BiLSTM-CNN-based noise reduction reconstruction method and system for micro-angle measurement time data

Abstract

The invention provides a BiLSTM-CNN-based noise reduction reconstruction method and system for micro-angle measurement time data. The method comprises the steps of obtaining micro-angle measurement time data of a visual axis pointing change micro-angle measurement system; analyzing data change characteristics and noise distribution rules of the micro-angle measurement time data; generating a micro-angle measurement data set according to the data change characteristics and the noise distribution rule; training a BiLSTM-CNN model used for realizing micro-angle measurement time sequence denoising based on the micro-angle measurement data set, and searching an optimal hyper-parameter combination of the BiLSTM-CNN model by using a Bayesian search algorithm to obtain a micro-angle measurement time sequence denoising model; the micro-angle measurement time sequence denoising model is adopted to carry out noise reduction reconstruction on the micro-angle measurement time data to be processed, so that the noise influence of the micro-angle measurement time data is reduced, and the precision of a micro-angle measurement system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing of a micro-angle measurement system for changes in visual axis pointing, and in particular to a denoising and reconstruction method and system for micro-angle measurement time data based on BiLSTM-CNN. Background Art

[0002] The change of line of sight can be measured and fed back with high precision using a goniometer based on the principle of optical autocollimation. The optical autocollimation angle measurement system generally measures the change of line of sight angle by calculating the position change of the reflected light spot on the image sensor. In actual engineering applications, the micro-angle measurement system may cause errors and noise in optical and electronic components due to satellite on-orbit environments or special harsh external environments. Among them, the thermal noise and shot noise generated by the random operation of the resistors in the image sensor acquisition circuit will cause deviations when solving the center of mass of the light spot. The deviations in these center of mass solutions will cause deviations in the calculation results of the line of sight change, reducing the reliability and accuracy of the micro-angle measurement system. When deploying a high-precision center of mass solution algorithm, it is necessary to consider the solution accuracy, resource occupancy and real-time performance of the algorithm at the same time. Therefore, it is a feasible and effective strategy to develop a method that can reduce the influence of noise from the micro-angle measurement time series of the line of sight change received in the end. Summary of the invention

[0003] In order to solve the above technical problems or at least partially solve the above technical problems, the present invention provides a denoising and reconstruction method and system for micro-angle measurement time data based on BiLSTM-CNN.

[0004] One aspect of the present invention provides a denoising and reconstruction method for micro-angle measurement time data based on BiLSTM-CNN, the method comprising:

[0005] Acquire micro-angle measurement time data of the micro-angle measurement system of the line of sight pointing change;

[0006] Analyze the data variation characteristics and noise distribution law of micro-angle measurement time data;

[0007] A standard micro-angle measurement data set is constructed according to the data variation characteristics of the micro-angle measurement time data, and according to the noise distribution law of the micro-angle measurement time data, noise data conforming to the noise distribution law is added to each standard micro-angle measurement time data in the standard micro-angle measurement data set to form a micro-angle measurement data set;

[0008] Based on the micro-angle measurement data set, a BiLSTM-CNN model for realizing micro-angle measurement time series denoising is trained, and a Bayesian search algorithm is used to find the optimal hyperparameter combination of the BiLSTM-CNN model to obtain a micro-angle measurement time series denoising model;

[0009] The micro-angle measurement time series denoising model is used to perform denoising and reconstruction on the micro-angle measurement time data to be processed to output the denoised micro-angle measurement time data.

[0010] Wherein, obtaining micro-angle measurement time data of the micro-angle measurement system of the line of sight pointing change includes:

[0011] Receive micro-angle measurement time data sent to the ground station in binary format by a micro-angle measurement system for line of sight pointing changes deployed on the satellite;

[0012] Decode the micro-angle measurement time data in binary format and convert it into decimal data including positive and negative signs.

[0013] The analysis of the data variation characteristics and noise distribution law of the micro-angle measurement time data includes:

[0014] The annual variation law analysis of the micro-angle measurement time data is carried out in the year, the daily variation law analysis is carried out in the day, and the time variation law analysis is carried out in the hour to obtain the variation law of the micro-angle measurement time series over time;

[0015] A noise distribution analysis is performed on the micro-angle measurement time data to obtain the noise distribution law of the micro-angle measurement time series.

[0016] Wherein, performing noise distribution analysis on the micro-angle measurement time data comprises:

[0017] The Monte Carlo simulation method is used to analyze the distribution law of the centroid solution deviation caused by image sensor noise, and the obtained distribution law of the centroid solution deviation is used as the noise distribution law of the micro-angle measurement time series.

[0018] The micro-angle measurement time data includes a first micro-angle measurement time series and a second micro-angle measurement time series output by different measurement circuits of the micro-angle measurement system for changing the visual axis pointing;

[0019] The variation pattern of each micro-angle measurement time series over time is: periodic change in years, periodic change in days and periodic change in hours;

[0020] The noise distribution law of each micro-angle measurement time series conforms to the normal distribution law.

[0021] The micro-angle measurement time series denoising model includes a first BiLSTM layer, a second BiLSTM layer arranged in parallel with the first BiLSTM layer, and a CNN layer connected to the output ends of the first BiLSTM layer and the second BiLSTM layer;

[0022] The first BiLSTM layer and the second BiLSTM layer both use the Tanh activation function, and the first BiLSTM layer and the second BiLSTM layer are respectively used to extract features of two micro-angle measurement time series samples included in each micro-angle measurement time data sample in the micro-angle measurement data set, so as to obtain long-term dependency features in each micro-angle measurement time series sample; each micro-angle measurement time data sample includes a first micro-angle measurement time series sample and a second micro-angle measurement time series sample representing different measurement circuit outputs of the micro-angle measurement system with a change in the line of sight pointing;

[0023] The CNN layer is composed of a first convolutional layer, a first activation function layer, a first dropout layer and a second convolutional layer. The CNN network uses a ReLU activation function. The CNN layer is used to extract and fuse features of the output results of the first BiLSTM layer and the second BiLSTM layer, and reconstruct the micro-angle measurement time series according to the fused features to obtain denoised micro-angle measurement time data.

[0024] Wherein, the method further comprises:

[0025] Constructing a loss function of the micro-angle measurement time series denoising model, the loss function includes a first loss function and a second loss function, the first loss function is a deviation loss loss1 between the micro-angle measurement time series after model denoising and the original micro-angle time series corresponding to the first micro-angle measurement time series sample, and the second loss function is a deviation loss loss2 between the micro-angle measurement time series after model denoising and the original micro-angle time series corresponding to the second micro-angle measurement time series sample;

[0026] The dynamic gradient balancing algorithm is used to dynamically weigh the gradients of the first loss function and the second loss function according to the preset error standard level. The specific formula is as follows:

[0027]

[0028] loss=weight·loss1+(1-weight)·loss2 (2)

[0029] Where: grad1 and grad2 are the average values ​​of the gradient norms generated by the first loss function and the second loss function, respectively; loss is the total loss of the micro-angle measurement time series denoising model; and weight is the weight coefficient.

[0030] Among them, the hyperparameters of the BiLSTM-CNN model include: the number of BiLSTM layers and the number of hidden layer nodes of the BiLSTM layer, the coefficient of the Dropout layer, the batch size, the learning rate, the convolution kernel size and the weight decay parameter.

[0031] Another aspect of the present invention further provides a denoising and reconstruction system for micro-angle measurement time data based on BiLSTM-CNN, the system comprising:

[0032] An acquisition module, used for acquiring micro-angle measurement time data of a micro-angle measurement system for changing the visual axis pointing direction;

[0033] An analysis module is used to analyze the data variation characteristics and noise distribution law of micro-angle measurement time data;

[0034] A data generation module is used to construct a standard micro-angle measurement data set according to the data change characteristics of the micro-angle measurement time data, and add noise data that conforms to the noise distribution law to each standard micro-angle measurement time data in the standard micro-angle measurement data set according to the noise distribution law of the micro-angle measurement time data to form a micro-angle measurement data set;

[0035] A model building module is used to train a BiLSTM-CNN model for realizing micro-angle measurement time series denoising based on the micro-angle measurement data set, and use a Bayesian search algorithm to find the optimal hyperparameter combination of the BiLSTM-CNN model to obtain a micro-angle measurement time series denoising model;

[0036] The reconstruction module is used to use the micro-angle measurement time series denoising model to perform denoising and reconstruction on the micro-angle measurement time data to be processed to output the denoised micro-angle measurement time data.

[0037] Wherein, the analysis module includes:

[0038] The first analysis unit is used to perform annual variation analysis based on years, daily variation analysis based on days, and time variation analysis based on hours on the micro-angle measurement time data, so as to obtain the variation pattern of the micro-angle measurement time series over time;

[0039] The second analysis unit is used to perform noise distribution analysis on the micro-angle measurement time data to obtain the noise distribution law of the micro-angle measurement time series.

[0040] The denoising and reconstruction method and system for micro-angle measurement time data based on BiLSTM-CNN provided by the present invention integrate BiLSTM and CNN, firstly use BiLSTM to capture the long-term dependency between micro-angle measurement time series, and then use CNN to further extract features, thereby being able to encode the contextual information of the micro-angle measurement time series to ensure that each data point can utilize its contextual information, thus providing CNN with more comprehensive time series features. The present invention can more accurately reconstruct the micro-angle measurement time series, effectively reduce the influence of noise, and improve the accuracy of the micro-angle measurement system. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0042] Figure 1 A flowchart of a method for denoising and reconstructing micro-angle measurement time data based on BiLSTM-CNN provided in an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the structure of the angle measurement system according to an embodiment of the present invention;

[0044] Figure 3 A schematic diagram of the deviation distribution result obtained by solving the centroid of the light spot during the Monte Carlo simulation process proposed in an embodiment of the present invention;

[0045] Figure 4 A specific structural principle diagram of the BiLSTM-CNN model provided in an embodiment of the present invention;

[0046] Figure 5 A Bayesian search optimization flow chart proposed in an embodiment of the present invention;

[0047] Figure 6 A schematic diagram showing the comparison of denoising and reconstruction results of micro-angle measurement time data using different models;

[0048] Figure 7 A schematic diagram showing the comparison of denoising effect evaluation indicators for denoising and reconstruction of micro-angle measurement time data using different models;

[0049] Figure 8 A structural block diagram of a denoising and reconstruction system for micro-angle measurement time data based on BiLSTM-CNN provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] Figure 1 FIG. 1 is a flowchart of a method for denoising and reconstructing micro-angle measurement time data based on BiLSTM-CNN according to an embodiment of the present invention. Figure 1 As shown, the denoising and reconstruction method of micro-angle measurement time data based on BiLSTM-CNN proposed in the present invention specifically includes the following steps:

[0052] S11, obtaining micro-angle measurement time data of the line of sight pointing change micro-angle measurement system. Specifically comprising: receiving micro-angle measurement time data sent to the ground station in binary format by the line of sight pointing change micro-angle measurement system deployed on the satellite; decoding the micro-angle measurement time data in binary format and converting it into decimal data including positive and negative signs. In a specific example, the line of sight pointing change micro-angle measurement system is provided with two measurement lines or at least two measurement lines. In a specific embodiment, the micro-angle measurement system structure is as follows Figure 2 The micro-angle measurement time data includes a first micro-angle measurement time series and a second micro-angle measurement time series output by different measurement circuits of a micro-angle measurement system with a change in visual axis pointing.

[0053] In this embodiment, the line of sight pointing change micro-angle measurement system deployed on the satellite transmits data in binary form to the host computer of the ground station. In order to facilitate analysis and processing, it is necessary to perform base conversion. These binary floating-point numbers need to be converted into decimal numbers, and the binary data needs to be decoded and converted into the corresponding decimal value. Not only the size of the value needs to be considered, but also the positive and negative signs of the value need to be processed.

[0054] Take the example of a micro-angle measurement system for line of sight direction change including two measurement lines. The measurement method of the micro-angle measurement time series of each measurement line includes the following implementation steps:

[0055] Step 1, such as Figure 2 As shown, four angle measuring instruments and their reflectors are arranged at corresponding positions on the satellite frame.

[0056] Step 2: When the on-orbit space environment changes, the satellite's line of sight will undergo a slight spatial three-axis angle offset, and the corresponding image point positions on the image planes of the angle measuring sensors will change.

[0057] Step 3: Obtain the coordinates of the image points on the image plane of the sensors on the three angle measuring instruments, and calculate the pitch and yaw angle deflection of the corresponding reflector surface, where J1, J2, and J3 are a group, and J1, J4, and J3 are a group. The positions of J2 and J4 are fixed in parallel at the same reference value, that is, the angle change of the same reference is measured.

[0058] Step 4: Use the algorithm to calculate the angle deflection measured by the three angle measuring instruments to obtain the spatial three-axis angle offset of the satellite's on-orbit line of sight. Assume that the spatial three-axis angle changes of rigid bodies A and B relative to the world coordinate system are expressed as (a 1 ,b 1 ,c 1 )、(a 2 ,b 2 ,c 2 ). Then the corresponding image point coordinates on the sensor image planes of the three angle measuring instruments J1, J2 and J3 are (d x1 ,d y1 )、(d x2 ,d y2 ) and (d x3 ,d y3 ). Then the corresponding image point coordinates on the sensor image planes of the three angle measuring instruments J1, J4 and J3 are (d x1 ,d y1 )、(d x4 ,d y4 ) and (d x3 ,d y4 ).

[0059] For the combination of J1, J2, and J3, the relative change in the direction at A (a 1 ,b 1 ,c 1 ) is the change in the satellite's orbital axis pointing direction. This can be obtained by solving the following set of equations. MS = D, where M is the parameter matrix, which is determined by the position and relative angle of each angle measuring instrument and reflector at the "zero position". Where S is [a 1 ,b 1 ,c 1 ,a 2 ,b 2 ,c 2 ] T , represents the relative spatial three-axis angle change of rigid bodies A and B, where D is [d x1 ,d y1 ,d x2 ,d y2 ,d x3 ,d y3 ] T, which represents the change in the coordinates of the image points on the image plane of the sensor measured by the three angle measuring instruments. Solving for S =

[0060] M -1 D, where S is [a 1 ,b 1 ,c 1 ,a 2 ,b 2 ,c 2 ] T The first three items (a 1 ,b 1 ,c 1 ) is the micro-angle change in the direction of the satellite's orbital axis.

[0061] For the combination of J1, J4, and J3, the relative orientation change at A (a 1 ,b 1 ,c 1 ) is the change in the satellite's orbital axis pointing direction. This can be obtained by solving the following set of equations. MS = D, where M is the parameter matrix, which is determined by the position and relative angle of each angle measuring instrument and reflector at the "zero position". Where S is [a 1 ,b 1 ,c 1 ,a 2 ,b 2 ,c 2 ] T , represents the relative spatial three-axis angle change of rigid bodies A and B, where D is [d x1 ,d y1 ,d x4 ,d y4 ,d x3 ,d y3 ] T , which represents the change in the coordinates of the image points on the image plane of the sensor measured by the three angle measuring instruments. Solving for S =

[0062] M -1 D, where S is [a 1 ,b 1 ,c 1 ,a 2 ,b 2 ,c 2 ] T The first three items (a 1 ,b 1 ,c 1 ) is the micro-angle change in the direction of the satellite's orbital axis.

[0063] S12. Analyze the data variation characteristics and noise distribution rules of micro-angle measurement time data.

[0064] The main reasons for the change of the line of sight pointing are external vibration, thermal deformation and other factors, which have strong correlation and coherence in time. Therefore, the time series of the line of sight pointing change measured by the micro-angle measurement system shows strong time dependence. The present invention analyzes the factors affecting the change of the line of sight pointing, analyzes the data characteristics and noise sources of the micro-angle measurement time data in combination with the working environment of the micro-angle measurement system and the measured data, and obtains the data change law and noise distribution law of the micro-angle measurement time data.

[0065] S13. Construct a standard micro-angle measurement data set according to the data change characteristics of the micro-angle measurement time data, and add noise data that conforms to the noise distribution law to each standard micro-angle measurement time data in the standard micro-angle measurement data set to form a micro-angle measurement data set.

[0066] The present invention obtains the data variation law and noise distribution law of micro-angle measurement time data based on theoretical analysis and received micro-angle measurement data, and then generates a micro-angle measurement data set according to the data variation law and noise distribution law. Each micro-angle measurement data in the micro-angle measurement data set includes initial data and noisy data. The micro-angle measurement data set is divided into a training set, a validation set, and a test set, which are respectively used for model training, validation, and testing.

[0067] S14. Based on the micro-angle measurement data set, a BiLSTM-CNN model for realizing micro-angle measurement time series denoising is trained, and a Bayesian search algorithm is used to find the optimal hyperparameter combination of the BiLSTM-CNN model to obtain a micro-angle measurement time series denoising model.

[0068] The embodiment of the present invention builds a BiLSTM-CNN model for micro-angle measurement time series denoising, trains the BiLSTM-CNN network, uses Bayesian search to find the optimal hyperparameter combination, and obtains and saves the micro-angle measurement time series denoising model. Among them, BiLSTM is the abbreviation of Bi-directional Long Short-Term Memory, which is composed of forward LSTM and backward LSTM. It is suitable for processing sequence labeling tasks with contextual relationships, can capture bidirectional semantic dependencies, better model contextual information, is suitable for processing sequence problems, and can capture dependencies over longer distances.

[0069] S15. Use the micro-angle measurement time series denoising model to perform denoising and reconstruction on the micro-angle measurement time data to be processed to output denoised micro-angle measurement time data.

[0070] The denoising and reconstruction method of micro-angle measurement time data based on BiLSTM-CNN provided by the present invention integrates BiLSTM and CNN, firstly uses BiLSTM to capture the long-term dependency between micro-angle measurement time series, and then uses CNN to further extract features, thereby being able to encode the contextual information of the micro-angle measurement time series to ensure that each data point can utilize its contextual information, thereby providing CNN with more comprehensive time series features. The present invention can more accurately reconstruct the micro-angle measurement time series, effectively reduce the influence of noise, and improve the accuracy of the micro-angle measurement system.

[0071] In the embodiment of the present invention, the analysis of the data variation characteristics and noise distribution law of the micro-angle measurement time data in step S12 specifically includes the following steps not shown in the drawings:

[0072] S121, respectively performing annual change law analysis on the micro-angle measurement time data in units of years, daily change law analysis in units of days, and time change law analysis in units of hours, so as to obtain the change law of the micro-angle measurement time series over time. Specifically, the change law of each micro-angle measurement time series over time is: periodic change in units of years, periodic change in units of days, and periodic change in units of hours.

[0073] Specifically, data analysis is performed based on the working principle and working environment of the angle measurement system. According to the analysis, the experimental data should have the following characteristics:

[0074] On the whole, it should show periodic changes with certain rules on a yearly basis. This is because the period of the Earth's revolution around the sun and rotation is about one year. Therefore, the light intensity and area of ​​the geostationary synchronous orbit satellite receiving solar radiation should change periodically on a yearly basis. Changes in orbit will cause periodic changes in the external heat flow.

[0075] Locally, it should show periodic changes with certain rules on a daily basis. This is because the earth is rotating while also revolving, causing the relative position of the satellite and the sun to be different every day and every moment. However, since the geostationary orbit satellite rotates simultaneously with the earth and their relative position does not change, although the data values ​​are not the same on different dates within a year, the changes on each day follow a fixed trend.

[0076] Combined with the analysis of the received micro-angle measurement data, the data range is within a certain positive and negative numerical range, and changes slowly and smoothly with a 24-hour period.

[0077] S122, performing noise distribution analysis on the micro-angle measurement time data to obtain the noise distribution law of the micro-angle measurement time series. Further, performing noise distribution analysis on the micro-angle measurement time data includes: using a Monte Carlo simulation method to analyze the distribution law of the centroid solution deviation caused by the image sensor noise, and using the obtained centroid solution deviation distribution law as the noise distribution law of the micro-angle measurement time series. Specifically, the noise distribution law of each micro-angle measurement time series conforms to the normal distribution law.

[0078] In this embodiment, Monte Carlo simulation and D'Agostino-Pearson normality test methods are used to analyze and verify the noise distribution characteristics. The Monte Carlo simulation method is specifically implemented as follows: the Monte Carlo simulation method is used to simulate and generate an initial standard simulation image, and the OTSU centroid method is used to locate the light spot and calculate the centroid. Since the noise mainly comes from shot noise and thermal noise generated by resistor movement, appropriate Poisson distributed shot noise and Gaussian distributed thermal noise are added to the initial ideal image, and the D'Agostino-Pearson normality test is used, and the P value standard is set to 0.05.

[0079] In a specific example, for noise source analysis, there is thermal noise error caused by random movement of resistors in image sensor electronics, which is mainly manifested as Gaussian white noise; shot noise composed of input photon shot noise, photocurrent shot noise and dark current shot noise is caused by the statistical uncertainty of a large number of single events and obeys Poisson distribution in time. It is selected to add Poisson distributed shot noise with a gain of 5 to the simulated image; it is selected to add Gaussian noise with a mean of 2000 and a standard deviation of 0.1 to the simulated image.

[0080] Specifically, an initial simulation image of size 1100x1100 is created, and the target light spot covers an area of ​​11x11 pixels. The grayscale value of the light spot is concentrated in the central 5x5pixel area and follows a two-dimensional Gaussian function distribution. The grayscale value of the center of the light spot is set to 20000, and the OTSU centroid method is used to locate and calculate the center of mass of the light spot. By adding the above noise to the ideal image, the Monte Carlo simulation method and the D'Agostino-Pearson normality test are used to verify the noise distribution characteristics of the image noise mapped to the light spot centroid algorithm solution. The P value standard is set to 0.05, and the offset distribution is observed. If the final calculation result P value is greater than 0.05, it can be considered that the distribution law of the deviation distribution of the imaging noise of the image sensor mapped to the light spot centroid solution conforms to the normal distribution. The results are shown in Figure 3The P value of the X-direction offset distribution is 0.812276, and the P value of the Y-direction offset distribution is 0.685084, indicating that the deviation distribution of the imaging noise of the image sensor mapped to the spot centroid solution conforms to the normal distribution, that is, the deviation distribution of the micro-angle measurement data conforms to the normal distribution.

[0081] Furthermore, the present invention can generate a data set based on theoretical analysis and received micro-angle measurement data, and divide it into a training set, a validation set, and a test set. In a specific example, a total of eight years of micro-angle measurement data can be generated, and according to the noise distribution characteristics in the noise analysis, the corresponding Gaussian white noise is added to form a micro-angle measurement data set, and the added standard deviation of the Gaussian noise is in the range of 0-0.8, and the step size is 0.05. The first five years of the micro-angle measurement data set are used as the training set data, the sixth and seventh years are used as the validation set, and the eighth year is used as the test set. The test set is used to test the generalization level of the denoising ability of the model.

[0082] like Figure 4 As shown, the micro-angle measurement time series denoising model for micro-angle measurement time series denoising constructed in an embodiment of the present invention includes a first BiLSTM layer, a second BiLSTM layer arranged in parallel with the first BiLSTM layer, and a CNN layer connected to the output ends of the first BiLSTM layer and the second BiLSTM layer;

[0083] Among them, the first BiLSTM layer and the second BiLSTM layer both use the Tanh activation function, and the first BiLSTM layer and the second BiLSTM layer are respectively used to extract features from the two micro-angle measurement time series samples included in each micro-angle measurement time data sample in the micro-angle measurement data set to obtain the long-term dependency characteristics in each micro-angle measurement time series sample; each micro-angle measurement time data sample includes the first micro-angle measurement time series sample and the second micro-angle measurement time series sample representing the output of different measurement lines of the micro-angle measurement system of the visual axis pointing change. The CNN layer is composed of the first convolution layer, the first activation function layer, the first dropout layer and the second convolution layer. The CNN network uses the ReLU activation function. The CNN layer is used to extract and fuse the output results of the first BiLSTM layer and the second BiLSTM layer, and reconstruct the micro-angle measurement time series according to the fused features to obtain the denoised micro-angle measurement time data. In this embodiment, the BiLSTM layer performs feature extraction on the input micro-angle measurement time series data to capture the long-term dependencies in the sequence. The CNN part further extracts and fuses features, and reconstructs the micro-angle measurement time series data through a denoising process. The optimizer of the model selects the Adam optimizer.

[0084] Furthermore, the denoising and reconstruction method of micro-angle measurement time data based on BiLSTM-CNN proposed in the present invention also includes: the step of constructing a loss function of the micro-angle measurement time series denoising model, the loss function of the model includes a first loss function and a second loss function, the first loss function is the deviation loss loss1 of the micro-angle measurement time series after model denoising corresponding to the first micro-angle measurement time series sample and the original micro-angle time series, and the second loss function is the deviation loss loss2 of the micro-angle measurement time series after model denoising corresponding to the second micro-angle measurement time series sample and the original micro-angle time series;

[0085] The present invention adopts a dynamic gradient balancing algorithm to dynamically weigh the gradients of the first loss function and the second loss function according to a preset error standard level. The specific formula is as follows:

[0086]

[0087] loss=weight·loss1+(1-weight)·loss2 (2)

[0088] Where: grad1 and grad2 are the average values ​​of the gradient norms generated by the first loss function and the second loss function, respectively; loss is the total loss of the micro-angle measurement time series denoising model; and weight is the weight coefficient.

[0089] The present invention adds a gradient balancing algorithm to dynamically weigh the gradients of two different loss functions according to the different levels of input error standards, so as to achieve dynamic adjustment of the weights of the losses of the two micro-angle change measurement time series in the model, ensuring that each loss contributes as equally as possible to the final gradient, thereby avoiding the model from over-learning one of the micro-angle time series.

[0090] The present invention uses the Bayesian search algorithm to tune the parameters of the BiLSTM-CNN model. The parameters that need to be adjusted include: the number of BiLSTM layers (num_layers), the number of BiSLTM hidden layer nodes (hidden_size), the coefficient of the Dropout layer (dropout_rate), the batch size (batch_size), the learning rate (lr), the convolution kernel size (kernel_size), and the weight decay parameter (weight_decay). The Bayesian search flow chart is as follows Figure 5As shown in the figure, the training set and the validation set are sent into the model. Through the Bayesian optimization algorithm, the model parameters are set as follows: the number of BiLSTM layers is set to 4, the number of BISLTM hidden layer nodes is set to 44, the coefficient of the Dropout layer is set to 0.15, the batch size is set to 6, the learning rate is set to 0.0002, the convolution kernel size is 13, the weight decay parameter is set to 0.0001, the optimizer uses the Adam optimizer, the loss function is set to MSE, and the number of iterations is 200.

[0091] The embodiment of the present invention also provides a model evaluation step, which specifically includes inputting test data into a trained micro-angle measurement time series denoising model and evaluating the results.

[0092] Specifically, the test set is fed into the trained micro-angle measurement time series denoising model. The specific model is as follows: Figure 4 As shown, the data is based on Figure 4 The output of the post-process model is the denoised and reconstructed micro-angle measurement time series. For ease of display, a portion of the model noise identification and removal result graphs with input standard deviations of 0.2 and 0.7 are intercepted, as shown in the figure below. Figure 6 As shown. MSE (mean square error, Mean Squared Error), RMSE (root mean square error, Root Mean Square Error), MAE (mean absolute error, Mean Absolute Error), R 2 (Determination coefficient, R-Square) four evaluation indicators were compared and analyzed, and the results are as follows Figure 7 As shown in the figure. The denoising effect of BiLSTM-CNN is significantly improved compared with traditional RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), and BiLSTM (Bi-directional Long Short-Term Memory). Compared with BiLSTM, which has the best performance among the four traditional models, BiLSTM-CNN has achieved significant reductions of 58.01%, 35.20%, and 35.19% in MSE, RMSE, and MAE, respectively. 2 It also showed an improvement of 0.062%. The model was able to better identify and remove noise and retain the original features of micro-angle data to the greatest extent.

[0093] Compared with the prior art, the present invention has the following advantages:

[0094] In the BiLSTM-CNN architecture proposed in the present invention, the bidirectional dependency of the micro-angle measurement time series is first captured by BiLSTM, and the forward and backward bidirectional time information of the sequence is fully integrated, thereby providing CNN with richer time series features, and being able to more comprehensively understand the changing trend of micro-angle measurement data when processing time series data, thereby enabling CNN to more effectively extract these time-series enhanced features. This structural optimization enables the BiLSTM-CNN model to show better performance and stronger generalization ability, especially in the task of denoising and reconstructing micro-angle measurement time series, which requires high long-term dependencies. Compared with traditional models, this model performs well in multiple performance indicators, proving its effectiveness in denoising and data reconstruction, and providing a novel and reliable way to denoise the time series of the line-of-sight pointing micro-angle measurement system, which is of great significance to improving the accuracy of the line-of-sight pointing micro-angle measurement system.

[0095] Figure 8 The structure block diagram of the denoising and reconstruction system of micro-angle measurement time data based on BiLSTM-CNN according to one embodiment of the present invention is shown. The system is used to implement the denoising and reconstruction method of micro-angle measurement time data based on BiLSTM-CNN as described above, such as Figure 8 As shown, the denoising and reconstruction system for micro-angle measurement time data based on BiLSTM-CNN provided by an embodiment of the present invention includes:

[0096] An acquisition module 701 is used to acquire micro-angle measurement time data of a micro-angle measurement system for changing the visual axis direction;

[0097] Analysis module 702, used to analyze the data variation characteristics and noise distribution law of micro-angle measurement time data;

[0098] The data generation module 703 is used to construct a standard micro-angle measurement data set according to the data variation characteristics of the micro-angle measurement time data, and add noise data that conforms to the noise distribution law to each standard micro-angle measurement time data in the standard micro-angle measurement data set according to the noise distribution law of the micro-angle measurement time data to form a micro-angle measurement data set;

[0099] A model building module 704 is used to train a BiLSTM-CNN model for realizing micro-angle measurement time series denoising based on the micro-angle measurement data set, and use a Bayesian search algorithm to find the optimal hyperparameter combination of the BiLSTM-CNN model to obtain a micro-angle measurement time series denoising model;

[0100] The reconstruction module 705 is used to use the micro-angle measurement time series denoising model to perform denoising and reconstruction on the micro-angle measurement time data to be processed to output denoised micro-angle measurement time data.

[0101] In the embodiment of the present invention, the analysis module 702 includes a first analysis unit and a second analysis unit, wherein:

[0102] The first analysis unit is used to perform annual variation analysis based on years, daily variation analysis based on days, and time variation analysis based on hours on the micro-angle measurement time data, so as to obtain the variation pattern of the micro-angle measurement time series over time;

[0103] The second analysis unit is used to perform noise distribution analysis on the micro-angle measurement time data to obtain the noise distribution law of the micro-angle measurement time series.

[0104] In this embodiment, the micro-angle measurement time series denoising model includes a first BiLSTM layer, a second BiLSTM layer arranged in parallel with the first BiLSTM layer, and a CNN layer connected to the output ends of the first BiLSTM layer and the second BiLSTM layer;

[0105] The first BiLSTM layer and the second BiLSTM layer both use the Tanh activation function, and the first BiLSTM layer and the second BiLSTM layer are respectively used to extract features of two micro-angle measurement time series samples included in each micro-angle measurement time data sample in the micro-angle measurement data set, so as to obtain long-term dependency features in each micro-angle measurement time series sample; each micro-angle measurement time data sample includes a first micro-angle measurement time series sample and a second micro-angle measurement time series sample representing different measurement circuit outputs of the micro-angle measurement system with a change in the line of sight pointing;

[0106] The CNN layer is composed of a first convolutional layer, a first activation function layer, a first dropout layer and a second convolutional layer. The CNN network uses a ReLU activation function. The CNN layer is used to extract and fuse features of the output results of the first BiLSTM layer and the second BiLSTM layer, and reconstruct the micro-angle measurement time series according to the fused features to obtain denoised micro-angle measurement time data.

[0107] Furthermore, the model construction module 704 is also used to construct the loss function of the micro-angle measurement time series denoising model, the loss function includes a first loss function and a second loss function, the first loss function is the deviation loss loss1 of the micro-angle measurement time series after model denoising and the original micro-angle time series corresponding to the first micro-angle measurement time series sample, and the second loss function is the deviation loss loss2 of the micro-angle measurement time series after model denoising and the original micro-angle time series corresponding to the second micro-angle measurement time series sample; a dynamic gradient balancing algorithm is used to dynamically weigh the gradients of the first loss function and the second loss function according to a preset error standard level, and the specific formula is as follows:

[0108]

[0109] loss=weight·loss1+(1-weight)·loss2 (2)

[0110] Where: grad1 and grad2 are the average values ​​of the gradient norms generated by the first loss function and the second loss function, respectively; loss is the total loss of the micro-angle measurement time series denoising model; and weight is the weight coefficient.

[0111] The denoising and reconstruction method and system for micro-angle measurement time data based on BiLSTM-CNN provided by the present invention integrate BiLSTM and CNN, firstly use BiLSTM to capture the long-term dependency between micro-angle measurement time series, and then use CNN to further extract features, thereby being able to encode the contextual information of the micro-angle measurement time series to ensure that each data point can utilize its contextual information, thus providing CNN with more comprehensive time series features. The present invention can more accurately reconstruct the micro-angle measurement time series, effectively reduce the influence of noise, and improve the accuracy of the micro-angle measurement system.

[0112] In addition, those skilled in the art will appreciate that, although some embodiments herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present invention and form different embodiments. For example, any one of the claimed embodiments may be used in any combination.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A denoising and reconstruction method for micro-angle measurement time data based on BiLSTM-CNN, characterized in that: The method comprises: Acquire micro-angle measurement time data of the micro-angle measurement system of the line of sight pointing change; Analyze the data variation characteristics and noise distribution law of micro-angle measurement time data; A standard micro-angle measurement data set is constructed according to the data variation characteristics of the micro-angle measurement time data, and according to the noise distribution law of the micro-angle measurement time data, noise data conforming to the noise distribution law is added to each standard micro-angle measurement time data in the standard micro-angle measurement data set to form a micro-angle measurement data set; Based on the micro-angle measurement data set, a BiLSTM-CNN model for realizing micro-angle measurement time series denoising is trained, and a Bayesian search algorithm is used to find the optimal hyperparameter combination of the BiLSTM-CNN model to obtain a micro-angle measurement time series denoising model; The micro-angle measurement time series denoising model is used to perform denoising and reconstruction on the micro-angle measurement time data to be processed to output the denoised micro-angle measurement time data.

2. The method according to claim 1, characterized in that The micro-angle measurement time data of the micro-angle measurement system for obtaining the change of the line of sight pointing includes: Receive micro-angle measurement time data sent to the ground station in binary format by a micro-angle measurement system for line of sight pointing changes deployed on the satellite; Decode the micro-angle measurement time data in binary format and convert it into decimal data including positive and negative signs.

3. The method according to claim 1, characterized in that: The analysis of the data variation characteristics and noise distribution rules of the micro-angle measurement time data includes: The annual variation law analysis of the micro-angle measurement time data is carried out in the year, the daily variation law analysis is carried out in the day, and the time variation law analysis is carried out in the hour to obtain the variation law of the micro-angle measurement time series over time; A noise distribution analysis is performed on the micro-angle measurement time data to obtain the noise distribution law of the micro-angle measurement time series.

4. The method according to claim 3, characterized in that Performing noise distribution analysis on the micro-angle measurement time data includes: The Monte Carlo simulation method is used to analyze the distribution law of the centroid solution deviation caused by image sensor noise, and the obtained distribution law of the centroid solution deviation is used as the noise distribution law of the micro-angle measurement time series.

5. The method according to claim 3, characterized in that: The micro-angle measurement time data includes a first micro-angle measurement time series and a second micro-angle measurement time series output by different measurement circuits of a micro-angle measurement system for changing the line of sight direction; The variation pattern of each micro-angle measurement time series over time is: periodic change in years, periodic change in days and periodic change in hours; The noise distribution law of each micro-angle measurement time series conforms to the normal distribution law.

6. The method according to claim 1, characterized in that The micro-angle measurement time series denoising model includes a first BiLSTM layer, a second BiLSTM layer arranged in parallel with the first BiLSTM layer, and a CNN layer connected to the output ends of the first BiLSTM layer and the second BiLSTM layer; The first BiLSTM layer and the second BiLSTM layer both use the Tanh activation function, and the first BiLSTM layer and the second BiLSTM layer are respectively used to extract features of two micro-angle measurement time series samples included in each micro-angle measurement time data sample in the micro-angle measurement data set, so as to obtain long-term dependency features in each micro-angle measurement time series sample; Each micro-angle measurement time data sample includes a first micro-angle measurement time series sample and a second micro-angle measurement time series sample representing outputs of different measurement circuits of the micro-angle measurement system for changing the line of sight direction; The CNN layer is composed of a first convolutional layer, a first activation function layer, a first dropout layer and a second convolutional layer. The CNN network uses a ReLU activation function. The CNN layer is used to extract and fuse features of the output results of the first BiLSTM layer and the second BiLSTM layer, and reconstruct the micro-angle measurement time series according to the fused features to obtain denoised micro-angle measurement time data.

7. The method according to claim 6, characterized in that The method further comprises: Constructing a loss function of the micro-angle measurement time series denoising model, the loss function includes a first loss function and a second loss function, the first loss function is a deviation loss loss1 between the micro-angle measurement time series after model denoising and the original micro-angle time series corresponding to the first micro-angle measurement time series sample, and the second loss function is a deviation loss loss2 between the micro-angle measurement time series after model denoising and the original micro-angle time series corresponding to the second micro-angle measurement time series sample; The dynamic gradient balancing algorithm is used to dynamically weigh the gradients of the first loss function and the second loss function according to the preset error standard level. The specific formula is as follows: loss=weight·loss1+(1-weight)·loss2 (2) Where: grad1 and grad2 are the average values ​​of the gradient norms generated by the first loss function and the second loss function, respectively; loss is the total loss of the micro-angle measurement time series denoising model; and weight is the weight coefficient.

8. The method according to claim 1, characterized in that The hyperparameters of the BiLSTM-CNN model include: the number of BiLSTM layers and the number of hidden layer nodes of the BiLSTM layer, the coefficient of the Dropout layer, the batch size, the learning rate, the convolution kernel size, and the weight decay parameter.

9. A denoising and reconstruction system for micro-angle measurement time data based on BiLSTM-CNN, characterized in that: The system comprises: An acquisition module, used for acquiring micro-angle measurement time data of a micro-angle measurement system for changing the visual axis pointing direction; An analysis module is used to analyze the data variation characteristics and noise distribution law of micro-angle measurement time data; A data generation module is used to construct a standard micro-angle measurement data set according to the data change characteristics of the micro-angle measurement time data, and add noise data that conforms to the noise distribution law to each standard micro-angle measurement time data in the standard micro-angle measurement data set according to the noise distribution law of the micro-angle measurement time data to form a micro-angle measurement data set; A model building module is used to train a BiLSTM-CNN model for realizing micro-angle measurement time series denoising based on the micro-angle measurement data set, and use a Bayesian search algorithm to find the optimal hyperparameter combination of the BiLSTM-CNN model to obtain a micro-angle measurement time series denoising model; The reconstruction module is used to use the micro-angle measurement time series denoising model to perform denoising and reconstruction on the micro-angle measurement time data to be processed to output the denoised micro-angle measurement time data.

10. The system according to claim 9, characterized in that The analysis module comprises: The first analysis unit is used to perform annual variation analysis based on years, daily variation analysis based on days, and time variation analysis based on hours on the micro-angle measurement time data, so as to obtain the variation pattern of the micro-angle measurement time series over time; The second analysis unit is used to perform noise distribution analysis on the micro-angle measurement time data to obtain the noise distribution law of the micro-angle measurement time series.

Images