BP neural network based on Adam to predict the landing point deviation

Abstract

The present invention discloses a BP neural network drop point offset prediction method based on Adam, which belongs to the field of drop point prediction and includes the following steps: S1, data collection and preprocessing; S2, constructing a BP neural network, and using the collected data as the input variable of the BP neural network, and setting the prediction target as the offset and offset angle of the source drop point; S3, BP neural network initialization; S4, optimizing the BP neural network based on the Adam optimizer; S5, iterative training until the optimized BP neural network converges or reaches the maximum number of iterations. The present invention adopts the above-mentioned BP neural network drop point offset prediction method based on Adam, which has the advantages of high precision, strong adaptability, data-driven, fast and stable training process, as well as technical integration and scalability, and brings important technical breakthroughs and application prospects to the field of drop point prediction for drone mounted source throwing.

Description

Technical Field

[0001] The present invention relates to the technical field of landing point prediction, and in particular to a landing point offset prediction method based on Adam's BP neural network. Background Art

[0002] Drone seismic source exploration refers to the use of drones for geological exploration, resource exploration, disaster monitoring and other activities, in which accurate prediction of the source point is crucial to improving exploration efficiency and accuracy. In this field, known parameters such as wind speed, wind direction, throwing height, throwing angle, initial velocity and source quality are all key factors affecting the prediction of the point of impact.

[0003] Accurately predicting the source drop point is crucial for exploration missions. First, it can improve the accuracy of exploration and ensure that exploration equipment or materials accurately reach the target area, thereby improving the accuracy and reliability of exploration data. Secondly, accurately predicting the drop point helps improve operational efficiency, reduce the search range of drones, and save time and resources. In addition, when conducting exploration in dangerous or inaccessible areas, accurately predicting the drop point can ensure the safe delivery of drones and sources and reduce potential safety risks. Finally, accurately throwing the source helps reduce interference and damage to the environment, especially in exploration operations in ecologically sensitive areas, which is beneficial to environmental protection. Therefore, accurately predicting the source drop point is not only crucial to the exploration mission itself, but also has important significance in maintaining safety and protecting the environment.

[0004] The existing method for predicting earthquake source locations is to use traditional kinematic equations for calculations. This method has the following limitations: First, these equations often ignore the impact of complex environmental factors on the movement of objects, such as wind speed changes and air resistance. Second, for the convenience of calculation, kinematic models usually simplify the real world, which may cause the prediction results to deviate from the actual situation. In addition, kinematic equations perform poorly when dealing with nonlinear factors, such as the impact of variable wind speed and irregular terrain. Summary of the invention

[0005] To solve the above problems, the present invention provides a BP neural network drop point offset prediction method based on Adam, which has the advantages of high precision, strong adaptability, data-driven, fast and stable training process, technical integration and scalability, etc., which brings important technological breakthroughs and application prospects to the field of UAV mounted seismic source drop point prediction.

[0006] To achieve the above object, the present invention provides a method for predicting the landing point offset based on the BP neural network of Adam, comprising the following steps:

[0007] S1. Data collection and preprocessing, where the collected data include the throwing angle, initial velocity, initial wind speed, initial wind direction and the actual landing point of the earthquake source under the throwing height;

[0008] S2, constructing a BP neural network, and using the collected data as input variables of the BP neural network, and setting the prediction target as the offset and offset angle of the earthquake source landing point;

[0009] S3, BP neural network initialization;

[0010] S4, optimize BP neural network based on Adam optimizer;

[0011] S5. Iterate the training until the optimized BP neural network converges or reaches the maximum number of iterations.

[0012] Preferably, the preprocessing described in step S1 includes data cleaning, feature engineering, and data standardization or normalization, and for the initial wind direction, the wind direction angle is converted into sine and cosine values.

[0013] Preferably, the BP neural network described in step S2 includes an input layer, a hidden layer and an output layer, and the number of neurons in the output layer is the same as the number of prediction targets.

[0014] Preferably, in step S2, cross features that capture the relationship between input variables are introduced into the input variables.

[0015] Preferably, the BP neural network initialization method described in step S3 is random initialization, Xavier initialization or He initialization.

[0016] Preferably, in step S4, the parameters of the Adam optimizer include learning rate, and Two exponential decay rates and Whitening parameters, where and They are used to calculate the exponential decay of the first and second order moments of the gradient, respectively.

[0017] Preferably, step S5 specifically includes the following steps:

[0018] S51, obtaining training samples from the collected data and performing forward propagation;

[0019] S52. Calculate the loss function and its gradient to the parameters, and use the mean square error as the loss function to measure the difference between the predicted value and the actual value:

[0020] ;

[0021] ;

[0022] In the formula, represents the loss function; represents the number of samples; represents the true output of the training data; Represents the BP neural network for training data The predicted value of the loss function is about the model parameters The gradient of is expressed as the loss function Parameters The partial derivative of is the parameter vector of the model, and m is the total number of parameters; represents the true output of the training data; Represents the BP neural network for training data The predicted value of

[0023] S53, update parameters using Adam algorithm;

[0024] S54, looping steps S51-S53 until all training samples have been processed, at which point if the BP neural network performance meets the set requirements, save the BP neural network, otherwise return to S4 for tuning.

[0025] The present invention has the following beneficial effects:

[0026] 1. High-precision prediction capability: Through the BP neural network model optimized based on the Adam algorithm, the landing point offset of the drone-mounted seismic source thrown from high altitude can be accurately predicted, that is, the target landing point can be accurately predicted before throwing, thus improving the accuracy of throwing;

[0027] 2. Strong adaptability: It can adapt to the dynamic changes under various throwing conditions, including changes in throwing angle, initial velocity, wind speed, wind direction, throwing height and other factors. Whether under complex meteorological conditions or in different environments, the model can effectively make predictions and has strong adaptability;

[0028] 3. Data-driven: By collecting a large amount of historical throwing data and combining it with environmental factor data for analysis and prediction, the model's prediction ability is continuously improved as data is continuously accumulated and updated, with good real-time and sustainability.

[0029] 4. Fast and stable training process: The Adam algorithm is used as the optimizer, which can achieve adaptive learning rate in the training process of BP neural network, making the network more stable and faster during the training process, that is, the model can be trained more efficiently, reducing training time and improving training efficiency;

[0030] 5. It has certain scalability and can be applied to other similar throwing prediction problems, such as rocket launch, missile delivery and other fields.

[0031] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The present invention is a flowchart of the Adam-based BP neural network landing point deviation prediction method. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical scheme and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention are further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions.

[0034] It should be noted that the terms "include", "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or devices.

[0035] Like reference numerals and letters denote similar items in the following drawings, and thus, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0036] The BP neural network has the following advantages and benefits: First, it is highly adaptable to complex environments, including taking into account factors such as wind speed changes and air resistance, thereby improving the accuracy of predictions. Second, the BP neural network has a strong nonlinear modeling capability, can better handle nonlinear problems, and is closer to physical phenomena in the real world. In addition, it has self-learning and adaptive capabilities. As the amount of data increases, the neural network can continuously improve its prediction capabilities through learning and adapt to new data and environmental changes. Finally, once training is completed, the BP neural network can predict unseen data and has good generalization capabilities, which makes it have broad application prospects in practical applications.

[0037] Based on the above analysis, the present invention is designed as follows: Figure 1 As shown, the BP neural network landing point offset prediction method based on Adam includes the following steps:

[0038] S1. Data collection and preprocessing, where the collected data include the throwing angle, initial velocity, initial wind speed, initial wind direction and the actual landing point of the earthquake source under the throwing height;

[0039] The preprocessing described in step S1 includes data cleaning (for removing outliers and missing values), feature engineering (such as feature selection and feature conversion), and data standardization or normalization. For the initial wind direction, the wind direction angle is converted into sine and cosine values ​​to better represent the directionality and avoid the discontinuity problem between the angles of 0° and 360°, which helps to improve the training efficiency and prediction accuracy of the model.

[0040] S2, constructing a BP neural network, and using the collected data as input variables of the BP neural network, and setting the prediction target as the offset and offset angle of the earthquake source landing point;

[0041] The BP neural network described in step S2 includes an input layer, a hidden layer, and an output layer, and the number of neurons in the output layer is the same as the number of predicted targets, that is, the output layer has 2 neurons. The number of hidden layers and the number of neurons in each layer need to be determined through experiments. Properly increasing the number of hidden layers and the number of neurons can improve the representation ability of the network, but care should be taken to prevent overfitting. Therefore, in this embodiment, the number of hidden layers is set to 5 layers, the number of neurons in each layer is 10, the number of input layer neurons is 6, and the dimension of the weight matrix is ​​(6,5).

[0042] In step S2, cross features that capture the relationship between input variables are introduced into the input variables. Cross features are input at the input end, and the representation ability of the model is enhanced by introducing cross terms or combinations of features. For example, the input features can be combined in pairs or crossed in higher order to capture the nonlinear relationship between the input features. In this embodiment, the throwing height mainly determines the landing time, the wind speed affects the offset, and the throwing height affects the effective action time of the wind force and thus affects the offset, so the product of wind speed and throwing angle or a polynomial can be introduced.

[0043] S3, BP neural network initialization;

[0044] The BP neural network initialization method described in step S3 is random initialization, Xavier initialization or He initialization.

[0045] In this embodiment, initializing the weights and biases of the neural network is a key step in training the neural network. It is specifically divided into weight initialization and bias initialization. In this embodiment, there are 6 input neurons and 2 output neurons, so the dimension of the weight matrix is ​​(6,5). If a normal distribution is adopted to generate random weight values, the generated random weight values ​​are assigned to the corresponding positions of the weight matrix to complete the weight initialization process. In this embodiment, the number of neurons in the hidden layer is 10, the dimension of the bias is 10, and the initial value of the bias is set to 0 (as the training progresses, the bias value will be adjusted according to the characteristics of the data; choosing a smaller constant value can avoid the problem of activation function output saturation or gradient disappearance caused by the initial value of the bias being too large).

[0046] S4, optimize BP neural network based on Adam optimizer;

[0047] In step S4, the parameters of the Adam optimizer include the learning rate, and Two exponential decay rates and Whitening parameters, where and They are used to calculate the exponential decay of the first and second order moments of the gradient, respectively.

[0048] The learning rate is set to 0.001 to control the step size of parameter updates, affecting the speed at which the model moves in the parameter space. Set to 0.9. Set to 0.999, Set to , used to prevent division by zero errors, is added to the denominator to ensure numerical stability.

[0049] S5. Iterate the training until the optimized BP neural network converges or reaches the maximum number of iterations.

[0050] Step S5 specifically includes the following steps:

[0051] S51, obtaining training samples from the collected data and performing forward propagation;

[0052] S52. The loss function measures the difference between the predicted value of the model on a given training sample and the true label, and the gradient of the parameters needs to adjust the parameters in a specific direction in the parameter space to minimize the loss function. Therefore, the loss function and its gradient to the parameters are calculated, and the mean square error is used as the loss function to measure the difference between the predicted value and the actual value:

[0053] ;

[0054] ;

[0055] In the formula, represents the loss function; represents the number of samples; represents the true output of the training data; Represents the BP neural network for training data The predicted value of the loss function is about the model parameters The gradient of is expressed as the loss function Parameters The partial derivative of is the parameter vector of the model, and m is the total number of parameters; represents the true output of the training data; Represents the BP neural network for training data The predicted value of

[0056] S53, use the Adam algorithm to update parameters; use the Adam algorithm to update parameters. The Adam algorithm is a commonly used optimization algorithm that combines the idea of ​​momentum method and adaptive learning rate. When updating parameters, it considers the estimation of the first-order moment (mean) and second-order moment (variance) of the gradient, and uses exponential decay averaging to dynamically adjust the learning rate;

[0057] S54, looping steps S51-S53 until all training samples have been processed, at which point if the BP neural network performance meets the set requirements, save the BP neural network, otherwise return to S4 for tuning.

[0058] In general, although the traditional kinematic equations can be used to predict the location of the earthquake source of the UAV under certain simplified conditions, in practical applications, considering the changing environmental factors and complex influencing factors, the BP neural network provides a more accurate and flexible solution. That is, the BP neural network can more accurately predict the location of the earthquake source by learning the patterns and relationships in a large amount of data.

[0059] Therefore, the present invention adopts the above-mentioned Adam-based BP neural network landing point offset prediction method, which has the advantages of high precision, strong adaptability, data-driven, fast and stable training process, technical integration and scalability, etc., which brings important technological breakthroughs and application prospects to the field of UAV-mounted seismic source throwing landing point prediction.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. The BP neural network landing point offset prediction method based on Adam is characterized by: The following steps are involved: S1. Data collection and preprocessing, where the collected data include the throwing angle, initial velocity, initial wind speed, initial wind direction and actual landing point of the earthquake source at the throwing height when the earthquake source is mounted on the drone; S2, constructing a BP neural network, and using the collected data as input variables of the BP neural network, and setting the prediction target as the offset and offset angle of the earthquake source landing point; S3, BP neural network initialization; S4, optimize BP neural network based on Adam optimizer; S5, iterative training until the optimized BP neural network converges or reaches the maximum number of iterations; Step S5 specifically includes the following steps: S51, obtaining training samples from the collected data and performing forward propagation; S52. Calculate the loss function and its gradient to the parameters, and use the mean square error as the loss function to measure the difference between the predicted value and the actual value: ； ； In the formula, represents the loss function; ; represents the true output of the training data; Represents the BP neural network for training data The predicted value of Loss function with respect to model parameters Parameters The partial derivative of is the parameter vector of the model, and m is the total number of parameters; S53, update parameters using Adam algorithm; S54, looping steps S51-S53 until all training samples have been processed, at which point if the BP neural network performance meets the set requirements, save the BP neural network, otherwise return to S4 for tuning.

2. The method for predicting the landing point deviation based on the BP neural network of Adam according to claim 1 is characterized in that: The preprocessing described in step S1 includes data cleaning, feature engineering, and data standardization or normalization, and for the initial wind direction, the wind direction angle is converted into sine and cosine values.

3. The method for predicting the landing point deviation based on the Adam BP neural network according to claim 1 is characterized in that: The BP neural network described in step S2 includes an input layer, a hidden layer and an output layer, and the number of neurons in the output layer is the same as the number of prediction targets.

4. The method for predicting the landing point deviation based on the Adam BP neural network according to claim 3 is characterized in that: In step S2, cross features that capture the relationship between input variables are introduced into the input variables.

5. The method for predicting the landing point deviation based on the Adam BP neural network according to claim 1 is characterized in that: The BP neural network initialization method described in step S3 is random initialization, Xavier initialization or He initialization.

6. The method for predicting the landing point deviation based on the BP neural network based on Adam according to claim 1 is characterized in that: In step S4, the parameters of the Adam optimizer include the learning rate, and Two exponential decay rates and Whitening parameters, where and They are used to calculate the exponential decay of the first and second order moments of the gradient, respectively.

Images