A deep learning-based ship positioning method for coastal scenarios

Abstract

The present invention discloses a deep learning-based ship positioning method for coastal scenarios, which generates preliminary positioning features through data enhancement, and then effectively identifies the positioning features of the NLOS noise environment after data enhancement through the DenseNet model, thereby obtaining more accurate target ship positioning. Data enhancement is to generate multiple pre-positioning coordinate features through a group pre-positioning method, group the original distance data according to base stations, and use a mathematical model to perform preliminary positioning on each group of distance data to generate preliminary positioning features, which maximizes the use of distance features and base station position features when the number of base stations is small. DenseNet is a feature learning method for strong noise or highly redundant data. The model can effectively identify the positioning features of the NLOS noise environment after data enhancement, and can effectively alleviate NLOS errors in coastal scenarios with fewer base stations.

Description

Technical Field

[0001] The present invention belongs to the field of ship positioning technology, and specifically relates to a deep learning-based ship positioning method for coastal scenarios. The method is applicable to NLOS (Non-Light Of Sight) error elimination in coastal scenarios with a small number of base stations (e.g., four base stations) and targets at special GDOP (Geometric Dilution Precision). Background Art

[0002] The marine environment plays a vital role in sustaining the global economy and our daily lives. As we all know, the ocean is not only home to numerous marine species but also a venue for human activities such as sailing, cruising, fishing, and cargo transportation. In recent years, the increasing number of ships, traffic density, and the amount of dangerous goods carried have posed potential threats to maritime safety. Frequent marine accidents result in significant loss of life and property, and severely damage the marine ecosystem in adjacent waters.

[0003] For nearly two decades, AIS (Automatic Identification System) has been widely used in the maritime sector to monitor maritime traffic and prevent collisions. However, a major practical challenge is that some ships fail to maintain their AIS systems in proper working order, inputting false information into the equipment. Consequently, AIS signals cannot be effectively received by other ships and land-based service agencies, directly impacting the effectiveness of AIS equipment and posing a serious threat to maritime traffic safety. This has led to numerous problems with maritime surveillance.

[0004] To ensure coastal navigation safety, the World Radionavigation Plan developed by the International Maritime Organization (IMO) calls on member states worldwide to add ranging and positioning capabilities, known as AIS-R mode, to existing AIS ship stations. AIS-R mode uses wireless electronic measurements to determine the distance between the ship and a land-based base station for positioning. However, in practical applications, NLOS paths are a significant factor affecting positioning accuracy. Due to the lack of a line-of-sight (LOS) path, signals can only reach the receiver via penetration, reflection, diffraction, or scattering. Signal propagation under NLOS conditions can result in positive ranging errors, severely impacting positioning accuracy. Most current research is based on cellular networks, assuming a target location with sufficient base stations and good GDOP conditions. However, limited research has addressed positioning challenges in coastal scenarios with limited base station availability and the increased severity of NLOS propagation. This scenario has strong engineering application and practical value. Summary of the Invention

[0005] To address the problem of low positioning accuracy due to NLOS errors in special coastal scenarios, the present invention provides a deep learning-based ship positioning method for coastal scenarios to obtain more accurate target ship positioning.

[0006] The specific steps of the present invention are as follows:

[0007] Step 1: Use N b The base stations measure the distance to the test target, calculate the slope k and intercept b that minimize the system error, and obtain the measured distance d after eliminating the system error of each base station. ′ mi ; The least squares fitting formula is as follows:

[0008]

[0009] Among them, k represents the slope of the line after the least squares fitting, b represents the intercept of the line after the least squares fitting, and d ti represents the actual distance between the i-th base station and the test target, d mi represents the measured distance of the i-th base station before eliminating the system error.

[0010] Step 2: Let the minimum number of base stations required be m. Since the total number of base stations is N b , then the minimum number of base stations required is obtained Then, for each base station combination, the measured distance d after eliminating the system error of each base station is ′ mi, a three-side positioning method based on PSO is used, and the particle swarm algorithm is introduced to find the optimal positioning position of the test target determined by each base station combination.

[0011] Step 3: Obtain an N according to the position coordinates of each base station, the optimal positioning position of the test target determined for each base station combination in step 2, and the test target position coordinates. pos ×2 coordinate matrix:

[0012]

[0013] in The optimal positioning position of the test target determined for the p-th base station combination, n in {m,m+1,…,N b}, (x ms ,y ms ) is the test target position coordinate, (x bsi ,y bsi ) is the location coordinate of the i-th base station.

[0014] The coordinate matrix M is converted into a two-dimensional pixel matrix through a plane rectangular coordinate system, and the two-dimensional pixel matrix is ​​used as the data set of the Densenet model.

[0015] Step 4: Train the DenseNet model using the dataset constructed in step 3, and use the trained DenseNet model to estimate the position of the target ship.

[0016] Furthermore, the steps of the particle swarm algorithm are as follows:

[0017] Step 1: Take the test target position as the optimization target, use X pj =(x pj ,y pj ) and V pj =(v pj ,v pj ) to represent the position and velocity of the jth particle, initialize the particle swarm, and the velocity of each particle V pj =(v pj ,v pj ) and position X pj =(x pj ,y pj ) takes a random value during initialization.

[0018] Step 2: Establish a fitness function to determine the fitness value of each iteration of the particles in the particle swarm. The fitness function of the jth particle is defined as follows:

[0019]

[0020] in, n in {m,m+1,…,N b}, the measured distance expression of the p-th base station after eliminating the system error is:

[0021]

[0022] Where (x ms ,y ms ) represents the test target position coordinates, (x bsp ,y bsp ), p∈(1,2,…,s) represents the location coordinates of the p-th base station, n p represents the NLOS error of the p-th base station.

[0023] Step 3: The particle swarm tracks the optimal value X of the particle during the iteration process prej =(x prej ,y prej ) and the particle swarm global optimal value X g =(x g ,y g ), if according to the current particle swarm global optimal value X g If the calculated fitness value is less than the preset error value, the test ends and the optimal positioning position (x esp ,y esp ), p∈{1,2,3,…,s}, otherwise, it is determined whether the current fitness value of each particle in the particle swarm is better than the optimal value X recorded in the last iteration of the particle. prej The obtained fitness value is better, then the optimal value X of the particle is updated. prej , and judge whether the current fitness value of each particle in the particle swarm is better than the global optimal value X of the particle swarm recorded in the last iteration g The obtained fitness value is better, then the global optimal value of the particle swarm is updated. g .

[0024] Furthermore, the speed and position update rules of the jth particle are as follows:

[0025]

[0026] Where t represents the number of iterations, c1 and c2 represent learning factors, and r1 and r2 are random probability values ​​between [0,1].

[0027] Furthermore, a part of the dataset constructed in step 3 is selected as the training set and the other part is selected as the validation set to train and validate the Densenet model.

[0028] The beneficial effects of the present invention are:

[0029] The present invention generates preliminary positioning features through data enhancement, and then effectively identifies the positioning features of the NLOS noise environment after data enhancement through the DenseNet model, thereby obtaining more accurate positioning of the target ship. Among them, data enhancement is to generate multiple pre-positioning coordinate features through a group pre-positioning method, group the original distance data according to base stations, and use a mathematical model to perform preliminary positioning on each group of distance data to generate preliminary positioning features. When the number of base stations is small, the distance features and base station position features are maximized. DenseNet is a feature learning method for strong noise or highly redundant data. It uses information reuse to make the extracted data more complete. The model can effectively identify the positioning features of the NLOS noise environment after data enhancement, and can effectively alleviate NLOS errors in coastal scenarios with fewer base stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a positioning model diagram in a coastal scene.

[0031] Figure 2 It is the feature image generated after the coordinate matrix is ​​converted into a two-dimensional pixel matrix.

[0032] Figure 3 It is a diagram of the DenseNet structure with three dense blocks and two transition layers. DETAILED DESCRIPTION

[0033] The present invention will be described in further detail below with reference to the accompanying drawings.

[0034] Systematic error and NLOS error are the main factors affecting base station ranging accuracy. In coastal scenarios, where there are fewer base stations and the target is located outside the polygon formed by multiple base stations, LOS base stations have difficulty compensating for NLOS error, which is generally much larger than systematic error. Therefore, this situation can seriously affect positioning accuracy.

[0035] The measured distance d of the i-th base station mi It can be expressed as:

[0036] d mi =d ti +e i +n i , i=1,2,3,…,N b

[0037] Among them, d ti represents the actual distance between the i-th base station and the target; e i represents the system error of the i-th base station, which follows a Gaussian random variable with a mean of 0; n irepresents the NLOS error of the i-th base station and is assumed to obey 0.05d ti with 0.9d ti The independent and identically distributed between .

[0038] The present invention provides a deep learning-based ship positioning method for coastal scenarios, and the specific steps are as follows:

[0039] Step 1: Eliminate systematic errors in distance measurement by each base station. Systematic errors primarily include laser emission delay error and detection circuit delay error. Because LiDAR uses a photoelectric-to-electro-optical conversion method for ranging, there is a delay between the electrical signal driving the laser and when it activates the laser and when it emits light. This delay is related to the characteristics of the laser itself and constitutes a linear error. Therefore, systematic errors can be adjusted through bias correction to eliminate the laser emission delay error component. Laser emission delay error is the primary form of systematic error, and eliminating it can be considered to eliminate systematic errors.

[0040] The present invention adopts N b The base stations measure the distance to the test target and use the least squares method to correct the bias. That is, the least squares method is used to fit the distance to calculate the slope k and intercept b that minimize the system error, and the measured distance d of each base station after eliminating the system error is obtained. ′ mi , to improve the ranging accuracy; the least squares fitting formula is as follows:

[0041]

[0042] Among them, k represents the slope of the line after the least squares fitting, b represents the intercept of the line after the least squares fitting, and d ti represents the actual distance between the i-th base station and the test target, d mi N represents the measured distance of the i-th base station before eliminating the system error, b Indicates the number of base stations.

[0043] Step 2: Group positioning. The positioning model in the coastal positioning scenario is as follows: Figure 1 As shown. Assuming that the total number of base stations is greater than the minimum number required for positioning, multiple groups of positioning base stations can be obtained in a combined manner. In a two-dimensional positioning scenario, at least three base stations are required to estimate the position of the target. Assuming that the minimum number of base stations required is m and the total number of base stations is N b , we get different base station combinations. Assume S gi Represents a combination of base stations with the number gi, gi∈{N b ,N b -1…m}. With N bTaking m=4 and m=3 as an example, there are 5 available base station ranging value combinations, such as the selected base station combination S4=(1,2,3,4), or the base station combination S3=[(2,3,4), (1,2,3), (1,2,4), (1,2,3)].

[0044] The present invention measures the distance d after eliminating the system error of each base station in each base station combination. ′ mi , the PSO-based three-sided positioning method is used to preliminarily estimate the position of the test target, and at this time m=3. The PSO-based three-sided positioning method is a distance-based wireless positioning algorithm. Assuming that there are three non-collinear base stations with known coordinates in space, three circles can be created with the base station coordinates as the center and the measured distance as the radius. The intersection of the circles is the positioning coordinates of the target. However, due to various errors in most real-life scenarios, the three circles will not intersect at a specific point. Therefore, when the present invention preliminarily estimates the target position based on the PSO-based three-sided positioning method, the particle swarm algorithm is introduced to find the optimal positioning position of the test target determined by each base station combination. The basic principle of the particle swarm algorithm is to utilize the information sharing of individuals in the group, so that the entire group produces an evolutionary process from disorder to order in the problem-solving space, and obtains the optimal solution of the group. The particle swarm algorithm mainly focuses on the position and velocity of particles. The present invention takes the test target position as the optimization target and uses X pj =(x pj ,y pj ) and V pj =(v pj,x ,v pj,y ) to represent the position and velocity of the jth particle.

[0045] The main steps of the particle swarm algorithm are as follows:

[0046] Step 1: Initialize the number of particles to N p The particle group, the speed of each particle V pj =(v pj ,v pj ) and position X pj =(x pj ,y pj ) takes a random value during initialization.

[0047] Step 2: Establish a fitness function to determine the fitness value of each iteration of the particles in the particle swarm. The fitness function of the jth particle is defined as follows:

[0048]

[0049] in, n in {m,m+1,…,N b}, the measured distance expression of the p-th base station after eliminating the system error is:

[0050]

[0051] Where (x ms ,y ms ) represents the test target position coordinates, (x bsp ,y bsp ), p∈(1,2,…,s) represents the location coordinates of the p-th base station, n p represents the NLOS error of the p-th base station.

[0052] Step 3: The particle swarm tracks the optimal value X of the particle itself during the iteration process prej =(x prej ,y prej ) and the particle swarm global optimal value X g =(x g ,y g ), if the current particle swarm global optimal value X g If the calculated fitness value is less than the preset error value, the test ends and the optimal positioning position (x esp ,y esp ), p∈{1,2,3,…,s}, otherwise, it is determined whether the current fitness value of each particle in the particle swarm is better than the optimal value X recorded in the last iteration of the particle. prej The obtained fitness value is better, then the optimal value X of the particle is updated. prej , and judge whether the current fitness value of each particle in the particle swarm is better than the global optimal value X of the particle swarm recorded in the last iteration g The obtained fitness value is better, then the global optimal value of the particle swarm is updated. g .

[0053] Among them, the speed and position update rules of the jth particle are as follows:

[0054]

[0055] Where t represents the number of iterations, c1 and c2 represent learning factors, and r1 and r2 are random probability values ​​between [0,1].

[0056] Step 3: Construct the dataset of Densenet model. b In the scenario of the positioning system, after group positioning, we can get an N pos ×2 matrix:

[0057]

[0058] in

[0059] The coordinate matrix M is converted into a two-dimensional pixel matrix through a plane rectangular coordinate system, and the two-dimensional pixel matrix is ​​used as the data set of the Densenet model. A feature image can be generated based on the two-dimensional pixel matrix. Each pixel in the feature image has the same size and different colors. The color of the pixel is used to distinguish the base station position coordinates, the optimal positioning position of the test target determined by the base station combination, and the test target position coordinates. For example, the base station position coordinates are black, RGB = (255, 255, 255), and the test target position coordinates are red, RGB = (255, 0, 0). Figure 2 shown.

[0060] Step 4. Train the Densenet model using the dataset constructed in step 3, and use the trained Densenet model to estimate the position of the target ship to obtain a more accurate positioning of the target ship. Furthermore, a part of the dataset constructed in step 3 can be selected as a training set and another part as a validation set to train and validate the Densenet model; during the training phase, a large amount of simulation data in different scenarios is used to train the model. DenseNet is a classic DRSN network model. Its main advantage is that it can effectively process data containing noise, and can establish jump connections between the front and back layers, so that the entire neural network is deep enough to avoid the gradient vanishing problem caused by the network layer being too deep, and then it can extract enough features to effectively process feature image data containing noise data. Due to the N b The environment in which LOS base stations and NLOS base stations are mixed is analyzed in this invention. b Some combinations of base stations do not contain NLOS ranging values ​​or contain fewer NLOS ranging values ​​than other combinations. The DenseNet model can assign a larger weight to the base station measurement distances in these combinations that are less affected by NLOS, thereby effectively reducing the impact of NLOS on the target ship positioning.

[0061] The neural network structure design of the DenseNet model is as follows Figure 3 As shown in the figure, the model consists of an input layer, a convolutional layer, three dense blocks (DenseBlock), two transition layers (TransitionLayer), a pooling layer, a linear connection layer, and an output layer; the input layer takes a two-dimensional pixel matrix as input, and the output layer outputs the final estimated position of the target ship; the main function of the convolutional layer is to generate feature maps, and the pooling layer is to alleviate the convolutional layer's excessive sensitivity to position. The core of the network structure is the dense block, Figure 3The black box in the dense block represents a convolution layer. The output data of each layer maintains the same data size and can be connected in the channel dimension. The input of each layer is composed of the outputs of all previous convolution layers. The nonlinear combination function in the dense block is a combination operation, which uses the structure of normalization + rectified linear unit + 3*3 convolution kernel to make the extracted data more complete by reusing information. Among them, the output X of the qth layer in the dense block is q The expression is as follows:

[0062] X q =H([X1,X2,…,X q-1 ])

[0063] Where H represents the nonlinear combination function, X1, X2,…, X q-1 are the outputs of each convolutional layer before the qth layer respectively.

[0064] The transition layer is placed between two Dense Blocks because the number of output channels after each Dense Block is large and a 1*1 convolution kernel is needed to reduce the dimension.

Claims

1. A deep learning-based ship positioning method for coastal scenarios, characterized by: The specific steps of this method are as follows: Step 1: Use N b Each base station measures the distance to the test target, calculates the slope k and intercept b that minimize the system error, and obtains the measured distance d′ after eliminating the system error of each base station mi ; The least squares fitting formula is as follows: Among them, k represents the slope of the line after the least squares fitting, b represents the intercept of the line after the least squares fitting, and d ti represents the actual distance between the i-th base station and the test target, d mi represents the measured distance of the i-th base station before eliminating the system error; Step 2: Let the minimum number of base stations required be m. Since the total number of base stations is N b , then the minimum number of base stations required is obtained different base station combinations; then, for each base station combination, the measured distance d′ after eliminating the system error mi , a three-sided positioning method based on PSO is used, and the particle swarm algorithm is introduced to find the optimal positioning position of the test target determined by each base station combination; Step 3: Obtain an N according to the position coordinates of each base station, the optimal positioning position of the test target determined for each base station combination in step 2, and the test target position coordinates. pos ×2 coordinate matrix: in The optimal positioning position of the test target determined for the p-th base station combination; n in {m,m+1,…,N b }; (x ms ,y ms ) is the test target position coordinate; (x bsi ,y bsi ) is the location coordinate of the i-th base station; The coordinate matrix M is converted into a two-dimensional pixel matrix through a plane rectangular coordinate system, and the two-dimensional pixel matrix is ​​used as the data set of the Densenet model; Step 4: Train the DenseNet model using the dataset constructed in step 3, and use the trained DenseNet model to estimate the position of the target ship.

2. The deep learning-based ship positioning method for coastal scenarios according to claim 1, characterized in that: The steps of the particle swarm algorithm are as follows: Step 1: Take the test target position as the optimization target, use X pj =(x pj ,y pj ) and V pj =(v pj ,v pj ) to represent the position and velocity of the jth particle, initialize the particle swarm, and the velocity of each particle V pj =(v pj ,v pj ) and position X pj =(x pj ,y pj ) takes a random value at initialization; Step 2: Establish a fitness function to determine the fitness value of each iteration of the particles in the particle swarm. The fitness function of the jth particle is defined as follows: in, n in {m,m+1,…,N b }, the measured distance expression of the p-th base station after eliminating the system error is: Where (x ms ,y ms ) represents the test target position coordinates, (x bsp ,y bsp ), p∈(1,2,…,s) represents the location coordinates of the p-th base station, n p represents the NLOS error of the p-th base station; Step 3: The particle swarm tracks the optimal value X of the particle during the iteration process prej =(x prej ,y prej ) and the particle swarm global optimal value X g =(x g ,y g ), if according to the current particle swarm global optimal value X g If the calculated fitness value is less than the preset error value, the test ends and the optimal positioning position (x esp ,y esp ), p∈{1,2,3,…,s}, otherwise, it is determined whether the current fitness value of each particle in the particle swarm is better than the optimal value X recorded in the last iteration of the particle. prej The obtained fitness value is better, then the optimal value X of the particle is updated. prej , and judge whether the current fitness value of each particle in the particle swarm is better than the global optimal value X of the particle swarm recorded in the last iteration g The obtained fitness value is better, then the global optimal value of the particle swarm is updated. g .

3. The deep learning-based ship positioning method for coastal scenarios according to claim 2, characterized in that: The speed and position update rules of the jth particle are as follows: Where t represents the number of iterations, c1 and c2 represent learning factors, and r1 and r2 are random probability values ​​between [0,1].

4. The deep learning-based ship positioning method for coastal scenarios according to claim 1, 2 or 3, characterized in that: A part of the dataset constructed in step 3 is selected as the training set and the other part is selected as the validation set to train and validate the Densenet model.

Images