Underwater robot visual servo control method based on improved neural network

Abstract

The invention relates to the technical field of automatic control, in particular to an underwater robot visual servo control method based on an improved neural network. Comprising the steps of determining input and output of a neural network model according to the principle of an IBVS algorithm; taking the input and the output of the neural network model as a training sample set of the neural network model and performing preprocessing; optimizing a neural network model algorithm flow based on a pseudo-inverse weight direct determination method, and directly calculating a threshold-containing equivalent weight corresponding to a minimum error of the neural network model algorithm in a back propagation process; determining the number of optimal hidden layer neurons of the neural network model by using a dual-stage structure self-determination algorithm; and the trained neural network model is used to replace the operation of an image Jacobi matrix in an IBVS algorithm, and visual servo control of the underwater robot is realized. According to the invention, the precision and stability of visual servo control of the underwater robot can be improved.

Description

Technical Field

[0001] This invention relates to the field of automation control technology, and in particular to a visual servo control method for underwater robots based on an improved neural network. Background Technology

[0002] Image-Based Visual Servoing (IBVS) is a key technology in the field of automation. Its core function is to acquire real-time visual feedback information through the vision module of an autonomous system (such as an underwater robot) and dynamically adjust the system's motion state to achieve high-precision and highly adaptable task execution. It is currently widely used in practical scenarios such as fixed-point landing and target tracking of underwater robots. The IBVS algorithm uses the image Jacobian matrix as the core computing unit. Its working principle is as follows: First, the actual coordinates of feature points in the image are extracted and the feature point error is calculated by comparing them with the preset expected coordinates. Then, the image Jacobian matrix is ​​constructed based on the feature point depth information. Through this matrix, a mapping relationship is established between the "feature point error" and the "underwater robot's motion speed". Finally, a control signal is generated based on the mapping result to drive the robot to adjust its motion, forming a visual servoing control closed loop.

[0003] However, the IBVS algorithm faces inherent technical challenges in practical applications, making it difficult for its control performance to meet the needs of complex underwater scenarios. First, the algorithm is highly dependent on the depth information of feature points, but underwater environments have interference factors such as light attenuation and water turbidity, making it difficult to accurately obtain the depth information of feature points. This directly leads to insufficient accuracy in constructing the image Jacobian matrix, which in turn affects the accuracy of the control signal. Second, the IBVS algorithm is prone to problems such as local minima and singular interaction matrices, which can cause underwater robots to fall into a non-optimal control state and fail to stably and efficiently complete the preset servo control tasks. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a visual servo control method for underwater robots based on an improved neural network, in order to solve the problems of insufficient accuracy of the image Jacobian matrix caused by the difficulty in obtaining feature point depth information in the IBVS algorithm, and the easy occurrence of local minima and singular interaction matrix, which in turn affect the accuracy and stability of the visual servo control of underwater robots.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a visual servo control method for underwater robots based on an improved neural network, comprising the following steps:

[0006] S1. Based on the principle of the IBVS algorithm, determine the input and output of the neural network model;

[0007] S2. Use the input and output of the neural network model as the training sample set of the neural network model and perform preprocessing.

[0008] S3. The method of directly determining weights based on pseudo-inverses optimizes the algorithm flow of neural network models and directly calculates the equivalent weights with thresholds corresponding to the minimum error in the backpropagation process of neural network model algorithms.

[0009] S4. Based on the threshold equivalent weights corresponding to the minimum error, the optimal number of hidden layer neurons in the neural network model is determined using a two-stage self-determining algorithm.

[0010] S5. Use the trained neural network model to replace the image Jacobian matrix operation in the IBVS algorithm to realize visual servo control of the underwater robot.

[0011] Preferably, the input and output of the neural network model are specifically: the error vector between the actual coordinates and the expected coordinates of four image feature points. , , , The error vector As input to the neural network model, the underwater robot's motion velocity, specifically including angular velocity... Translation speed , as the output of the neural network model; where, , , and For actual coordinates, , , and These are the desired coordinates.

[0012] Preferably, the training sample set of the neural network model is ,in, The total number of samples, For the first The input vector of the group, For the first The group corresponds to the output vector. It is the input vector The number of dimensions, It is the dimension space of the input vector. It is the output The number of dimensions, It is the dimension space of the output vector.

[0013] Preferably, the preprocessing uses a deviation standardization method, and the specific formula for the deviation standardization method is as follows:

[0014]

[0015] in, The original data, This is the standardized data.

[0016] Preferably, the specific steps of the method for directly determining weights based on pseudo-inverse to optimize the neural network model algorithm are as follows: First, randomly generate hidden layer nodes, construct the hidden layer output matrix, and establish an objective function with the goal of minimizing the approximate squared difference between the neural network model output and the sample label; second, use generalized matrix theory to calculate the Moore-Penrose generalized pseudo-inverse matrix of the hidden layer output matrix to obtain the threshold-equivalent weights corresponding to the minimum error of the neural network model algorithm during backpropagation.

[0017] Preferably, the specific steps for determining the optimal number of hidden layer neurons in a neural network model using a two-stage self-determining algorithm are as follows: In the first stage of the search, multiple hidden layer neurons are added at each step, the training set data is imported into the neural network model for training, and the initial prediction accuracy of the model under different hidden layer neurons is compared to quickly pinpoint the approximate range where the optimal number of hidden layer neurons is located; In the second stage of the search, within the approximate range determined in the first stage, only one hidden layer neuron is added at each step, the model is trained again and its performance is evaluated, and finally the number of hidden layer neurons that makes the neural network model perform optimally is found.

[0018] The beneficial effects of this invention are as follows: Compared with the prior art, this invention replaces the image Jacobian matrix operation in the IBVS algorithm with a trained neural network model, avoiding dependence on feature point depth information, realizing calibration-free visual servo control, and improving the accuracy and stability of underwater robot visual servo control; it adopts a pseudo-inverse-based direct weight determination method to optimize the neural network model algorithm process, aiming to minimize the approximate squared difference between the neural network model output and the sample label to quickly solve for the optimal threshold equivalent weights, accelerating the training speed and avoiding local optima problems; it uses a two-stage self-determining algorithm to determine the optimal number of hidden layer neurons in the neural network model, further improving the model's accuracy and generalization ability, ultimately enabling underwater robot visual servo control to have higher control accuracy, response efficiency, and environmental adaptability in complex underwater environments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the process of the present invention;

[0021] Figure 2 This is a schematic diagram of the basic process of the neural network model of the present invention;

[0022] Figure 3 This is a schematic diagram illustrating the visual servo control principle based on a neural network model algorithm of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains.

[0024] like Figure 1 The diagram shown is a flowchart of the present invention. A visual servo control method for underwater robots based on an improved neural network includes:

[0025] S1. Based on the principle of the IBVS algorithm, determine the input and output of the neural network model.

[0026] First, based on the principle of the IBVS algorithm, in image coordinates, for a single feature point, if its actual coordinates are... The expected coordinates are Then the error vector of this feature point can be calculated as: .

[0027] This invention selects four image feature points and calculates the error vector between the actual coordinates and the expected coordinates of these four image feature points. , , , Composition of error vector , as input to the neural network model.

[0028] For the model's output, the underwater robot's motion velocity is selected, specifically including angular velocity. Translation speed As the output of the neural network model; the basic flowchart of the neural network model of this invention is shown below. Figure 2 As shown, the specific process is explained in detail through the following steps.

[0029] S2. Use the input and output of the neural network model as the training sample set of the neural network model and perform preprocessing.

[0030] First, a training sample set is constructed. Through simulation experiments of visual servo control, sample data corresponding to the error vector between the actual and desired coordinates of feature points and the motion speed of the underwater robot are collected to form the training sample set. ,in, The total number of samples, For the first The input vector of the group, For the first The group corresponds to the output vector. It is the input vector The number of dimensions, It is the dimension space of the input vector. It is the output The number of dimensions, It is the dimension space of the output vector.

[0031] Secondly, data preprocessing is performed. Before inputting the training sample data into the neural network model, to eliminate the influence of differences in the dimensions of different variables on the training of the neural network model, deviation standardization is used to preprocess the training sample data. The specific formula for deviation standardization is:

[0032]

[0033] in, For the first Group of original data, For the first The data is standardized. After this change, the data is compressed to... The interval is dimensionless, ensuring the stability and convergence of subsequent neural network model training.

[0034] S3. The method of directly determining weights based on pseudo-inverses optimizes the algorithm flow of neural network models and directly calculates the equivalent weights with thresholds corresponding to the minimum error in the backpropagation process of neural network model algorithms.

[0035] The algorithm for optimizing neural network models is based on a direct weight determination method using pseudo-inverses. The threshold is equivalent to the weights of hidden layer neurons with a constant input of 1. This allows for the direct calculation of the threshold-equivalent weights corresponding to the minimum error during backpropagation of the neural network model. The specific steps are as follows:

[0036] (1) Randomly generate hidden layer nodes and define the mapping relationship from hidden layer to output layer.

[0037] Randomly generate hidden layer nodes, the first The hidden layer neurons to the first The activation output of each output layer neuron is ,in, For the first Input sample, For the first The weight vector of each hidden layer neuron. For the first The threshold of each hidden layer neuron The activation function for the hidden and output layers is the Sigmoid activation function, which is used to introduce nonlinear transformations.

[0038] Based on the data from the training sample set, construct the hidden layer output matrix. :

[0039]

[0040] in, The number of neurons in the hidden layer is denoted by , and the matrix dimension is determined by the number of samples and the number of hidden layer nodes.

[0041] If the weight vector of the hidden layer output layer is , Each hidden layer neuron corresponds to a threshold-equivalent weight in the output layer; therefore, the vector of predicted data output by the output layer of the neural network model is... Compare the outputs of neural network models With sample labels The objective function is established by minimizing their approximate squared difference:

[0042]

[0043] in, These are the actual output values ​​from the training data, corresponding to the actual measured values ​​of the underwater robot's motion speed, i.e., the corresponding angular velocity. Translation speed The six data points.

[0044] (2) Solve for the optimal equivalent weights with threshold using generalized matrix theory.

[0045] By calculating the hidden layer output matrix Moore–Penrose generalized pseudoinverse matrix This directly yields the optimal solution for the weight vector:

[0046]

[0047] The formula for calculating the pseudo-inverse is:

[0048]

[0049] Under the current hidden layer neuron structure, this method can output the threshold equivalent weights corresponding to the minimum error of the neural network model algorithm during backpropagation, thus completing the training of a neural network model.

[0050] S4. Based on the threshold equivalent weights corresponding to the minimum error, the optimal number of hidden layer neurons in the neural network model is determined using a two-stage self-determining algorithm.

[0051] Based on the threshold-equivalent weights corresponding to the minimum error obtained in step S3, the optimal number of hidden layer neurons in the neural network model is determined using a two-stage self-determining algorithm. The specific steps are as follows:

[0052] The first stage involves quickly locating the optimal interval. During the first stage search, multiple hidden layer neurons are added at each step. The training sample set data is imported into the neural network model for training. The initial prediction accuracy of the model under different hidden layer neurons is compared to quickly pinpoint the approximate interval where the optimal number of hidden layer neurons is located.

[0053] The second stage involves a meticulous search for the optimal number of hidden layer neurons. During this search, within the approximate range determined in the first stage, only one hidden layer neuron is added at each step. The model is then retrained and its performance is evaluated, including prediction error and convergence speed. Ultimately, the optimal number of hidden layer neurons that maximizes the performance of the neural network model is found, thus determining the optimal structure of the neural network model.

[0054] S5. Use the trained neural network model to replace the image Jacobian matrix operation in the IBVS algorithm to realize visual servo control of the underwater robot.

[0055] like Figure 3 The diagram illustrates the flow chart of the visual servo control principle based on a neural network model algorithm of this invention. The specific steps are as follows: First, acquire the camera image of the current state to obtain visual information about the underwater robot's current environment and its own position; second, perform image processing and feature point coordinate extraction to obtain the actual coordinates of the feature points at the current moment; then, compare the actual coordinates with the expected coordinates of the feature points in the image coordinate system to obtain the feature point error; finally, input the feature point error into the neural network-based visual servo controller, which outputs corresponding control signals to the underwater robot's motion controller to drive the underwater robot's position update. This method uses the optimal structure of the trained neural network model to replace the image Jacobian matrix calculation in the IBVS algorithm, directly establishing a mapping relationship between the actual coordinates of the feature points, the expected coordinate error, and the underwater robot's motion speed, thereby achieving visual servo control of the underwater robot.

[0056] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0057] This invention aims to cover all such substitutions, modifications, and variations that fall within the scope of protection. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. An improved neural network-based visual servoing control method for an underwater robot, characterized in that, The method comprises the following steps: S1, determining the input and output of a neural network model according to the principle of an IBVS algorithm; S2, taking the input and output of the neural network model as a training sample set of the neural network model and performing preprocessing; S3, optimizing the algorithm process of the neural network model based on a weight direct determination method based on pseudo-inverse, and directly calculating the equivalent weights with thresholds corresponding to the minimum error in the back propagation process of the neural network model algorithm; S4, determining the optimal number of hidden layer neurons of the neural network model using a two-stage structure self-determination algorithm based on the equivalent weights with thresholds corresponding to the minimum error; S5, replacing the operation of an image Jacobian matrix in the IBVS algorithm with the trained neural network model to realize visual servo control of the underwater robot.

2. The method of claim 1, wherein, The input and output of the neural network model are specifically: taking an error value vector of four image feature point actual coordinates and expected coordinates , , , composed of the error vector as the input of the neural network model; taking the motion speed of the underwater robot, specifically including angular velocity and translation speed as the output of the neural network model; wherein, , , and are actual coordinates, , , and are expected coordinates.

3. The method of claim 1, wherein, The training sample set of the neural network model is ,in, The total number of samples, For the first The input vector of the group, For the first The group corresponds to the output vector. It is the input vector The number of dimensions, It is the dimension space of the input vector. It is the output The number of dimensions, It is the dimension space of the output vector.

4. The method of claim 1, wherein, The preprocessing adopts a dispersion standardization method, and the specific formula of the dispersion standardization method is: wherein, is the original data, is the standardized data.

5. The method of claim 1, wherein, The specific steps of optimizing the algorithm process of the neural network model based on the weight direct determination method based on pseudo-inverse are as follows: first, randomly generating hidden layer nodes, constructing a hidden layer output matrix, and taking the minimization of the approximate square difference between the output of the neural network model and sample labels as an objective, a target function is established; second, the Moore-Penrose generalized pseudo-inverse matrix of the hidden layer output matrix is calculated by using the generalized matrix theory, and the equivalent weights with thresholds corresponding to the minimum error in the back propagation process of the neural network model algorithm are obtained.

6. The method of claim 1, wherein, The specific steps of determining the optimal number of hidden layer neurons of the neural network model using the two-stage structure self-determination algorithm are as follows: in the first stage of search, a plurality of hidden layer neurons are added at each step, the training set data is imported into the neural network model for training, the initial prediction accuracy of the model under different numbers of hidden layer neurons is compared, and the approximate interval of the optimal number of hidden layer neurons is quickly locked; in the second stage of search, only one hidden layer neuron is added at each step within the approximate interval determined in the first stage, the model is trained again and its performance is evaluated, and finally the number of hidden layer neurons that optimizes the performance of the neural network model is found.

Images