GPS-INS combined positioning method based on RNN-CNN neural network

Abstract

The present invention relates to a GPS-INS combined positioning method based on an RNN-CNN neural network, comprising the following steps: constructing a GPS data set and an INS data set; building a GRU-based recurrent neural network (RNN) model; designing a time-series-based convolutional neural network (CNN) model; designing a stacking integrated module; and designing a GPS-INS combined navigation method: when a GPS signal is available, using the neural network to train an error model of the INS system; and when the GPS signal is missing, using the neural network to predict the INS error and correct the INS output.

Description

Technical Field

[0001] The present invention relates to a GPS-INS combined positioning and navigation technology, which mainly utilizes a neural network system to predict positioning information when a GPS signal is missing in the GPS-INS combined positioning and navigation. Background Art

[0002] In recent years, with the increasing development of transportation, integrated positioning and navigation technology has emerged. It can guide us to our destinations and provide real-time navigation information such as attitude, speed, and location during navigation. In integrated positioning and navigation, when the GPS signal is available, integrated navigation is used to provide navigation information to the vehicle. However, complex real-world environments often cause GPS signal quality to deteriorate or even disappear. When the GPS signal is lost, the GPS / INS integrated navigation system degenerates to a single inertial navigation system, and the navigation system's output accuracy rapidly decreases. To address this issue, many studies have designed neural network models to predict navigation signals when the GPS signal is lost. In this case, the navigation data predicted by the neural network model continues to provide reliable navigation information to the vehicle.

[0003] At present, there has been a lot of research on the application of neural networks in GPS-INS combined positioning and navigation at home and abroad. Wang Guangcai et al. proposed a new back propagation neural network model. When the GPS signal is good enough, the neural network learns the working mode of GPS / INS and establishes a corresponding model to map the vehicle's dynamic factors (attitude, speed, position) and the corresponding errors. The error obtained by the neural network is used to compensate for the INS drift when the GPS signal is interrupted. [1] Wang Guocheng and others from the Chinese Academy of Sciences introduced radial basis function neural networks into GPS / INS integrated navigation. They used a tight combination model and assisted by a three-layer radial basis function neural network (RBFNN) to effectively improve the error accumulation problem when INS works alone. [2] M Malleswaran et al. designed a GPS / INS integrated navigation system using convolutional neural networks. They were able to achieve relatively accurate positioning even when the GPS signal was interrupted for a long time. The accuracy was better than that of the extended Kalman filter and the unscented Kalman filter. [2] In order to solve the navigation problem when GPS signal is lost, Zhao Xuefeng designed a static BP neural network combined with a dynamic Elman neural network prediction model to learn the error model of INS. [4]Domestic and international research has confirmed that neural networks can play an important role in GPS / INS integrated navigation, addressing the impact of multipath effects on GPS signals and improving positioning accuracy. Similarly, these methods also face computational complexity and modeling difficulties. One of the current research focuses is how to simply and effectively use neural networks to address the impact of GPS signal loss in integrated systems. Based on neural networks, this paper proposes a method based on an RNN-CNN neural network model to predict the positioning results of the integrated system when the GPS signal is lost during GPS-INS integrated positioning and navigation.

[0004] References:

[0005] [1]Wang G, Xu

[0006] [2] Wang Guocheng, Liu Lintao, Xu Aigong, et al. Application of radial basis function neural network in GPS satellite clock error prediction [J]. Acta Geodaetica et Cartographica Sinica, 2014, 43(08):803-807.

[0007] [3]Malleswaran M, Vaidehi V, Deborah S A.CNN based GPS / INS dataintegration using new dynamic learning algorithm[C].International Conference on Recent Trends in Information Technology, 2011:211-216.

[0008] [4] Zhao Xuefeng. Research and implementation of integrated navigation technology when satellite signals are temporarily lost[D]. Chengdu: University of Electronic Science and Technology of China, 2017. Summary of the Invention

[0009] This invention addresses the computational complexity and modeling difficulties of existing neural network methods used in GPS-INS combined positioning and navigation. It provides a more effective GPS-INS combined positioning method. This method uses an RNN-CNN neural network model within a neural network to predict positioning information in the combined navigation system, thereby eliminating the impact of GPS signal loss on the combined system. The technical solution is as follows:

[0010] A GPS-INS combined positioning method based on RNN-CNN neural network includes the following steps:

[0011] The first step is to construct GPS and INS datasets and perform data preprocessing.

[0012] The second step is to build a recurrent neural network (RNN) model based on GRU: the input time series INS data first flows into the first layer of the GRU layer expanded in a one-to-one form. After passing through the first layer of the GRU layer, the output length remains unchanged, and the memory state of each time step is an updated sequence. Then this sequence flows into the second layer of the GRU layer expanded in a many-to-one form. The second layer of the GRU layer outputs the new memory state of the next time step and serves as the input of the third layer of the GRU layer; the third layer of the GRU layer is equipped with a fully connected neural network Dense layer to adjust the stability of the neural network, and the number of neurons is 1 / 2 of the previous layer; the fourth layer of the GRU layer is continued to be equipped with a Dense layer, and the number of neurons is also 1 / 2 of the previous layer; finally, the output layer outputs the INS error value at the target moment, and the various GRU layers are connected in a stacked manner;

[0013] The third step is to design a time-series-based convolutional neural network (CNN) model: The data with N-dimensional attributes, including INS position, velocity, and accumulated error, from the last L moments are expanded and compared to grid data. This data is processed through a convolutional neural network with a convolution kernel size of 3×3 and a stride of 1. In the pooling layer, a maximum pooling function is selected with a pooling kernel size of 2×2 and a stride of 2. After performing the maximum pooling operation on the input data, the resulting features are first flattened into one dimension and fed into a fully connected layer for ReLU activation, which outputs the INS error.

[0014] The fourth step is the design of the Stacking integration module: In the Stacking integration module, the models that need to be combined are called individual learners, and the model used to combine other learners is called a meta-learner. RNN and CNN are used as individual learners, and the fully connected neural network is used as a meta-learner. The RNN-CNN prediction model is divided into two levels. Before training the RNN-CNN, the training set is first divided into two parts. The first part is used for the RNN and CNN in the first level respectively. The model trained in the first level is then used to predict the second part of the training set. The obtained output prediction values are spliced as new inputs to train the second level fully connected neural network model. The final output is the measured value of the second part of the training set at the corresponding time.

[0015] Step 5: GPS-INS integrated navigation design: When GPS signals are available, the neural network is used to train the INS system's error model; when GPS signals are missing, the neural network is used to predict INS errors and correct INS outputs.

[0016] Furthermore, in the fifth step, when the GPS signal is available, the neural network is in training mode. At this time, the three-dimensional position information output by the INS system is used as the input sample for network training, and the INS error compensation value output after fusion by the adaptive fading Kalman filter is used as the expected output sample for network training. When the error value between the actual output and the expected output is greater than the set threshold, the network's weight update algorithm is used to update the weight information between each neuron until a set of optimal weight sets is obtained, so that the error between the actual output of the network and the expected output is less than the set threshold.

[0017] The implementation of the present invention's combined GPS-INS positioning method based on an RNN-CNN neural network has the following beneficial effects: To address this issue, an RNN-CNN neural network model is proposed, and five modules within the RNN-CNN model are designed: a GRU-based recurrent neural network (RNN) model, a time-series-based convolutional neural network (CNN) model, a stacking integration module, and a GPS-INS combined navigation module. This solves the problem of severely reduced positioning accuracy due to GPS signal loss in GPS-INS combined positioning and navigation. Furthermore, the present method is simple, low-cost, and can be directly applied to vehicle-mounted positioning modes. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Structure diagram of the INS error model training process.

[0019] Figure 2 Two gridded representations of time series data.

[0020] Figure 3 Structure diagram of the time series-based CNN prediction model.

[0021] Figure 4 Prediction process structure diagram based on the Stacking integration framework.

[0022] Figure 5 Structure diagram of the meta-learner based on a fully connected neural network.

[0023] Figure 6 Figure 2. Structure diagram of the missing GPS prediction process based on the RNN-CNN integrated deep learning model.

[0024] Figure 7 GPS / INS integrated navigation system training process structure diagram.

[0025] Figure 8 INS prediction process structure diagram.

[0026] Figure 9RMSE experimental results of 100 non-continuous missing points.

[0027] Figure 10 RMSE experimental results of 300 non-continuous missing points. DETAILED DESCRIPTION

[0028] The proposed method mainly includes: GRU-based recurrent neural network (RNN) model design, time-series-based convolutional neural network (CNN) model design, stacking integration module design, and GPS-INS integrated navigation design. The main steps include:

[0029] The first step is to build a data set and perform data preprocessing;

[0030] A GPS dataset and an INS dataset were constructed, and data preprocessing was performed on the GPS position, velocity, and satellite information in the GPS system, and on the velocity, acceleration, and attitude information in the INS system. The data used in this paper was derived from INS data collected by the DB502 board. A total of 56,890 sets of GPS information data were collected, with the actual INS data corresponding to each location point as output. 40,000 sets of data were used as training data, of which 20,000 sets were used as the first training set, 20,000 sets as the second training set, and the remaining 16,890 sets of data were used as test data for the experiment.

[0031] The second step is to build a GRU-based recurrent neural network (RNN) model. The main design idea is that the input time series INS data first flows into the first layer of the GRU layer, which is expanded in a one-to-one manner. The output of the first GRU layer is a sequence of constant length with updated memory state at each time step. This sequence then flows into the second layer of the GRU layer, which is expanded in a many-to-one manner, and outputs the new memory state at the last time step. The effect of superimposing the first two GRU layers is equivalent to first making a prediction for each time step, then performing a trend prediction on the sequence of preliminary predictions, and finally updating the memory result of the time step. In addition, the number of neurons in the second layer of GRU is set to the same, 256. To prevent instability when superimposing two GRU layers, a fully connected neural network (Dense) layer is added to the third layer for adjustment, with the number of neurons being 1 / 2 of the previous layer. The fourth layer is also equipped with a Dense layer, with the number of neurons also being 1 / 2 of the previous layer. Finally, the output layer outputs the error value at the target moment. The network layers are connected in a stacked manner. The structural diagram of the INS error model training process is shown below. Figure 1 .

[0032] The third step is to design a convolutional neural network (CNN) model based on time series. This step first involves the expansion of time series data. In the application scenario of the present invention, it is first necessary to expand the data with N-dimensional attributes (including INS position, speed, and cumulative error, etc.) at the latest L moments and then compare them to grid data to facilitate convolutional neural network processing. Since the present invention uses data from the past 10 cycles to predict the future INS error of the site, L = 10. The expansion form can generally be divided into two types: expanding from top to bottom in chronological order and from left to right in attributes, and expanding from left to right in chronological order and from top to bottom in attributes. The present invention chooses the expansion form of expanding from top to bottom in chronological order and from left to right in attributes. See for details. Figure 2 . Finally, the architecture design. The larger the convolution kernel size, the larger the receptive field, and the local features extracted are of a larger spatial range. However, there are more parameters in the convolution calculation, and the amount of calculation is also greater. Considering that the GPS data grid data dimension in this article is not high, and considering various factors, the convolution kernel size selected is 3×3, and the stride is 1; the pooling layer selects the maximum pooling function, the pooling kernel size is 2×2, and the stride s is 2. After performing the maximum pooling operation on the input data, it is necessary to first pull the obtained features into one dimension, and then input them into the fully connected layer for ReLU activation, and finally output, such as Figure 3 .

[0033] The fourth step is the design of Stacking integration module. The basic idea of ensemble learning is to fuse multiple learners through a certain strategy. Compared with a single learner, ensemble learning has better generalization performance and occupies an important position in many famous machine learning competitions. In the Stacking integration framework, the model that needs to be combined is called an individual learner, and the model used to combine other learners is called a meta-learner. In the Stacking integration framework, you first need to divide the data set into three parts: the first part is the training set, the second part is the training set, and the test set; then use the first part of the training set to train different individual learners respectively; then input the second part of the training set into the previously trained individual learner to obtain the prediction results, and splice the different prediction results into new inputs. The measured values of the second part of the training set are used as output to train the meta-learner; finally, the test set is injected into the individual learner, and the prediction results output by the individual learner flow into the meta-learner to output the final prediction results. The prediction process based on the Stacking integration framework is as follows: Figure 4In order to give full play to the temporal memory advantage of RNN and the automatic feature extraction advantage of CNN, the Stacking integration model designed in this paper uses the above RNN and CNN as individual learners and the fully connected neural network as a meta-learner. Taking into account the simplicity, balance and nonlinearity of the meta-learner, this paper builds two fully connected layers in the fully connected neural network structure of the second level. The number of neurons in the first layer is 128, and ReLU is selected for nonlinear activation. The number of neurons in the second layer is 1 / 2 of the first layer, and linear activation is selected. Finally, the output layer is connected as follows Figure 5 As shown. The entire prediction model is an RNN-CNN prediction model. RNN-CNN is divided into two levels, and before training RNN-CNN, the training set is first divided into two parts. The first part is used for RNN and CNN in the first level respectively, and then the model trained in the first level is used to predict the second part of the training set. The output prediction value is spliced as a new input to train the second level fully connected neural network model. The final output is the measured value of the second part of the training set at the corresponding moment. The specific flow chart is as follows Figure 6 shown.

[0034] The fifth step is the design of GPS-INS combined navigation. When the GPS signal is available, combined navigation is used to provide navigation information to the carrier. However, the complex real environment often causes the GPS signal quality to deteriorate or be lost. When the GPS signal is lost, the GPS / INS combined navigation system degenerates into a single inertial navigation, and the output accuracy of the navigation system will drop rapidly. In order to solve this problem, many studies have designed neural network models to predict the navigation signal when the GPS signal is lost. When the GPS signal is lost, the navigation data predicted by the neural network model continues to provide reliable navigation information to the carrier. The main idea of the present invention is to use a neural network to train the error model of the INS system when the GPS signal is available; when the GPS signal is lost, the neural network is used to predict the error of the INS and correct the output of the INS. The application of neural networks in combined navigation systems is mainly divided into two processes, the INS error model training process and the INS error model prediction process. The principle of the INS error model training process is as follows: Figure 7 As shown in the figure. When the GPS signal is available, the neural network is in training mode. The three-dimensional position information output by the INS system is used as the input sample for network training, and the INS error compensation value output after fusion through the adaptive fading Kalman filter is used as the expected output sample for network training. When the error between the actual output and the expected output exceeds the set threshold, the network's weight update algorithm is used to update the weight information between each neuron until an optimal set of weights is obtained, so that the error between the actual and expected outputs of the network is less than the set threshold.

[0035] The prediction process of the neural network is as follows Figure 8 When the GPS signal is lost, the neural network enters prediction mode. Similarly, it uses the navigation position information output by the INS system as the network input, uses the previously trained neural network model to predict the INS system error value, and uses this error value to correct the INS navigation output, providing corrected navigation information to the mobile vehicle.

[0036] The main process of the adaptive fading Kalman filter algorithm used in GPS-INS integrated positioning and navigation is:

[0037] One-step estimate of the state:

[0038] X k / k-1 =Φ k / k-1 X k-1 / k-1 (1)

[0039] State Estimation:

[0040] X k / k =X k / k-1 +K k (Z k -H k X k / k-1 ) (2)

[0041] Filter gain:

[0042]

[0043] Forecast state error covariance:

[0044]

[0045] Estimated mean squared error:

[0046] P k =(IK k H k )P k / k-1 (5)

[0047] The only difference from the Kalman filter algorithm process is that the fading factor λ is introduced into the predicted state covariance of formula (4) k (λ k >=1). Therefore, the selection method of the fading factor is the key to this algorithm. The following is the selection method of the fading factor.

[0048] In Kalman filtering, the information sequence at time k is:

[0049] r k =Z k -H k X k / k-1 (6)

[0050] The actual measurement variance matrix is:

[0051]

[0052] R k is the sequence V of the measurement noise vector k The variance matrix of , and it is a positive definite matrix.

[0053] The information autocovariance function is:

[0054]

[0055] Substituting the filter gain of formula (3) and the information sequence of formula (6) into formula (8), we can obtain: It shows that when the gain matrix K k When is the optimal gain matrix, the autocorrelation function of the information is 0, that is, the information sequence remains orthogonal. Further, we can get:

[0056]

[0057] Fading factor λ k The role of is to use the fading factor λ in actual situations when the information sequence is not orthogonal. k Adjust K k Make the information sequence orthogonal.

[0058] Let the new gain matrix be:

[0059]

[0060] In the above formula, are the prediction error covariance, C k Predicted value, fading factor and scalar factor. Therefore, when λ k , α k When the optimal value is reached, the gain matrix is optimal. The expression is:

[0061]

[0062] Substitute the above formula into We can get α k The expression is:

[0063]

[0064] The current estimated information covariance can also be expressed as:

[0065]

[0066] It can be seen from this that the information covariance can be obtained by λ kTo adjust, we can get the following equation (13):

[0067]

[0068] From formula (14), we can see that α k The value of is very close to 1, so (1-α k )R k Close to 0, it can be ignored, and then we can get α k and λ k Approximately equal, let α k =λ k , we can get the fading factor λ k It can be expressed as:

[0069]

[0070] Bringing λk into the Kalman filter equation can dynamically increase or reduce the influence of the current measurement value on the state variables in the recursive process, reducing the influence of unclear system model establishment or other noise.

[0071] Step 6: Experimental comparison and verification. In order to verify the RNN-CNN neural network model, the present invention uses the model of the present invention to compare with the RNN neural network model. Randomly select 100 missing points and randomly select 300 missing points to use the traditional RNN prediction algorithm and RNN-CNN neural network model algorithm for simulation respectively. The simulation results are as follows: Figure 9 , as shown in Figure 10. When the number of missing points is 100, the longitude and latitude errors of the RNN-CNN algorithm proposed in this paper increase by 9.62%, while the error of the RNN algorithm increases by 16.35%; when the number of missing points is 300, the longitude and latitude errors of the RNN-CNN algorithm proposed in this paper increase by 15.38%, while the error of the RNN algorithm increases by 55.74%.

Claims

1. A GPS-INS combined positioning method based on RNN-CNN neural network, comprising the following steps: The first step is to construct GPS and INS datasets and perform data preprocessing. The second step is to build a recurrent neural network (RNN) model based on GRU: the input time series INS data first flows into the first layer of the GRU layer expanded in a one-to-one form. After passing through the first layer of the GRU layer, the output length remains unchanged, and the memory state of each time step is an updated sequence. Then this sequence flows into the second layer of the GRU layer expanded in a many-to-one form. The second layer of the GRU layer outputs the new memory state of the next time step and serves as the input of the third layer of the GRU layer; the third layer of the GRU layer is equipped with a fully connected neural network Dense layer to adjust the stability of the neural network, and the number of neurons is 1 / 2 of the previous layer; the fourth layer of the GRU layer is continued to be equipped with a Dense layer, and the number of neurons is also 1 / 2 of the previous layer; finally, the output layer outputs the INS error value at the target moment, and the various GRU layers are connected in a stacked manner; The third step is to design a time-series-based convolutional neural network (CNN) model: The data with N-dimensional attributes, including INS position, velocity, and accumulated error, from the last L moments are expanded and compared to grid data. This data is processed through a convolutional neural network with a convolution kernel size of 3×3 and a stride of 1. In the pooling layer, a maximum pooling function is selected with a pooling kernel size of 2×2 and a stride of 2. After performing the maximum pooling operation on the input data, the resulting features are first flattened into one dimension and fed into a fully connected layer for ReLU activation, which outputs the INS error. The fourth step is the design of the Stacking integration module: In the Stacking integration module, the models that need to be combined are called individual learners, and the model used to combine other learners is called a meta-learner. RNN and CNN are used as individual learners, and the fully connected neural network is used as a meta-learner. The RNN-CNN prediction model is divided into two levels. Before training the RNN-CNN, the training set is first divided into two parts. The first part is used for the RNN and CNN in the first level respectively. The model trained in the first level is then used to predict the second part of the training set. The obtained output prediction values are spliced as new inputs to train the second level fully connected neural network model. The final output is the measured value of the second part of the training set at the corresponding time. Step 5: GPS-INS integrated navigation design: When the GPS signal is available, the neural network is used to train the error model of the INS system; when the GPS signal is missing, the neural network is used to predict the INS error and correct the INS output.

2. The GPS-INS combined positioning method according to claim 1, in the fifth step, when the GPS signal is available, the neural network is in training mode, at which time the three-dimensional position information output by the INS system is used as the input sample for network training, and the INS error compensation value output after fusion by the adaptive fading Kalman filter is used as the expected output sample for network training. When the error value between the actual output and the expected output is greater than a set threshold, the network's weight update algorithm is used to update the weight information between each neuron until a set of optimal weight sets is obtained so that the error between the actual output of the network and the expected output is less than the set threshold.

Images