity vector feature recognition model to extract features from current standard gravity measurement data, and obtain a current gravity vector feature vector; Taking the current gravity vector feature vector as observation input, executing the Kalman filter to perform gravity vector estimation correction, and obtaining an aviation gravity vector evaluation result; Wherein, the design target deep learning network includes: Performing multi-dimensional feature extraction on the characteristic information of the gravity vector data to determine data noise characteristics, spatial correlation and temporal correlation; Selecting a deep learning network structure according to the data noise characteristics, spatial correlation and temporal correlation; Performing a processing requirement analysis on the characteristic information of the gravity vector data to obtain a gravity vector data processing requirement target; The parameters of the deep learning network structure are designed based on the gravity vector data processing requirement target to obtain the target deep learning network.

2. The method for evaluating aerial gravity vector based on deep learning according to claim 1, characterized in that: The method of obtaining a standard aerial gravity measurement data set includes: Performing abnormal data identification on the aerial gravity measurement data set to obtain abnormal gravity measurement data, wherein the abnormal gravity measurement data includes inconsistent data and data outliers; Performing preprocessing analysis on the abnormal gravity measurement data to determine data preprocessing steps; Preprocessing the abnormal gravity measurement data based on the data preprocessing step to obtain a usable aviation gravity measurement data set; The INS data and GNSS data in the available aerospace gravity measurement data set are time synchronized and coordinate transformed and fused to obtain the standard aerospace gravity measurement data set.

3. The method for evaluating aerial gravity vector based on deep learning according to claim 1, characterized in that: The construction of the gravity vector feature recognition model includes: Using the standard aerial gravity measurement data set and the gravity vector associated feature vector set as a gravity vector sample data set; Dividing the gravity vector sample data set into data proportions to obtain a data training set, a data verification set, and a data test set; Using the target deep learning network to perform feature recognition training on the data training set to generate a basic vector feature recognition model; The data verification set and the data test set are used to verify and optimize the basic vector feature recognition model to construct the gravity vector feature recognition model.

4. The method for evaluating aerial gravity vector based on deep learning according to claim 3, characterized in that: The step of constructing the gravity vector feature recognition model includes: Using the data validation set and the data test set to test and verify the basic vector feature recognition model to obtain model performance parameter information; Selecting a model optimizer according to the model performance parameter information; The model optimizer is used to optimize and update the model parameter information of the basic vector feature recognition model to obtain the gravity vector feature recognition model.

5. The method for evaluating aerial gravity vector based on deep learning according to claim 1, characterized in that: The initialization obtains the Kalman filter, comprising: Establishing a state equation and an observation equation according to a physical model of aerial gravity vector measurement and characteristic information of the gravity vector data; Set the Kalman filter initial state estimate and error covariance matrix; The Kalman filter is initialized based on the state equation and the observation equation, as well as the Kalman filter initial state estimation and the error covariance matrix.

6. The method for evaluating aerial gravity vector based on deep learning according to claim 1, characterized in that: The method comprises: Verifying and evaluating the aerial gravity vector evaluation result to obtain gravity vector evaluation performance parameters, wherein the gravity vector evaluation performance parameters include accuracy and stability; Evaluate performance parameters based on the gravity vector and set a parameter closed-loop feedback adjustment strategy; The parameters of the gravity vector feature recognition model and the Kalman filter are adjusted and optimized based on the parameter closed-loop feedback adjustment strategy.

7. A deep learning-based aviation gravity vector assessment platform, characterized in that: The platform is used to implement the deep learning-based aviation gravity vector assessment method according to any one of claims 1 to 6, and the platform comprises: A measurement data set acquisition module, the measurement data set acquisition module is used to acquire an aviation gravity measurement data set through an aviation gravity vector measurement system, pre-process the aviation gravity measurement data set, and obtain a standard aviation gravity measurement data set, wherein the standard aviation gravity measurement data set includes INS data and GNSS data; A correlation feature extraction module, wherein the correlation feature extraction module is used to extract correlation features from the standard aerial gravity measurement data set to obtain a gravity vector correlation feature vector set, wherein the gravity vector correlation feature vector set includes a carrier attitude, a velocity, and an acceleration; A training optimization module, wherein the training optimization module is used to select and design a target deep learning network according to the characteristic information of the gravity vector data, and to train and optimize the standard aerial gravity measurement data set and the gravity vector associated feature vector set using the target deep learning network to construct a gravity vector feature recognition model; A gravity vector feature vector acquisition module, wherein the gravity vector feature vector acquisition module is used to initialize and obtain a Kalman filter, and to extract features from current standard gravity measurement data using the gravity vector feature recognition model to obtain a current gravity vector feature vector; The vector evaluation result acquisition module is used to use the current gravity vector feature vector as observation input, execute the Kalman filter to perform gravity vector estimation correction, and obtain the aviation gravity vector evaluation result.

Images