A multi-source vehicle speed fusion method based on hunting algorithm optimized radial basis function neural network

Abstract

The invention discloses a multi-source vehicle speed fusion method based on a hunting algorithm to optimize a radial basis neural network, which includes: obtaining the location vehicle speed of a certain section of a target road section through three location vehicle speed collection methods, and dividing the location vehicle speed into a fusion group and a control group; clustering the fusion group using K-means clustering, and determining the optimal cluster center number K by the elbow method; using the optimal cluster center number K as the number of neurons of the radial basis neural network; constructing the radial basis neural network using a Gaussian function as a kernel function according to the number of neurons; using the fusion group as the input of the radial basis neural network, comparing the output of the neural network with the corresponding control group, and using the comparison result as a fitness function; (S5) using a hunting algorithm, the fitness function optimizes the parameters of the radial basis neural network to obtain the parameters of the radial basis neural network, and using the radial basis neural network to fuse multi-source vehicle speeds. The method of the invention can accurately and efficiently fuse the location vehicle speeds.

Description

Technical Field

[0001] The present invention relates to the field of traffic information and control, and in particular to a multi-source vehicle speed fusion method based on hunting algorithm optimization of radial basis neural network. Background Art

[0002] With the development of economy, the number of motor vehicles in cities has increased steadily, and urban road traffic control has become increasingly complex. Obtaining accurate traffic data is the basis for urban traffic control. Vehicle speed includes location speed, driving speed, and trip speed. The location speed represents the congestion degree of the target road section. At the same time, the change of speed indicates the change of parameters such as future traffic flow and density of the road section. Therefore, obtaining accurate location speed is of great significance for estimating the traffic conditions of the target road section and adjusting the timing of traffic signals. At present, more and more new technologies are applied to the measurement of location speed, such as manual measurement method, road detector method, radar speed measurement method and electro-optical tube measurement method. The rise of intelligent networked vehicles also makes it more convenient to directly obtain the location speed of a single vehicle. Since the speed obtained by each measurement method has a certain error and the speed data collected by the intelligent networked vehicle can only cover a part of the vehicles, how to integrate the speeds from various sources to obtain a more accurate speed measurement result is an urgent problem to be solved.

[0003] Classical evolutionary optimization algorithms have strong optimization capabilities, such as ant colony algorithms and particle swarm algorithms. However, they are prone to falling into local optimality during the calculation process. At the same time, some particles participating in the calculation tend to maintain the current state after several rounds of iterations, thereby reducing the calculation speed and causing a waste of computing resources. The hunting algorithm can effectively solve this problem. As a type of neural network, the radial basis neural network has only one intermediate layer compared to traditional deep neural networks such as BP neural networks. The intermediate layer contains several neuron structures. The neuron kernel function is a radial basis function of different forms. The neuron kernel function only responds to stimuli in a specific range. Therefore, the radial basis neural network can approximate nonlinear functions of arbitrary precision and is suitable for multi-source data fusion. At the same time, because there is only one intermediate neuron, the radial basis neural network converges faster. However, in the process of neural network training, the determination of the parameters of each neuron (center position and width) and the determination of the weight of each neuron output are problems that need to be solved. In response to this problem, the particle swarm algorithm is used to optimize the radial basis neural network to determine the parameters of the neural network. At the same time, this method is used to fuse the vehicle speeds at different locations obtained by different sensors to improve data accuracy. Summary of the invention

[0004] In view of the above defects of the prior art, the purpose of the present invention is to provide a multi-source vehicle speed fusion method based on a hunting algorithm to optimize a radial basis function neural network. The hunting algorithm is used to solve the parameters in the radial basis function neural network used for multi-source vehicle speed fusion.

[0005] The above technical objectives of the present invention are achieved through the following technical solutions:

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A multi-source vehicle speed fusion method based on hunting algorithm optimization radial basis neural network, comprising:

[0008] (S1) obtaining a location speed of a certain section of a target road segment by at least three location speed collection methods, and dividing the location speeds obtained by the multiple methods into a fusion group and a control group, wherein the fusion group includes the location speeds obtained by at least two methods, and the control group includes the location speeds obtained by at least one method;

[0009] (S2) clustering the fusion group using K-means clustering, and determining the optimal number of cluster centers K using the elbow method; obtaining the coordinates of each cluster center under the optimal number of cluster centers K, and using the optimal number of cluster centers K as the number of neurons in the radial basis neural network;

[0010] (S3) constructing a radial basis neural network using a Gaussian function as a kernel function according to the number of neurons obtained in step S2, and determining the number of parameters that need to be optimized in the radial basis neural network;

[0011] (S4) using the fusion group as the input of the radial basis function neural network, comparing the output of the radial basis function neural network with the corresponding control group, and using the comparison result as the fitness function;

[0012] (S5) Using a hunting algorithm, the parameters of the radial basis function determined in step S3 are optimized according to the fitness function obtained in step S4 to obtain the parameters of the radial basis function, and the radial basis function is used to perform multi-source vehicle speed fusion.

[0013] A further improvement of the present invention is that the input data is normalized before being input into the radial basis function neural network.

[0014] A further improvement of the present invention is that after the output data of the radial basis neural network is obtained, denormalization is performed according to the corresponding input data.

[0015] A further improvement of the present invention is that determining the optimal number of cluster centers K and the coordinates of each cluster center by the elbow method in step S2 specifically includes:

[0016] (S21) Perform K-means clustering on the fusion group according to the number of cluster centers, and record the sum of the distances C between each sample and the nearest cluster center under the number of cluster centers. k ;

[0017] (S22) Increasing the number of cluster centers from small to large within the range of the number of cluster centers, and executing step S21 for each number of cluster centers to obtain the corresponding C k ; Plot the number of cluster centers and C k The curve between , takes the number of cluster centers at the inflection point of the curve as the optimal number of cluster centers K.

[0018] A further improvement of the present invention is that the expression of the Gaussian function in step S3 is: The coordinates of the kernel function of each neuron i are denoted by W i , width b i , each neuron i also has a corresponding weight w i ; The coordinates are denoted as W i and width b i The dimension n is the same as the data dimension of the fusion group.

[0019] A further improvement of the present invention is that step S5 specifically includes:

[0020] Determine the dimension of the particle coordinates in the hunting algorithm. The coordinate dimension is the number of parameters of the radial basis neural network in step S3 plus 1, that is, add a dimension to each particle i, denoted as k i , k i In each iteration, it is set to a random number between [0,1];

[0021] Perform several iterations, and in each iteration process, grasp the advantages and disadvantages of the positions of each particle in this iteration according to the fitness function in S4; the optimal position of a particle i in the tth round of iteration is recorded as p it-best , the optimal position among all particles is denoted as p gt-best , and finally get the global optimal position p g This is the optimal solution; then update the speed and position coordinates of each particle according to the hunting algorithm formula; then check the updated speed of each particle;

[0022] If the speed is less than a certain speed threshold E, start checking the parameter k of particle i i The value of k is set to the threshold K1. iWhen K1, "hunting" is initiated for the particle; if the particle is captured, the particle is deleted; if the particle escapes capture, the particle performs self-update, and during the update process, the particle resets its own velocity, which is set as a random number, and the position randomly changes to the optimal position p of an individual in the particle's iteration history it-be ;

[0023] Set a threshold K2 to determine whether the particle is captured during the "hunting" process. When k i < K2, the particle is successfully captured; otherwise, the particle escapes capture and self-update is initiated;

[0024] Every several rounds of iteration, when the accumulated number of deleted particles reaches the first threshold, new particles with the same number as the deleted particles are randomly generated to supplement the particle swarm quantity and ensure the calculation efficiency.

[0025] The beneficial effects of the present invention are as follows: The present invention combines the theoretical framework of general multi-source data fusion, utilizes the existing data on the road section, and adopts the hunting algorithm to solve the optimization parameters of the radial basis neural network. This method can enable the neural network to fit the specific law of the vehicle speed on the target road section as much as possible, realizing a comprehensive and non-linear mapping from various collected data to accurate data, and avoiding complex formula derivation and large computational volume problems.

[0026] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent description, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be learned from the practice of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0028] Figure 1 is the flow chart of the present invention.

[0029] Figure 2 is the structure diagram of the radial basis neural network.

[0030] Figure 3 is the flow chart of the hunting algorithm. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0032] In the present invention, firstly, the "elbow method" and "K-means" clustering are used to determine the parameters of the intermediate layer of the radial basis neural network. The data distribution characteristics of different samples are different, and the number of neurons in the radial basis neural network needs to be set in a targeted manner.

[0033] Then, during the training process, a batch of particles with coordinates and speeds in the hunting algorithm are randomly generated, the coordinates of the particles are brought into the neural network, and the speed samples collected by methods A and B on the target road section are used as the input of the neural network. After the neural network operation, the fused speed value is obtained, and the speed collected by method C is regarded as the real speed value. The fused speed value and the real speed value are compared to use this as the evaluation function of the particle coordinates. This is also the "fitness function" in the particle evolution process.

[0034] Next, the particle updates its position and velocity according to its historical optimum and global optimum in the evolution process. Different from the classical particle swarm algorithm, the control of some particles is added in each round of iteration. The specific operation is: add a dimension to each particle i, with values ​​of k and i In the process of traversing all particles in this round of iteration, operations on the newly added dimensions are added, "particles with poor running speed" are eliminated, and "particles with relatively poor running speed" are reset to ensure the excellent performance of the particle group. This speeds up the convergence speed and gradually converges to the global optimum.

[0035] Finally, the optimized neural network is applied to the fusion of vehicle speeds on the target road section.

[0036] See also Figure 1 , Figure 2 , Figure 3 As shown, a method for solving a multi-source vehicle speed fusion method based on a hunting algorithm to optimize a radial basis function neural network of the present invention comprises the following steps:

[0037] S1: When the ratio V / C of the road traffic volume (V) to the road capacity (C) is large, obtain the location speed of a certain section of the target road section through at least three location speed collection methods. The location speeds obtained by multiple methods are divided into a fusion group and a control group, wherein the fusion group includes the location speeds obtained by at least two methods, and the control group includes the location speeds obtained by at least one method.

[0038] In this embodiment, the location speed is obtained by three different methods (method A, method B, and method C), and the location speed data obtained each time is a three-element array. In the process of processing each three-element array, the location speed data obtained by method A and method B are classified into the fusion group as a two-element array, and the location speed obtained by method C is classified into the control group. The final fusion group is a two-dimensional point set, and the control group is a one-dimensional point set.

[0039] S2: Use K-means clustering to cluster the fusion group, and determine the optimal number of cluster centers K by the elbow method; obtain the coordinates of each cluster center under the optimal number of cluster centers K, and use the optimal number of cluster centers K as the number of neurons in the radial basis neural network. This step specifically includes:

[0040] (S21) Perform K-means clustering on the fusion group according to the number of cluster centers, and record the sum of the distances C between each sample and the nearest cluster center under the number of cluster centers. k ;

[0041] (S22) Increasing the number of cluster centers from small to large within the range of the number of cluster centers, and executing step S21 for each number of cluster centers to obtain the corresponding C k ; Plot the number of cluster centers and C k The curve between , takes the number of cluster centers at the inflection point of the curve as the optimal number of cluster centers K. This step is the elbow method. By adopting the elbow method, the subsequent radial basis neural network can have better pertinence and adaptability to the specific fusion group data.

[0042] The "K-means" clustering method is an existing technology, which specifically includes:

[0043] ① Assume that the number of cluster centers K = 2, and randomly select the center coordinates;

[0044] ② Traverse each sample point, calculate its distance to the two cluster centers respectively, and classify this point into the category with the closest Euclidean distance until this round of classification is completed;

[0045] ③After classification, find the mean of the horizontal and vertical coordinates of all sample points in each category to form a new center point;

[0046] ④ Repeat steps 2 and 3 until the sum of all distances of all sample points from the latest cluster center no longer decreases;

[0047] The “elbow method” is also a prior art, which specifically includes:

[0048] ① Assuming that the number of cluster centers K = 2, after performing the "K-means" clustering operation, record the sum of all distances of all sample points from the latest cluster center C2;

[0049] ② Increase the number of assumed cluster centers, that is, set the number of cluster centers K = 3, 4, 5, etc., repeat step 1, and then record the sum of all distances C of all sample points from the latest cluster center. k ;

[0050] ③The number of cluster centers K and the corresponding C k Draw a curve graph, and the k value corresponding to the obvious inflection point of the curve graph is the optimal K value.

[0051] (S3) According to the number of neurons obtained in step S2, a radial basis neural network is constructed using a Gaussian function as a kernel function, and the number of parameters to be optimized for the radial basis neural network is determined. In this embodiment, the structure of the radial basis neural network is as follows: Figure 2 As shown, the expression of Gaussian function is Among them, the coordinates Wi and width b of the kernel function i of each neuron i In this embodiment, i is the number of the neuron; i , width b i The dimension of is the same as the data dimension in the fusion group, and in this embodiment, the dimension is 2. Since each neuron has its own weight corresponding to the output value, it is also necessary to determine the weight w of each neuron i , weight w i The number of parameters that the neural network needs to optimize is N=K*(n+1+1), where n is the data dimension in the fusion group, and n is 2 in this embodiment.

[0052] In order to improve the performance of the radial basis neural network, each input data (two-dimensional in this embodiment) is normalized before being input into the radial basis neural network; and the output data of the radial basis neural network is denormalized in combination with the corresponding input data.

[0053] (S4) Using the fusion group as the input of the radial basis neural network, the output of the radial basis neural network is compared with the corresponding control group, and the comparison result is used as the output of the fitness function. In step S5, a hunting algorithm is used to obtain the various parameters of the radial basis neural network, and the above fitness function is used to evaluate the iterative process of the hunting algorithm. In some embodiments, the comparison of the fitness function refers to obtaining the Euclidean distance between the output generated by the radial basis neural network for a certain fusion group data input and the control group data corresponding to the fusion group data. When the output of the radial basis neural network and the control group data are both one-dimensional, obtaining the Euclidean distance is equivalent to subtracting them.

[0054] (S5) Using a hunting algorithm, the parameters of the radial basis function neural network determined in step S3 are optimized according to the fitness function obtained in step S4, so as to obtain a final radial basis function neural network for multi-source vehicle speed fusion. The specific process of the hunting algorithm is introduced below.

[0055] In the hunting algorithm, an N-dimensional space is constructed for target search. A particle swarm consisting of M particles is initialized in this space. In order to improve the computational efficiency, the particles are given a motion range [-x min ,x max ] (This range is obtained based on the coordinates of each cluster center under the optimal cluster center number K in step S2. Its dimension is the same as that of the particle. The one-dimensional representation is used here only for intuitiveness); particle i is recorded as X i =[ki,xi1,xi2,...,xi N ], whose velocity vector is denoted by V i =[k i ,vi1,vi2,...,vi N ].

[0056] The particle continuously updates its own position and current speed: each round of iteration needs to record the optimal position of the particle, denoted as p it-b =[k i ,pi1,pi2,...,pi N ], each round of iteration also needs to record the optimal position of all particles, denoted as p gt-best =[k i ,gi1,gi2,...,gi N ], it should be noted that k i The value of is randomly generated at the beginning of each iteration and is only used as a marker to determine whether the particle is captured in the next stage. It does not participate in the update here.

[0057] Each calculation will be X i (except k i) Substitute into the radial basis neural network to fill in each parameter. Use all the integrated combined data as the input, calculate the objective function using the difference between the output and the control group, and then iterate the particles according to the step size t using the velocity and coordinate iteration formulas of the i-th particle. The formulas are as follows:

[0058] Particle velocity iteration formula:

[0059]

[0060] Particle coordinate update formula:

[0061]

[0062] Among them, rand() is a random number generation function that generates random numbers between (0, 1); δ has only two cases of 0 and 1. ① If the velocity of particle i is greater than E, δ takes 1. ② If the velocity of particle i is less than E and K1 < k i When, δ takes 1. ③ If the velocity of particle i is less than E and K2 < k i < K1, δ takes 0. ④ If the velocity of particle i is less than E and k i < K2, delete the particle; represents a random number, represents the historical best position p of this particle it-best ; x i (t + 1), x i (t) represent the position coordinates of the i-th particle at times t + 1 and t in space; v i (t + 1), v i (t) represent the velocities of the i-th particle at times t + 1 and t in space; w is the inertia weight, which represents the inheritance of the motion state of each particle in the previous moment. The larger the value, the stronger the global optimization ability, and the smaller the value, the stronger the local optimization ability. In this paper, since the feasible region is small, w takes 0.1; c1 and c2 are the group learning rate and the individual learning rate, which represent the degree of importance of the particle to the optimal position it has passed through itself and the optimal position passed through by the group in the previous moment. It is generally recommended to take 0.2; r1 and r2 are random numbers between 0 and 1, which increase the randomness of the velocity change during the particle movement process.

[0063] In the specific implementation process, step S5 includes the following steps:

[0064] Determine the dimension of the particle coordinates in the hunting algorithm. The coordinate dimension is the number of parameters of the radial basis neural network in step S3 plus 1, that is, add a dimension to each particle i, denoted as k i , k i is set as a random number between [0, 1] in each round of iteration;

[0065] Perform several iterations. During each iteration, determine the quality of the positions of each particle according to the fitness function in S4. The individual best position of a certain particle i at the t-th iteration is denoted as p it-best , and the best position among all particles is denoted as p gt-best , and finally obtain the global best position p g , which is the optimal solution. Then update the velocity and position coordinates of each particle according to the hunting algorithm formula. Immediately afterwards, check the updated velocity of each particle;

[0066] If the velocity is less than a certain velocity threshold E, start checking the parameter k i of particle i, set the threshold K1. When k i <K1, initiate "hunting" for the particle. If the particle is captured, delete the particle. If the particle escapes capture, perform self-update on the particle. During the update process, the particle resets its own velocity, which is set to a random number, and the position randomly changes to an individual best position p it-be in the iteration history of the particle;

[0067] Set the threshold K2 to determine whether the particle is captured during the "hunting" process. When k i <K2, the particle is successfully captured; otherwise, the particle escapes capture and initiates self-update;

[0068] Every few iterations, when the accumulated number of deleted particles reaches the first threshold, the total number of the particle swarm becomes smaller, and new particles with the same number as the deleted particles need to be randomly generated to supplement the number of the particle swarm and ensure the calculation efficiency.

[0069] Through the above process, the optimal solutions of the various parameters of the radial basis neural network can be obtained. Substituting the optimal solutions into the radial basis neural network can obtain the radial basis neural network for multi-source vehicle speed fusion. In this embodiment, before inputting the two location vehicle speeds to be fused into the radial basis neural network, it is necessary to normalize them; after corresponding denormalization of the output of the radial basis neural network, the result of multi-source vehicle speed fusion is obtained, and this result can be used to guide traffic control.

[0070] The above embodiments only illustrate the principles and effects of the present invention by way of example, and are not used to limit the present invention. Any person familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A multi-source vehicle speed fusion method based on hunting algorithm optimized radial basis neural network, characterized in that The following steps are involved: (S1) obtaining a location speed of a certain section of a target road segment by at least three location speed collection methods, and dividing the location speeds obtained by the multiple methods into a fusion group and a control group, wherein the fusion group includes the location speeds obtained by at least two methods, and the control group includes the location speeds obtained by at least one method; (S2) clustering the fusion group using K-means clustering, and determining the optimal number of cluster centers K using the elbow method; obtaining the coordinates of each cluster center under the optimal number of cluster centers K, and using the optimal number of cluster centers K as the number of neurons in the radial basis neural network; (S3) constructing a radial basis neural network using a Gaussian function as a kernel function according to the number of neurons obtained in step S2, and determining the number of parameters that need to be optimized in the radial basis neural network; (S4) using the fusion group as the input of the radial basis function neural network, comparing the output of the radial basis function neural network with the corresponding control group, and using the comparison result as the fitness function; (S5) Using a hunting algorithm, the parameters of the radial basis function determined in step S3 are optimized according to the fitness function obtained in step S4 to obtain the parameters of the radial basis function, and the radial basis function is used to perform multi-source vehicle speed fusion.

2. According to claim 1, a multi-source vehicle speed fusion method based on hunting algorithm optimized radial basis neural network is characterized in that: Before the data is input into the radial basis neural network, the input data is normalized.

3. The multi-source vehicle speed fusion method based on hunting algorithm optimized radial basis neural network according to claim 2 is characterized in that: After the output data of the radial basis neural network is obtained, denormalization is performed according to the corresponding input data.

4. The multi-source vehicle speed fusion method based on hunting algorithm optimized radial basis neural network according to claim 1 is characterized in that: In step S2, the optimal number of cluster centers K and the coordinates of each cluster center are determined by the elbow method, specifically including: (S21) Perform K-means clustering on the fusion group according to the number of cluster centers, and record the sum of the distances C between each sample and the nearest cluster center under the number of cluster centers. k ; (S22) Increasing the number of cluster centers from small to large within the range of the number of cluster centers, and executing step S21 for each number of cluster centers to obtain the corresponding C k ; Plot the number of cluster centers and C k The curve between , takes the number of cluster centers at the inflection point of the curve as the optimal number of cluster centers K.

5. The multi-source vehicle speed fusion method based on hunting algorithm optimized radial basis neural network according to claim 1 is characterized in that: The expression of the Gaussian function in step S3 is: The coordinates of the kernel function of each neuron i are denoted by W i , width b i , each neuron i also has a corresponding weight w i ; The coordinates are denoted as W i And the dimension n of width bi is the same as the data dimension of the fusion group.

6. The multi-source vehicle speed fusion method based on hunting algorithm optimized radial basis neural network according to claim 1 is characterized in that: Step S5 specifically includes: Determine the dimension of the particle coordinates in the hunting algorithm. The coordinate dimension is the number of parameters of the radial basis neural network in step S3 plus 1, that is, add a dimension to each particle i, denoted as k i , k i In each iteration, it is set to a random number between [0,1]; Perform several iterations, and in each iteration process, grasp the advantages and disadvantages of the positions of each particle in this iteration according to the fitness function in S4; the optimal position of a particle i in the tth round of iteration is recorded as p it-best , the optimal position among all particles is denoted as p gt-best , and finally get the global optimal position p g This is the optimal solution; then update the speed and position coordinates of each particle according to the hunting algorithm formula; then check the updated speed of each particle; If the speed is less than a certain speed threshold E, start checking the parameter k of particle i i value, set the threshold K1, when k i < K1, start "hunting" for the particle; if the particle is captured, delete the particle; if the particle escapes capture, perform self-update on the particle, and during the update process, the particle resets its own speed, which is set to a random number, and the position randomly changes to the individual optimal position p in the iteration history of the particle it-best ; Set a threshold K2 to determine whether a particle is captured during the "hunting" process. When k i < K2, the particle is successfully captured; otherwise, the particle escapes capture and initiates self-update; After several rounds of iterations, when the deleted particles accumulate to the first threshold, new particles with the same number of deleted particles are randomly generated to supplement the number of particle swarms.

Images