Method for training mineral resource identification model based on data collected by unmanned aerial vehicle

Abstract

The present invention relates to the field of artificial intelligence, and particularly to a method for training a mineral resource identification model based on data collected by drones. The training method includes collecting real geological data, training a data generation model, and using the trained data generation model to generate simulated geological data, which together with the real geological data are used as training samples to train a feature extraction model and a classification model. Existing methods for training mineral resource identification models based on data collected by drones have problems such as low model accuracy and slow calculation speed of the mineral resource identification model. The method for training a mineral resource identification model based on data collected by drones provided by the present invention can enable the mineral resource identification model to have higher model accuracy and faster calculation speed.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and particularly to a method for training a mineral resource identification model based on data collected by an unmanned aerial vehicle (UAV). Background Art

[0002] Mineral resources, also known as mineral deposits, refer to aggregates of minerals or useful elements that are formed through geological ore-forming processes, naturally occurring within the Earth's crust, buried underground or exposed on the Earth's surface, in solid, liquid or gaseous states, and have value for development and utilization. Before mining mineral resources, regional general surveys and detailed explorations need to be carried out first.

[0003] Regional general surveys are comprehensive explorations of large areas through various methods such as geology, geophysics, and geochemistry, to preliminarily understand the stratigraphic structure, landforms, soil, rock types, tectonic features, and possible mineral signs in the area. The common methods of regional general surveys are manual exploration and ground instrument exploration. Currently, with the gradual maturity of UAV technology, using UAVs equipped with various sensors for exploration has become a new method for regional general surveys. The various sensors carried by the UAV collect geological data when the UAV flies over the mineral resource area. The geological data covers various geological information in the mining area, including multi-dimensional attributes such as spectrum, density, temperature, and humidity. The data collected by the UAV can provide more comprehensive geological data for the exploration area, greatly improving the exploration efficiency. However, because regional general surveys usually need to cover a vast area, exploration personnel need to process huge and complex geological data, which will consume a large amount of time and manpower. Therefore, a mineral resource identification model based on data collected by UAVs has been developed. This model can analyze geological data to obtain the possible types of mineral resources in the mineral resource area. The mineral resource identification model based on data collected by UAVs mainly includes a feature extraction model for extracting features that are more critical for classification and a classification model for analyzing the geological data after feature extraction to obtain its mineral resource category.

[0004] When training a mineral resource identification model based on data collected by UAVs, traditional neural network algorithms are usually used to train the feature extraction model. Traditional feature extraction methods have poor effects when dealing with complex, non-linear high-dimensional data, and it is difficult to maintain the content integrity and spatial structure consistency of the data, which affects the prediction accuracy of the model. Moreover, traditional neural network algorithms are prone to problems such as gradient disappearance or explosion, especially difficult to stably train in high-dimensional spaces, thereby affecting the feature extraction effect. When using deep neural network algorithms to train the classification model, it is easy to overfit when identifying complex mineral resource categories, and it is difficult to control the computational complexity of the model, resulting in slow model inference speed and poor practicability.

[0005] Therefore, the existing method for training a mineral resource identification model based on data collected by drones has problems such as low model accuracy and slow calculation speed of the mineral resource identification model. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the existing technology and provide a method for training a mineral resource identification model based on data collected by drones, which can make the model accuracy of the mineral resource identification model higher and the calculation speed faster.

[0007] To solve the above technical problems, the method for training a mineral resource identification model based on data collected by drones provided by the present invention includes:

[0008] S1. When the drone flies in the mineral resource area, real geological data of the mineral resource area are collected in real time by a plurality of different types of sensors equipped on the drone, the real geological data are preliminarily processed and stored, and the mineral resource categories of the real geological data after preliminary processing are labeled;

[0009] The labeling method is that professional geologists analyze the real geological data in combination with on-site samples in the mineral resource area for accurate labeling;

[0010] S2. The data generation model constructed by using the generative adversarial network algorithm based on dynamic gradient adjustment is trained with the labeled real geological data. When training the data generation model constructed by using the generative adversarial network algorithm based on dynamic gradient adjustment, the generative adversarial network algorithm based on dynamic gradient adjustment is adopted to obtain a trained data generation model. The simulated geological data are generated by using the trained data generation model, and the simulated geological data and the labeled real geological data are used as training samples together;

[0011] S3. Judge the data feature dimension of the training samples. When the data feature dimension is greater than or equal to 20, the feature extraction model constructed by using the autoencoder algorithm based on double constraints is trained with the training samples. When training the feature extraction model constructed by using the autoencoder algorithm based on double constraints, the autoencoder algorithm based on double constraints is adopted; when the data feature dimension is less than 20, the feature extraction model constructed by using the neural network algorithm based on dynamic population evolution optimization is adopted. When training the feature extraction model constructed by using the neural network algorithm based on dynamic population evolution optimization, the neural network algorithm based on dynamic population evolution optimization is adopted; a trained feature extraction model and the training samples after feature extraction are obtained;

[0012] S4. Use the training samples after the feature extraction to train the classification model constructed by the deep neural decision tree classification algorithm based on quantum transform pruning. When training the classification model constructed by the deep neural decision tree classification algorithm based on quantum transform pruning, use the deep neural decision tree classification algorithm based on quantum transform pruning to obtain the trained classification model.

[0013] As a further improvement of the present invention: The training of the data generation model constructed by the generative adversarial network algorithm based on dynamic gradient adjustment in S2 includes:

[0014] S201. Initialize the network parameters of the generator and discriminator of the generative adversarial network;

[0015] S202. During the adversarial training process, the generator and discriminator are alternately trained. The generator generates fake geological data, and the discriminator determines whether the input comes from the real geological data set or the generator. In this way, the generator learns to generate geological data that is increasingly difficult to distinguish by the discriminator. The objective functions of the generator and discriminator are as follows:

[0016]

[0017] In the formula, is the generator parameter corresponding to minimizing the expectation, is the weight of the generator, is the discriminator parameter corresponding to maximizing the expectation, is the weight of the discriminator, is the expectation operation, x is the real geological data sample, p data (x) is the distribution of the real geological data, D c (x) is the discriminator's judgment result on the real sample, z is the input noise of the generator, from the predefined noise distribution p z (z), D c () is the discriminator function, G c () is the generator function, G c (z) is the fake geological data generated by the generator;

[0018] In the adversarial training stage, as the discriminator's judgment result D c (x) of the discriminator output in the loss function and the discriminator's judgment result D c (G c (z)) of the discriminator on the fake geological data generated by the generator are implemented through a neural network layer with a Sigmoid activation function to ensure that the output value is between 0 and 1, expressed as:

[0019]

[0020] In the formula, fc () is the linear output of the discriminator network, is the weight of the discriminator;

[0021] S203. In each round of training, according to the feedback of the discriminator, dynamically adjust the learning rates of the generator and the discriminator. When the accuracy of the discriminator is too high, increase the learning rate of the generator and decrease the learning rate of the discriminator, and vice versa. This dynamic adjustment helps to avoid mode collapse during the training process and ensure the diversity of geological data generation. The dynamic gradient adjustment is achieved by adjusting the learning rate, and the adjustment strategy is expressed as:

[0022]

[0023] In the formula, is the learning rate of the generator, η0 is the base learning rate, is the adjustment coefficient of the generator learning rate, ∥∥ is the L2 norm, is the gradient of the generator's loss function with respect to its parameters, is the loss function of the generator, is the learning rate of the discriminator, is the adjustment coefficient of the discriminator learning rate, is the gradient of the discriminator's loss function with respect to its parameters, is the loss function of the discriminator;

[0024] S204. Use the self-similarity principle to guide the generator to focus on the local similarity of geological data, and improve the quality and practicality of the fake geological data. The self-similarity adjustment is achieved by adding a regularization term to the loss function of the generator to guide the model to pay attention to the self-similarity characteristics of local geological data. The calculation formula of the regularization term is:

[0025]

[0026] In the formula, R c () is the self-similarity regularization term calculation function, G c (z) is the fake geological data generated by the generator, β c is the regularization coefficient of the generative adversarial network, G c (z) i is the i-th feature region in the fake geological data, G c (z) j is the j-th feature region in the fake geological data, s c () is the self-similarity descriptor calculation function;

[0027] S205. Repeat the iterative steps S202 to S204 until the preset stop iteration condition is met.

[0028] Preferably, in the step S202, the the calculation formulas are respectively:

[0029]

[0030] In the formula, is the weight of the discriminator network, is the bias of the discriminator network, x i is the i-th characteristic attribute of the real geological data sample, G c,i (z) is the i-th output of the generator.

[0031] As a further improvement of the present invention: The training of the feature extraction model constructed based on the dual-constrained autoencoder algorithm in the S3 includes:

[0032] S311. Geological data passes through the autoencoder from the input layer through multiple hidden layers, and each layer performs a linear transformation and a non-linear activation, and finally outputs the extracted features. The transfer function of each layer is expressed as:

[0033] x p,i+1 = Sig(W p,i x p,i + b p,i ),

[0034] In the formula, x p,i+1 is the output vector of the (i + 1)-th layer of the autoencoder, Sig() is the Sigmoid activation function, W p,i is the weight matrix of the i-th layer of the autoencoder, x p,i is the input vector of the i-th layer of the autoencoder, b p,i is the bias vector of the autoencoder;

[0035] S312. The loss function of the autoencoder is designed as a dual-constrained loss. One part comes from the reconstruction loss of the geological data content to ensure that the extracted features can effectively reconstruct the original geological data; the other part is the structure-preserving loss to ensure that the original structure of the geological data is maintained in the feature space. The calculation formula of the loss function of the autoencoder is:

[0036] L p = λ p1 L p,recon + λ p2 L p,struct ,

[0037] In the formula, L p is the loss function of the autoencoder, λ p1 is the weight coefficient for adjusting the relative importance of the reconstruction loss, L p,recon is the reconstruction loss, λ p2To adjust the weight coefficient of the relative importance of the structure preservation loss,

[0038] L p,struct is the structure preservation loss;

[0039] S313. Calculate the gradient according to the loss function and update the network parameters using the least squares gradient descent method, considering the stability of the gradient and the adaptability of the update rate to avoid the gradient explosion or disappearance problem in the high-dimensional space. The parameter update rule is:

[0040]

[0041] where θ p,i is the parameter of the i-th layer of the autoencoder, ← represents the parameter update operation, and η p is the learning rate of the autoencoder, is the loss function L of the autoencoder p is the gradient of the parameter of the i-th layer of the autoencoder;

[0042] S314. Repeat the iterative steps S311 to S313 until the preset stop iteration condition is satisfied.

[0043] Preferably, the calculation formula of the reconstruction loss L p,recon in 4 S312 is:

[0044]

[0045] where x p is the original input geological data, is the output of the autoencoder;

[0046] The calculation formula of the structure preservation loss L p,struct is:

[0047] L p,struct = ∑ j,k W p,jk ∥z p,j - z p,k ∥ 2 ,

[0048] where W p,jk is the similarity matrix based on the input geological data, z p,j is the j-th geological data point mapped to the feature space, and z p,k is the k-th geological data point mapped to the feature space.

[0049] As a further improvement of the present invention: It is characterized in that the training of the feature extraction model constructed based on the dynamic population evolution optimization neural network algorithm in S3 includes:

[0050] S321. According to the initialization method of the bionic algorithm, in the initialization stage, an initial population is generated, where each individual represents a configuration of network weights. Let the population size be N. p , initialize the weights and biases for the i-th individual, expressed as:

[0051]

[0052] In the formula, is the weight matrix of the i-th individual in the initial state, ~ follows a specific distribution, is the normal distribution; is the normal distribution with a mean of 0 and a variance of σ 2 ; σ 2 is the initialized variance, is the bias of the i-th individual in the initial state; W pi is the weight of the neural network corresponding to the i-th individual, b pi is the bias of the neural network corresponding to the i-th individual;

[0053] S322. For each individual in the population, use its corresponding neural network configuration to process the input training data, calculate the output of the model, and evaluate its performance according to a predetermined loss function. For the i-th individual, calculate the loss on the training dataset using its weights and biases, expressed as:

[0054]

[0055] In the formula, L pi is the loss of the neural network corresponding to the i-th individual, mps is the number of samples in the current batch of input, is the composite loss function, fsig() is the neural network model function, is the feature of the j-th geological data sample, W pi is the weight of the neural network corresponding to the i-th individual, b pi is the bias of the neural network corresponding to the i-th individual, is the label of the j-th geological data sample;

[0056] S323. According to the fitness of the individuals, select the best-performing individuals from the current population to be retained as candidate solutions for the next generation. Selection is based on the fitness of the individuals, and excellent individuals have a higher probability of being selected. The calculation formula for the probability of an individual being selected is:

[0057]

[0058] In the formula, P select (i) is the probability of the i-th individual being selected, γ pse is the parameter controlling the selection pressure, Lpi is the loss of the neural network corresponding to the i-th individual, N p is the population size, L pk is the loss of the neural network corresponding to the k-th individual;

[0059] S324. Generate new individuals through crossover and mutation operations. The crossover operation allows two excellent individuals to exchange part of their genes to generate new offspring; the mutation operation randomly changes part of the genes in an individual to increase the diversity of the population. The crossover operation randomly selects two individuals for gene exchange, expressed as:

[0060] ′

[0061] W pi = α pcs W p1 +(1 - α pcs )W p2 ,

[0062] ′

[0063] b pi = α pcs b p1 +(1 - α pcs )b p2 ,

[0064] ′

[0065] In the formula, W pi is the weight of the neural network corresponding to the individual after the crossover operation, α pcs is the crossover rate,

[0066] W p1 is the weight of the neural network corresponding to the first selected individual, W p2 is the weight of the neural network corresponding to the second selected individual

[0067] ′, b pi is the bias of the neural network corresponding to the individual after the crossover operation, b p1 is the bias of the neural network corresponding to the first selected individual, b p2 is the bias of the neural network corresponding to the second selected individual;

[0068] The mutation operation performs a small random perturbation on the weights of the newly generated individual, expressed as:

[0069]

[0070] ″′

[0071] In the formula, W pi is the weight of the neural network corresponding to the individual after the mutation operation, W piis the weight of the neural network corresponding to the individual after the crossover operation, is a normal distribution with a mean of 0 and a variance of τ 2 where τ 2

[0072] ″′

[0073] is the variance of the mutation, and b pi is the bias of the neural network corresponding to the individual after the mutation operation, and b pi is the bias of the neural network corresponding to the individual after the crossover operation;

[0074] S325. Repeat the iterative steps S322 to S324 until the preset stop iteration condition is satisfied.

[0075] Preferably, in S322, the is calculated as:

[0076]

[0077] where is the composite loss function, fsig() is the neural network model function, is the feature of the j-th geological data sample, and W pi is the weight of the neural network corresponding to the i-th individual, and b pi is the bias of the neural network corresponding to the i-th individual, is the label of the j-th geological data sample, MSE() is the mean square error function, and λ ps is the regularization parameter, and Reg(W pi ) is the regularization term;

[0078] The regularization term Reg(W pi ) is calculated as:

[0079]

[0080] where W pi,k is the k-th weight of the neural network corresponding to the i-th individual.

[0081] As a further improvement of the present invention: The training in S4 for the classification model constructed based on the quantum transformation pruning deep neural decision tree classification algorithm includes:

[0082] S401. Initialize the parameters of the deep neural decision tree, including the weights of the decision nodes and the depth of the tree. Let the weight of the i-th node in the deep neural decision tree be W u,i , and the bias of the i-th node in the deep neural decision tree be b u,i , and the initialization method is expressed as:

[0083]

[0084] Wherein, is a normal distribution with a mean of 0 and a standard deviation of 1, and ~ follows a specific distribution;

[0085] S402. During the training process, the input geological data after feature extraction passes through multiple decision nodes of the decision tree. Each node makes a decision based on the weights and input features, guiding the geological data to flow to the left or right subtree until it reaches the leaf node. Let the feature vector input to the deep neural decision tree be x u , which is expressed as:

[0086] y u,n = Sig(W u,n x u + b u,n ),

[0087] Wherein, y u,n is the output of the nth node of the deep neural decision tree; Sig() is the Sigmoid activation function; W u,n is the weight of the nth node in the deep neural decision tree, and b u,n is the bias of the nth node in the deep neural decision tree;

[0088] S403. Considering the classification accuracy and the complexity of the tree, a loss function is used to evaluate the performance of the current decision tree. The loss function of the deep neural decision tree includes classification error and regularization term to prevent overfitting. The calculation formula is:

[0089]

[0090] Wherein, L u is the loss function of the deep neural decision tree, N u is the number of samples input to the deep neural decision tree in the current batch, t i is the target output of the ith sample, y u,i is the classification result of the deep neural decision tree for the ith sample, λ u is the regularization coefficient of the deep neural decision tree, and W u is the weight of the deep neural decision tree; ∥∥ is the L2 norm;

[0091] S404. In each training cycle, according to the quantum transformation principle, the utility and efficiency of each decision node are evaluated. Inefficient nodes will be pruned to reduce the model complexity while maintaining or improving the model performance. The present invention adopts the quantum transformation principle to evaluate and adjust the structure of the network. Let P u,i represent the importance of the ith node. If P u,i is less than the preset threshold θ u, then this node will be removed from the model, P u,i The calculation formula is:

[0092]

[0093] In the formula, L u is the loss function of the deep neural decision tree, W u,i is the weight of the i-th node in the deep neural decision tree, ψ r () is the quantum transformation function; Z r is the input feature vector of the current node of the deep neural decision tree;

[0094] S405. Repeat and iterate steps S402 to S404 until the preset stop iteration condition is satisfied, which indicates that the model training is completed.

[0095] Preferably, in S404, the ψ r (z r ) calculation formula is:

[0096]

[0097] In the formula, Z r is the input feature vector of the current node of the deep neural decision tree, λ r is the parameter controlling the quantum transformation sensitivity, z rj is the j-th element of z r , is the quantum noise level corresponding to z rj ;

[0098] The quantum noise level has different compression levels in different parts of the feature representation, so as to more finely control the information loss. The calculation formula is:

[0099]

[0100] In the formula, α r is the coefficient for adjusting the quantum noise response, z rj is the j-th element of z r .

[0101] The beneficial effects of the present invention are as follows: The method for training a mineral resource identification model based on data collected by an unmanned aerial vehicle provided by the present invention can make the mineral resource identification model have a relatively high model accuracy and a relatively fast calculation speed.

[0102] First, before training the feature extraction model and the classification model, the present invention first trains a data generation model constructed by the generative adversarial network algorithm based on dynamic gradient adjustment using the generative adversarial network algorithm based on dynamic gradient adjustment. Traditional generative adversarial networks may face the problem of mode collapse when simulating complex geological data distributions, resulting in insufficient diversity of the generated geological data. However, the generative adversarial network algorithm based on dynamic gradient adjustment adopts a dynamic gradient adjustment strategy, and realizes a more refined learning of geological data distributions by dynamically adjusting the learning rates of the generator and the discriminator during the adversarial process. Therefore, a large number of simulated geological data can be obtained through the trained data generation model, and using them together with the real geological data as training samples to train the feature extraction model and the classification model can effectively expand the training samples.

[0103] Secondly, the present invention selects feature extraction models based on different algorithms according to the number of data feature dimensions. When the data feature dimension is greater than or equal to 20, a feature extraction model based on the autoencoder algorithm with double constraints is used to process the geological data collected by the drone. Due to terrain and environmental factors, geological data has a high degree of nonlinearity and complexity. The double constraint mechanism ensures the integrity of mineral features and the spatial topological structure of geological data during the feature extraction process through the double mechanisms of content constraint and structure constraint, optimizes the representation of the feature space, and meets the requirements of high-dimensional complex geological data. For low-dimensional data with a data feature dimension less than 20, a feature extraction model based on the neural network algorithm with dynamic population evolution optimization is used. In traditional population evolution algorithms, the evolution of all individuals is based on fixed rules. The present invention uses a self-correcting mechanism to automatically adjust the evolution rules according to the characteristics of the current training data, and adjusts the probabilities of crossover and mutation according to the change trend of the loss function, so that the algorithm can more flexibly adapt to different geological data distributions, improve the generalization ability of the model and the training efficiency. By using the dynamic population evolution optimization algorithm and adjusting the evolution rules and parameters through the self-correcting mechanism, the adaptability of the neural network to variable geological data is improved, and the effectiveness of feature extraction and the robustness of the model are enhanced.

[0104] Finally, the present invention adopts a deep neural decision tree classification algorithm based on quantum transform pruning to improve the classification accuracy and optimize the complexity of the model, effectively control the complexity and computational amount of the classifier, enable it to accurately classify different mineral categories, and improve the computational efficiency at the same time.

[0105] Therefore, after adding a data augmentation model and optimizing the training method of the feature extraction model, the model accuracy of the mineral resource identification model based on drone-collected data is relatively high. After optimizing the training method of the classification model, the computational speed of the mineral resource identification model based on drone-collected data is relatively fast. Description of the Drawings

[0106] Figure 1This is the principle block diagram of the present invention. Detailed implementation manners

[0107] The following further elaborates in detail the detailed implementation manners of the present invention in conjunction with the accompanying drawings.

[0108] As Figure 1 shown, the method for training a mineral resource identification model based on data collected by an unmanned aerial vehicle provided by the present invention includes:

[0109] S1. When the unmanned aerial vehicle flies over the mineral resource area, multiple different types of sensors equipped on the unmanned aerial vehicle are used to collect real geological data of the mineral resource area in real time, the real geological data is preliminarily processed and stored, and the mineral resource categories of the preliminarily processed real geological data are labeled;

[0110] The multiple different types of sensors include a high-resolution camera, a multispectral sensor, an infrared scanner, an altimeter, etc.

[0111] The unmanned aerial vehicle is a specifically designed unmanned aerial vehicle according to usage requirements. Before flying, a flight path has been preset in advance. At the same time, by using an optical camera, an infrared thermal imager, a radar carried by the unmanned aerial vehicle and combining image recognition algorithms and target detection algorithms, an automatic target recognition function is realized to ensure that the specified mineral resource area can be comprehensively and accurately covered during flight.

[0112] In practical applications, the attributes of real geological data are usually more than ten, and can reach dozens or even hundreds. Common attributes of real geological data include:

[0113] Ra is the reflectivity, which describes the reflection ability of surface materials to light in a specific band;

[0114] Da is the density, which reflects the density information of the soil or rock in the mining area;

[0115] Ea is the height difference, which is obtained by the altimeter carried by the unmanned aerial vehicle;

[0116] Ta is the temperature index, which is the surface temperature measured by infrared scanning;

[0117] Pa is the pressure value, which is related to the degree of compaction of surface materials;

[0118] Fa is the fractal dimension, which describes the complexity of the surface texture;

[0119] Ga is the spectral signature, which is the feature vector extracted from multispectral data;

[0120] Ba is the magnetic index, which reflects the magnetic properties of surface minerals;

[0121] Ha is the humidity index, which characterizes the humidity level of the collection area.

[0122] The preliminary processing includes image correction, noise filtering, and format conversion. Format conversion refers to converting real geological data into a unified data format, namely HDF5 (Hierarchical Data Format).

[0123] Real geological data is stored in the form of discrete vectors.

[0124] The annotation method is that professional geologists accurately annotate real geological data by combining on-site samples in the mineral resource area.

[0125] The categories of mineral resources include iron ore, copper ore, gold ore, etc.

[0126] S2. Use the real geological data after annotation to train a data generation model constructed by the generative adversarial network algorithm based on dynamic gradient adjustment. After training is completed, a trained data generation model is obtained. Use the trained data generation model to generate simulated geological data, and use the simulated geological data and the real geological data after annotation as training samples together.

[0127] The generative adversarial network is a deep learning algorithm architecture composed of two neural networks, a generator and a discriminator, which compete and play against each other to achieve the purpose of generating new data. Dynamic gradient adjustment means that in each round of training, according to the feedback of the discriminator, using the gradients of the loss functions of the generator and the discriminator with respect to their parameters, dynamically adjust the learning rates of the generator and the discriminator. When the accuracy of the discriminator is too high, increase the learning rate of the generator and decrease the learning rate of the discriminator, and vice versa, to help avoid mode collapse during the training process and ensure the diversity of geological data generation. Training the data generation model constructed by the generative adversarial network algorithm based on dynamic gradient adjustment uses the generative adversarial network algorithm based on dynamic gradient adjustment. The training includes:

[0128] S201. Initialize the network parameters of the generator and discriminator of the generative adversarial network. The generator aims to simulate the distribution of geological data, and the goal of the discriminator is to distinguish real geological data from the fake geological data generated by the generator. Let the generator be G c , and the discriminator be D c . The initialization method is expressed as:[[]]

[0129]

[0130] In the formula, is the weight of the generator, represents a normal distribution with a mean of 0 and a variance of 0.01, is the weight of the discriminator;

[0131] S202. During the adversarial training process, the generator and the discriminator are alternately trained. The generator generates fake geological data, and the discriminator determines whether the input comes from the real geological data set or the generator. In this way, the generator learns to generate geological data that is increasingly difficult for the discriminator to distinguish. The objective functions of the generator and the discriminator are as follows:

[0132]

[0133] In the formula, is the generator parameter corresponding to minimizing the expectation, is the weight of the generator, is the discriminator parameter corresponding to maximizing the expectation, is the weight of the discriminator, is the expectation operation, x is the real geological data sample, p data (x) is the distribution of the real geological data, D c (x) is the discriminator's judgment result on the real geological data sample, z is the input noise of the generator, coming from the predefined noise distribution p z (z), D c () is the discriminator function, G c () is the generator function, G c (z) is the fake geological data generated by the generator;

[0134] During the adversarial training stage, as the discriminator output in the loss function, the discriminator's judgment result D c (x) on the real geological data sample and the discriminator's judgment result D c (G c (z)) are implemented through a neural network layer with a Sigmoid activation function to ensure that the output value is between 0 and 1, expressed as:

[0135]

[0136] In the formula, f c () is the linear output of the discriminator network, depending on the input x or G c (z) and its network parameters is the weight of the discriminator;

[0137] The linear output of the discriminator network is the weighted sum of the multi-layer perceptron, The calculation formulas of are respectively:

[0138]

[0139] In the formula, is the weight of the discriminator network, is the bias of the discriminator network, xi is the i-th characteristic attribute of the real geological data sample, G c,i (z) is the i-th output of the generator.

[0140] S203. In each round of training, according to the feedback of the discriminator, dynamically adjust the learning rates of the generator and the discriminator. When the accuracy of the discriminator is too high, increase the learning rate of the generator and decrease the learning rate of the discriminator, and vice versa. This dynamic adjustment helps to avoid mode collapse during the training process and ensure the diversity of geological data generation. The dynamic gradient adjustment is achieved by adjusting the learning rate, and the adjustment strategy is expressed as:

[0141]

[0142] In the formula, is the learning rate of the generator, η0 is the base learning rate, is the adjustment coefficient of the generator's learning rate, ∥∥ is the L2 norm, is the gradient of the generator's loss function with respect to its parameters, is the generator's loss function, is the learning rate of the discriminator, is the adjustment coefficient of the discriminator's learning rate, is the gradient of the discriminator's loss function with respect to its parameters, is the discriminator's loss function;

[0143] The adjustment coefficients of the learning rates of the generator and the discriminator are adjusted according to the second moment of the loss gradient to achieve the adaptive optimization of the learning rate of the generative adversarial network. The calculation method is expressed as:

[0144]

[0145] In the formula, is the adjustment coefficient of the generator's learning rate, σ() is the standard deviation calculation function of the gradient, G c () is the generator function, is the weight of the generator, μ() represents the mean calculation function of the gradient; ∈ is a preset small constant, and ∈ is set to 0.001; is the adjustment coefficient of the discriminator's learning rate, D c () is the discriminator function, is the weight of the discriminator;

[0146] The calculation methods of the standard deviation and mean of the gradient are expressed as:

[0147]

[0148] In the formula, N cg represents the number of samples input to the generative adversarial network in the current batch, is the gradient of the loss function of the generator for the n-th sample with respect to its parameters, represents the gradient of the loss function of the discriminator for the n-th sample with respect to its parameters.

[0149] S204. Utilize the self-similarity principle to guide the generator to focus on the local similarity of geological data, improve the quality and usability of the fake geological data, and combine the self-similarity process to enable the generated geological data to better simulate the local characteristics of geological data while maintaining the overall distribution consistency, enhancing the model's learning ability for small-sample geological data. By dynamically adjusting the learning rates of the generator and the discriminator and introducing the self-similarity regularization term, the problem of mode collapse in traditional generative adversarial networks is overcome, the generated data has higher diversity, and the generalization ability of the model is enhanced. The self-similarity adjustment guides the model to attach importance to the self-similarity characteristics of local geological data by adding a regularization term to the loss function of the generator. The calculation formula of the regularization term is:

[0150] R c (G c (z)) = β c ∑ i,j ∥G c (z) i - G c (z) j ∥ + s c (G c (z)),

[0151] where R c () is the self-similarity regularization term calculation function, G c (z) is the fake geological data generated by the generator, β c is the regularization coefficient of the generative adversarial network, G c (z) i is the i-th feature region in the fake geological data, G c (z) j is the j-th feature region in the fake geological data, s c () is the self-similarity descriptor calculation function;

[0152] The self-similarity descriptor measures the similarity within the generated samples, and the calculation method is expressed as:

[0153]

[0154] where |P| is the total number of geological data feature pairs participating in the comparison, P is all possible geological data feature pairs, is the mean of G c (z), G c (z) pis the p-th eigenpair of the fake geological data generated by the generator, τ c is the regularization scale parameter that controls the degree of influence of self-similarity.

[0155] S205. Repeat the iterative steps S202 to S204 until the preset iteration stop condition is met, which indicates that the model training is completed. The preset iteration stop condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0156] For the sake of distinction, the geological data generated by the data generation model during the training process is called fake geological data, and the geological data generated by the trained data generation model is called simulated geological data.

[0157] S3. Determine the data feature dimension of the training samples. When the data feature dimension is greater than or equal to 20, use the training samples to train the feature extraction model constructed by the autoencoder algorithm based on double constraints; when the data feature dimension is less than 20, use the neural network algorithm based on dynamic population evolution optimization to construct the feature extraction model; after the training is completed, the trained feature extraction model and the training samples after feature extraction are obtained;

[0158] The encoder of the autoencoder is responsible for mapping the high-dimensional input data to the low-dimensional feature space, and the decoder of the autoencoder is used to reconstruct the reduced-dimensional features back to the original space. The result of feature extraction is the output feature of the encoder. The double constraint mechanism is, firstly, the content constraint of the geological data to ensure the integrity and distinguishability of the mineral features during the feature extraction process; secondly, the structural constraint to maintain the geometry and topology of the geological data in the original space, thereby optimizing the representation of the geological data in the feature space. Training the feature extraction model constructed by the autoencoder algorithm based on double constraints uses the autoencoder algorithm based on double constraints, and the training includes:

[0159] S311. The geological data passes through the autoencoder from the input layer through multiple hidden layers, and each layer performs linear transformation and non-linear activation, and finally outputs the extracted features. The transfer function of each layer is expressed as:

[0160] x p,i+1 =Sig(W p,i x p,i +b p,i ),

[0161] In the formula, x p,i+1 is the output vector of the (i + 1)-th layer of the autoencoder, Sig() is the Sigmoid activation function, W p,i is the weight matrix of the i-th layer of the autoencoder, x p,i is the input vector of the i-th layer of the autoencoder, and b p,i is the bias vector of the autoencoder;

[0162] S312. The loss function of the autoencoder is designed as a dual-constraint loss. One part comes from the reconstruction loss of the geological data content to ensure that the extracted features can effectively reconstruct the original geological data; the other part is the structure-preserving loss to ensure that the original structure of the geological data is maintained in the feature space. The calculation formula of the loss function of the autoencoder is as follows:

[0163] L p = λ p1 L p,recon + λ p2 L p,struct ,

[0164] In the formula, L p is the loss function of the autoencoder, λ p1 is the weight coefficient for adjusting the relative importance of the reconstruction loss, L p,recon is the reconstruction loss, λ p2 is the weight coefficient for adjusting the relative importance of the structure-preserving loss,

[0165] L p,struct is the structure-preserving loss;

[0166] The calculation formula of the reconstruction loss L p,recon is as follows:

[0167]

[0168] In the formula, x p is the original input geological data, is the output of the autoencoder;

[0169] The structure-preserving loss aims to maintain the original topological structure of the geological data and is calculated using the Laplacian eigenmap method. The calculation formula of the structure-preserving loss L p,struct is as follows:

[0170] L p,struct = ∑ j,k W p,jk ∥z p,j - z p,k ∥ 2 ,

[0171] In the formula, W p,jk is the similarity matrix based on the input geological data, z p,j is the j-th geological data point mapped to the feature space, z p,k is the k-th geological data point mapped to the feature space.

[0172] Decompose the output of the autoencoder into the non-linear mapping relationship between the output and input of the network. The calculation method is expressed as:

[0173]

[0174] wherein, g p () is the decoding function of the autoencoder, responsible for mapping the representation h p (x p , W p,enc ) back to the original geological data space; h p () is the encoding function, which maps the original input to a latent space; W p,enc is the weight parameter of the encoder in the autoencoder, and W p,dec is the weight parameter of the decoder in the autoencoder.

[0175] The encoding function is a superposition of multi-layer networks, and the output of each layer is a non-linear transformation of the output of the previous layer, expressed as:

[0176] h p,k+1 = A p ⊙ Ser(W p,k h p,k + b p,k ),

[0177] wherein, h p,k+1 is the input of the (k + 1)-th layer of the encoding function, A p represents the attention score of the feature, ⊙ represents element-wise multiplication, Ser() is the dynamic adaptive activation function, W p,k is the weight of the k-th layer of the encoder, h p,k is the input of the k-th layer of the encoding function, and b p,k is the bias of the k-th layer of the encoder;

[0178] Based on the input geological data x p the calculation method of the similarity matrix W p,jk is expressed as:

[0179]

[0180] wherein, x p,j is the j-th sample in the original geological data, and x p,k is the k-th sample in the original geological data; σ p is the scale parameter used to control the similarity sensitivity, and σ p is set to 0.1.

[0181] Adopt a feature selection module based on the attention mechanism to further enhance the sensitivity of the autoencoder to the most informative part of the input geological data, thereby improving the ability of the overall model to identify key features in geological data. The attention score A p of the feature is expressed as:

[0182] A p = sof(W p,att ·h p,k-1 + b p,att ),

[0183] where sof() is the Softmax function, W p,att is the weight of the learned attention layer, h p,k-1 is the input of the (k - 1)-th layer of the encoding function, and b p,att is the bias of the learned attention layer;

[0184] In each layer of the autoencoder, a dynamic adaptive activation strategy is adopted to automatically adjust the shape and parameters of the activation function. Especially when dealing with highly non-linear mineral geology data, it can significantly improve the learning ability and generalization performance of the model. Let the input of the dynamic adaptive activation function be x puts , and the calculation method of the dynamic adaptive activation function is expressed as:

[0185]

[0186] where Ser() is the dynamic adaptive activation function; α p is the first dynamic adaptive activation function adjustment parameter, used to adjust the slope of the curve; β p is the second dynamic adaptive activation function adjustment parameter, used to adjust the center position of the function.

[0187] In each iteration, the updates of α p and β p not only depend on the partial derivatives of the loss function with respect to them, but also include adjustment terms based on historical gradients to ensure the stability and adaptability of the updates. The update method is expressed as:

[0188]

[0189] where η p is the learning rate of the autoencoder, L p is the loss function of the autoencoder, λ p is the hyperparameter that regulates the influence of historical gradients, Δα p is the change in the first dynamic adaptive activation function adjustment parameter between this iteration and the previous iteration, and Δβ p is the change in the second dynamic adaptive activation function adjustment parameter between this iteration and the previous iteration;

[0190] S313. Calculate the gradient according to the loss function and update the network parameters using the least squares gradient descent method, considering the stability of the gradient and the adaptability of the update rate to avoid the problem of gradient explosion or disappearance in high-dimensional space. The parameter update rule is:

[0191]

[0192] In the formula, θ p,i is the parameter of the i-th layer of the autoencoder, and ← represents the parameter update operation; η p is the learning rate of the autoencoder, and η p is set to 0.01; is the loss function L of the autoencoder p is the gradient of the parameter of the i-th layer of the autoencoder;

[0193] S314. Repeat the iterative steps S311 to S313 until the preset stop iteration condition is satisfied, which indicates that the model training is completed. The preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0194] Dynamic population evolution optimization is to use a self-correction mechanism to automatically adjust the evolution rules according to the characteristics of the current training data, and adjust the probabilities of crossover and mutation according to the change trend of the loss function. Training the feature extraction model constructed by the neural network algorithm based on dynamic population evolution optimization adopts the neural network algorithm based on dynamic population evolution optimization, and the training includes:

[0195] S321. According to the initialization method of the bionic algorithm, in the initialization stage, generate an initial population, and each individual represents a configuration of network weights. Let the population size be N p , initialize the weights and biases for the i-th individual, which is expressed as:

[0196]

[0197] In the formula, is the weight matrix of the i-th individual in the initial state, and ~ follows a specific distribution, is the normal distribution; is the normal distribution with a mean of 0 and a variance of σ 2 ; σ 2 is the initialized variance, is the bias of the i-th individual in the initial state; W pi is the weight of the neural network corresponding to the i-th individual, and b pi is the bias of the neural network corresponding to the i-th individual;

[0198] S322. For each individual in the population, use its corresponding neural network configuration to process the input training data, calculate the output of the model, and evaluate its performance according to the predetermined loss function. For the i-th individual, calculate the loss on the training data set using its weights and biases, which is expressed as:

[0199]

[0200] where L pi is the loss of the neural network corresponding to the i-th individual, mps is the number of samples in the current batch, is the composite loss function, fsig() is the neural network model function, is the feature of the j-th geological data sample, W pi is the weight of the neural network corresponding to the i-th individual, b pi is the bias of the neural network corresponding to the i-th individual, is the label of the j-th geological data sample;

[0201] The composite loss function includes a regularization term, which can increase the generalization ability of the model,

[0202] The calculation formula of

[0203]

[0204] is as follows: where is the composite loss function, fsig() is the neural network model function, pi is the weight of the neural network corresponding to the i-th individual, b pi is the bias of the neural network corresponding to the i-th individual, is the label of the j-th geological data sample, MSE() is the mean square error function; λ ps is the regularization parameter, γ pse is set to 2; Reg(W pi ) is the regularization term;

[0205] The calculation formula of the regularization term Reg(W pi ) is as follows:

[0206]

[0207] where W pi,k is the k-th weight of the neural network corresponding to the i-th individual.

[0208] S323. According to the fitness of the individuals, select the best-performing individuals from the current population for retention as candidate solutions for the next generation. Selection is based on the fitness of the individuals, and excellent individuals have a higher probability of being selected. The calculation formula for the probability of an individual being selected is:

[0209]

[0210] where P select (i) is the probability that the i-th individual is selected, γ pseParameter for controlling the selection pressure, L pi Loss of the neural network corresponding to the i-th individual, N p Population size, L pk Loss of the neural network corresponding to the k-th individual;

[0211] S324. Generate new individuals through crossover and mutation operations. The crossover operation allows two excellent individuals to exchange some genes to generate new offspring; the mutation operation randomly changes some genes in an individual to increase the diversity of the population. The crossover operation randomly selects two individuals for gene exchange, expressed as:

[0212] ′

[0213] W pi =α pcs W p1 +(1 - α pcs )W p2 ,

[0214] ′

[0215] b pi =α pcs b p1 +(1 - α pcs )b p2 ,

[0216] ′

[0217] In the formula, W pi is the weight of the neural network corresponding to the individual after the crossover operation; α pcs is the crossover rate, α pcs is set to 0.3; W p1 is the weight of the neural network corresponding to the first selected individual, W p2 is the ′

[0218] weight of the neural network corresponding to the second selected individual, b pi is the bias of the neural network corresponding to the individual after the crossover operation, b p1 is the bias of the neural network corresponding to the first selected individual, b p2 is the bias of the neural network corresponding to the second selected individual;

[0219] The mutation operation performs a small random perturbation on the weights of the newly generated individual, expressed as:

[0220]

[0221] ″′

[0222] In the formula, W pi is the weight of the neural network corresponding to the individual after the mutation operation, W piis the weight of the neural network corresponding to the individual after crossover operation, is a normal distribution with a mean of 0 and a variance of τ 2 ; τ 2 is the variance of mutation, τ 2 is set to 0.04; b ″ pi is the bias of the neural network corresponding to the individual after mutation operation, ′

[0223] b pi is the bias of the neural network corresponding to the individual after crossover operation;

[0224] S325. Repeat iteration steps S322 to S324 until the preset stop iteration condition is met, which indicates that the model training is completed. The preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0225] S4. Use the training samples after feature extraction to train the classification model constructed by the depth neural decision tree classification algorithm based on quantum transformation pruning. After training, the trained classification model is obtained;

[0226] The dynamic pruning strategy of quantum transformation (referred to as quantum transformation pruning) is to evaluate the node utility through quantum transformation, dynamically prune inefficient nodes, thereby optimizing the structure of the decision tree, improving the classification accuracy and computational efficiency of the model, being able to reduce the complexity of the model, and also enhancing the generalization ability of the model, avoiding overfitting, and improving the overall performance. Training the classification model constructed by the depth neural decision tree classification algorithm based on quantum transformation pruning uses the depth neural decision tree classification algorithm based on quantum transformation pruning, and the training includes:

[0227] S401. Initialize the parameters of the depth neural decision tree, including the weights of the decision nodes and the depth of the tree. Let the weight of the i-th node in the depth neural decision tree be W u,i and the bias of the i-th node in the depth neural decision tree be b u,i , and the initialization method is expressed as:

[0228]

[0229] In the formula, is a normal distribution with a mean of 0 and a standard deviation of 1, and ~ means following a specific distribution;

[0230] S402. During the training process, the input geological data after feature extraction passes through multiple decision nodes of the decision tree. Each node makes a decision based on the weights and input features, guiding the geological data to flow to the left or right subtree until reaching the leaf node. Let the feature vector input to the depth neural decision tree be x u , which is expressed as:

[0231] y u,n = Sig(W u,n x u + b u,n ),

[0232] where y u,n is the output of the nth node of the deep neural decision tree; Sig() is the Sigmoid activation function, which is used to map the result of the linear combination to the interval (0, 1) to simulate the probability output of the decision; W u,n is the weight of the nth node in the deep neural decision tree, and b u,n is the bias of the nth node in the deep neural decision tree;

[0233] S403. Considering the classification accuracy and the complexity of the tree, use the loss function to evaluate the performance of the current decision tree. The loss function of the deep neural decision tree includes the classification error and the regularization term to prevent overfitting. The calculation formula is:

[0234]

[0235] where L u is the loss function of the deep neural decision tree, N u is the number of samples input to the deep neural decision tree in the current batch, t i is the target output of the ith sample, y u,i is the classification result of the deep neural decision tree for the ith sample, λ u is the regularization coefficient of the deep neural decision tree, and W u is the weight of the deep neural decision tree; ∥∥ is the L2 norm, which is used to punish the overly large weight values to improve the generalization ability of the model;

[0236] S404. In each training cycle, evaluate the utility and efficiency of each decision node according to the quantum transformation principle. The inefficient nodes will be pruned to reduce the model complexity while maintaining or improving the performance of the model. The present invention uses the quantum transformation principle to evaluate and adjust the structure of the network. Let P u,i represent the importance of the ith node. If P u,i is less than the preset threshold θ u , then this node will be removed from the model. The calculation formula of P u,i is:

[0237]

[0238] where L u is the loss function of the deep neural decision tree, W u,i is the weight of the ith node in the deep neural decision tree, ψ r () is the quantum transformation function, which is used to simulate the quantum state; Zr is the input feature vector of the current node of the deep neural decision tree;

[0239] ψ r (z r ) is calculated as follows:

[0240]

[0241] In the formula, Z r is the input feature vector of the current node of the deep neural decision tree, λ r is the parameter controlling the sensitivity of the quantum transformation, z rj is z r 's j-th element, is the quantum noise level corresponding to z rj ;

[0242] The quantum noise level has different compression levels in different parts of the feature representation, so as to more finely control the information loss. The calculation formula is:

[0243]

[0244] In the formula, α r is the coefficient for adjusting the quantum noise response, α r is set to 0.2; z rj is z r 's j-th element.

[0245] S405. Repeat the iterative steps S402 to S404 until the preset stop iteration condition is satisfied, which means the model training is completed. The preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0246] So far, the training of the mineral resource identification model based on the data collected by the drone is completed.

[0247] Steps S5 to S7 are a brief description of the application process of this mineral resource identification model.

[0248] S5. Real-time collect new real geological data of the mineral resource area through multiple sensors;

[0249] S6. Judge the data feature dimension of the training samples. When the data feature dimension is greater than or equal to 20, use the trained feature extraction model constructed by the autoencoder algorithm based on double constraints to extract features from the new real geological data; when the data feature dimension is less than 20, use the trained feature extraction model constructed by the neural network algorithm based on dynamic population evolution optimization to extract features from the new real geological data to obtain the new real geological data after feature extraction;

[0250] S7. Use the trained classification model to label the mineral resource categories of new real geological data.

Claims

1. A method for training a mineral resource identification model based on data collected by drones, characterized in that, Including: S1. When the drone flies in the mineral resource area, real geological data of the mineral resource area are collected in real time by multiple different types of sensors equipped on the drone, the real geological data are preliminarily processed and stored, and the mineral resource categories of the real geological data after preliminary processing are labeled. The labeling method is that professional geologists accurately label the real geological data by combining on-site samples in the mineral resource area to analyze the real geological data. S2. Use the real geological data after labeling to train a data generation model constructed by a generative adversarial network algorithm based on dynamic gradient adjustment. When training the data generation model constructed by the generative adversarial network algorithm based on dynamic gradient adjustment, use the generative adversarial network algorithm based on dynamic gradient adjustment to obtain a trained data generation model. Use the trained data generation model to generate simulated geological data, and use the simulated geological data and the real geological data after labeling as training samples together. Training the data generation model constructed by the generative adversarial network algorithm based on dynamic gradient adjustment includes: S201. Initialize the network parameters of the generator and discriminator of the generative adversarial network. S202. During the adversarial training process, the generator and discriminator are alternately trained. The generator generates fake geological data, and the discriminator judges whether the input comes from the real geological data set or the generator. In this way, the generator learns to generate geological data that is increasingly difficult to distinguish by the discriminator. During the adversarial training phase, the discriminator's judgment result D c (x) of the real samples and the discriminator's judgment result D c (G c (z)) are implemented through a neural network layer with a Sigmoid activation function to ensure that the output values are between 0 and 1, expressed as: where f c () is the linear output of the discriminator network, is the weight of the discriminator; S203. In each round of training, according to the feedback of the discriminator, dynamically adjust the learning rates of the generator and discriminator. When the accuracy of the discriminator is too high, increase the learning rate of the generator and decrease the learning rate of the discriminator, and vice versa. This dynamic adjustment helps to avoid mode collapse during the training process and ensure the diversity of geological data generation. The dynamic gradient adjustment is achieved by adjusting the learning rate, and the adjustment strategy is expressed as: Wherein, is the learning rate of the generator, η0 is the base learning rate, is the adjustment coefficient of the generator learning rate, || || is the L2 norm, is the gradient of the loss function of the generator with respect to its parameters, is the loss function of the generator, is the learning rate of the discriminator, is the adjustment coefficient of the discriminator learning rate, is the gradient of the loss function of the discriminator with respect to its parameters, is the loss function of the discriminator; S204. Use the self-similarity principle to guide the generator to pay attention to the local similarity of geological data, and improve the quality and usability of fake geological data. The self-similarity adjustment guides the model to attach importance to the self-similarity characteristics of local geological data by adding a regularization term to the loss function of the generator. The calculation formula of the regularization term is: R c (G c (z)) = β c ∑ i,j ||G c (z)i - G c (z) j || + s c (G c (z)) where, R c () is the self - similarity regularization term calculation function, G c (z) is the fake geological data generated by the generator, β c is the regularization coefficient of the generative adversarial network, G c (z) i is the i - th feature region in the fake geological data, G c (z) j is the j - th feature region in the fake geological data, s c () is the self - similarity descriptor calculation function; S205. Repeat steps S202 to S204 until the preset stop iteration condition is met. S3. Judge the data feature dimension of the training samples. When the data feature dimension is greater than or equal to 20, use the training samples to train a feature extraction model constructed by an autoencoder algorithm based on double constraints. When training the feature extraction model constructed by the autoencoder algorithm based on double constraints, use the autoencoder algorithm based on double constraints. When the data feature dimension is less than 20, use a feature extraction model constructed by a neural network algorithm based on dynamic population evolution optimization. When training the feature extraction model constructed by the neural network algorithm based on dynamic population evolution optimization, use the neural network algorithm based on dynamic population evolution optimization. Obtain a trained feature extraction model and the training samples after feature extraction. S4. Use the training samples after the above feature extraction to train a classification model constructed by a deep neural decision tree classification algorithm based on quantum transformation pruning. When training the classification model constructed by the deep neural decision tree classification algorithm based on quantum transformation pruning, use the deep neural decision tree classification algorithm based on quantum transformation pruning to obtain a trained classification model.

2. The method for training a mineral resource identification model based on data collected by an unmanned aerial vehicle according to claim 1, wherein In the step S202, the said calculation formulas are respectively: wherein, are the weights of the discriminator network, are the biases of the discriminator network, x i is the i-th characteristic attribute of the real geological data sample, G c,i (z) is the i-th output of the generator.

3. The method for training a mineral resource identification model based on data collected by an unmanned aerial vehicle according to any one of claims 1 to 2, characterized in that The feature extraction model constructed by training the dual-constraint autoencoder algorithm in S3 includes: S311. Geological data passes through the autoencoder from the input layer through multiple hidden layers. Each layer performs a linear transformation and a non-linear activation, and finally outputs the extracted features. The transfer function of each layer is expressed as: x p,i+1 = Sig(W p,i x p,i + b p,i ), where x p,i+1 is the output vector of the (i + 1)-th layer of the autoencoder, Sig() is the Sigmoid activation function, W p,i is the weight matrix of the i-th layer of the autoencoder, x p,i is the input vector of the i-th layer of the autoencoder, and b p,i is the bias vector of the autoencoder; S312. The loss function of the autoencoder is designed as a dual-constraint loss. One part comes from the reconstruction loss of the geological data content to ensure that the extracted features can effectively reconstruct the original geological data; the other part is the structure-preserving loss to ensure that the original structure of the geological data is maintained in the feature space. The calculation formula of the loss function of the autoencoder is: L p = λ p1 L p,recon + λ p2 L p,struct , where L p is the loss function of the autoencoder, λ p1 is the weight coefficient for adjusting the relative importance of the reconstruction loss, L p,recon is the reconstruction loss, λ p2 is the weight coefficient for adjusting the relative importance of the structure preservation loss, L p,struct is the structure preservation loss; S313. Calculate the gradient according to the loss function and use the least squares gradient descent method to update the network parameters. Consider the stability of the gradient and the adaptability of the update rate to avoid the problem of gradient explosion or disappearance in the high-dimensional space. The parameter update rule is: where θ p,i is the parameter of the i-th layer of the autoencoder, ← is the parameter update operation, and η p is the learning rate of the autoencoder, is the loss function L of the autoencoder p is the gradient of the parameter of the i-th layer of the autoencoder; S314. Repeat steps S311 to S313 until the preset stop iteration condition is met.

4. The method for training a mineral resource identification model based on data collected by an unmanned aerial vehicle according to claim 3, wherein The reconstruction loss L in S312 p,recon has the following calculation formula: where x p is the original input geological data, is the output of the autoencoder; The structural retention loss L p,struct is calculated by the formula: L p,struct = ∑ j,k W p,jk ||z p,j -z p,k || 2 , where, W p,jk is the similarity matrix based on the input geological data, z p,j is the j-th geological data point mapped to the feature space, z p,k is the k-th geological data point mapped to the feature space.

5. The method for training a mineral resource identification model based on data collected by an unmanned aerial vehicle according to any one of claims 1 to 2, characterized in that The feature extraction model constructed by training the neural network algorithm based on dynamic population evolution optimization in S3 includes: S321. According to the initialization method of the bionic algorithm, in the initialization stage, an initial population is generated, where each individual represents a configuration of network weights. Let the population size be N p , initialize the weights and biases for the i-th individual, denoted as: In the formula, is the weight matrix of the i-th individual in the initial state, and ~ follows a specific distribution. is a normal distribution. is a normal distribution with a mean of 0 and a variance of σ 2 . σ 2 is the initialized variance. is the bias of the i-th individual in the initial state; W pi is the weight of the neural network corresponding to the i-th individual, and b pi is the bias of the neural network corresponding to the i-th individual. S322. For each individual in the population, use its corresponding neural network configuration to process the input training data, calculate the output of the model, and evaluate its performance according to a predetermined loss function. For the i-th individual, calculate the loss on the training data set using its weights and biases, which is expressed as: Where L pi is the loss of the neural network corresponding to the i-th individual, mps is the number of samples in the current batch input, l() is the composite loss function, and fsig() is the neural network model function, is the feature of the j-th geological data sample, and W pi is the weight of the neural network corresponding to the i-th individual, and b pi is the bias of the neural network corresponding to the i-th individual, is the label of the j-th geological data sample; S323. According to the fitness of the individual, select the best-performing individual from the current population to be retained as a candidate solution for the next generation. Selection is based on the fitness of the individual, and excellent individuals have a higher probability of being selected. The calculation formula for the probability of an individual being selected is: where P select (i) is the probability that the i-th individual is selected, and γ pse is the parameter for controlling the selection pressure, L pi is the loss of the neural network corresponding to the i-th individual, N p is the population size, and L pk is the loss of the neural network corresponding to the k-th individual; S324. Generate new individuals through crossover and mutation operations. The crossover operation allows two excellent individuals to exchange part of their genes to generate new offspring; the mutation operation randomly changes part of the genes in an individual to increase the diversity of the population. The crossover operation randomly selects two individuals for gene exchange, which is expressed as: W′ pi = α pcs W p1 +(1 - α pcs )W p2 , b′ pi = α pcs b p1 +(1 - α pcs )b p2 , Where, W′ pi is the weight of the neural network corresponding to the individual after the crossover operation, α pcs is the crossover rate, W p1 is the weight of the neural network corresponding to the first selected individual, W p2 is the weight of the neural network corresponding to the second selected individual, b′ pi is the bias of the neural network corresponding to the individual after the crossover operation, b p1 is the bias of the neural network corresponding to the first selected individual, b p2 is the bias of the neural network corresponding to the second selected individual; The mutation operation performs a small random perturbation on the weights of the newly generated individual, which is expressed as: Where, W″ pi is the weight of the neural network corresponding to the individual after the mutation operation, W′ pi is the weight of the neural network corresponding to the individual after the crossover operation, is a normal distribution with a mean of 0 and a variance of τ 2 , τ 2 is the variance of the mutation, b″ pi is the bias of the neural network corresponding to the individual after the mutation operation, b′ pi is the bias of the neural network corresponding to the individual after the crossover operation; S325. Repeat steps S322 to S324 until the preset stop iteration condition is met.

6. The method for training a mineral resource identification model based on data collected by an unmanned aerial vehicle according to claim 5, wherein In S322, the calculation formula is as follows: where \(l()\) is the composite loss function, \(fsig()\) is the neural network model function, is the feature of the \(j\)-th geological data sample, \(W\) pi is the weight of the neural network corresponding to the \(i\)-th individual, \(b\) pi is the bias of the neural network corresponding to the \(i\)-th individual, is the label of the \(j\)-th geological data sample, \(MSE()\) is the mean square error function, \(\lambda\) ps is the regularization parameter, \(Reg(W\) pi ) is the regularization term; The regularization term Reg(W pi ) has the following calculation formula: Where, W pi,k is the k-th weight of the neural network corresponding to the i-th individual.

7. The method for training a mineral resource identification model based on data collected by an unmanned aerial vehicle according to any one of claims 1 to 2, characterized in that The training of the classification model constructed by the deep neural decision tree classification algorithm based on quantum transformation pruning in S4 includes: S401. Initialize the parameters of the deep neural decision tree, including the weights of the decision nodes and the depth of the tree. Let the weight of the $i$-th node in the deep neural decision tree be $W$ u,i , and the bias of the $i$-th node in the deep neural decision tree be $b$ u,i . The initialization method is expressed as: wherein, is a normal distribution with a mean of 0 and a standard deviation of 1, and ~ means subject to a specific distribution; S402. During the training process, the geological data after input feature extraction passes through multiple decision nodes of the decision tree. Each node makes a decision based on the weights and input features, guiding the geological data to flow to the left or right subtree until reaching the leaf node. Let the feature vector input to the deep neural decision tree be x u , which is expressed as: y u,n = Sig(W u,n x u + b u,n ) where y u,n is the output of the n-th node of the deep neural decision tree; Sig() is the Sigmoid activation function; W u,n is the weight of the n-th node in the deep neural decision tree, and b u,n is the bias of the n-th node in the deep neural decision tree; S403. Considering the classification accuracy and the complexity of the tree, use a loss function to evaluate the performance of the current decision tree. The loss function of the deep neural decision tree includes classification error and a regularization term to prevent overfitting. The calculation formula is: where L u is the loss function of the deep neural decision tree, N u is the number of samples input to the deep neural decision tree in the current batch, t i is the target output of the i-th sample, y u,i is the classification result of the deep neural decision tree for the i-th sample, λ u is the regularization coefficient of the deep neural decision tree, W u is the weight of the deep neural decision tree; || || is the L2 norm; S404. In each training cycle, evaluate the utility and efficiency of each decision node according to the principle of quantum transformation. Inefficient nodes will be pruned to reduce the model complexity while maintaining or improving the model performance. The structure of the network is evaluated and adjusted using the principle of quantum transformation. Let P u,i represent the importance of the i-th node. If P u,i is less than the preset threshold θ u , then this node will be removed from the model. The calculation formula of P u,i is: Where, L u is the loss function of the deep neural decision tree, W u,i is the weight of the i-th node in the deep neural decision tree, ψ r () is the quantum transformation function; Z r is the input feature vector of the current node of the deep neural decision tree; S405. Repeat steps S402 to S404 until the preset stop iteration condition is met, which means the model training is completed.

8. The method for training a mineral resource identification model based on data collected by an unmanned aerial vehicle according to claim 7, wherein The ψ described in S404 r (z r ) has the following calculation formula: where, Z r is the input feature vector of the current node of the deep neural decision tree, λ r is the parameter for controlling the sensitivity of the quantum transformation, Z rj is the j-th element of Z r , is the quantum noise level corresponding to Z rj ; The quantum noise level has different compression levels in different parts of the feature representation, thus more finely controlling the information loss. The calculation formula is as follows: where α r is the coefficient for adjusting the quantum noise response, and Z rj is the j-th element of Z r .

Images