Method for detecting content of iron element in iron ore

Abstract

The invention belongs to the technical field of mineral product detection, and particularly relates to an iron ore iron element content detection method which is used for iron ore detection and comprises the steps that 1, an unmanned aerial vehicle is used for carrying out detection path planning, and communication is carried out in cooperation with Internet of Things data in a region so as to reach a designated cargo position; step 2, acquiring spectral data and performing data processing; 3, performing spectral line identification and classification by using a machine learning model; and 4, improving the learning model, on one hand, improving the efficiency and reducing the time consumed by manpower, and on the other hand, learning through the machine model, and improving the accuracy and efficiency of ore detection, so that the requirements of commodity transactions are better met, and the efficiency of production activities is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of mineral detection, and particularly relates to a method for detecting the iron content of iron ore. Background Art

[0002] Iron ore grade testing is a vital part of commodity trading. Testing can ensure that the grade of the iron ore being traded meets the contract requirements and avoid trade fraud and economic losses. However, the current port iron ore grade testing has great limitations, including:

[0003] Chemical analysis: This is the most mature detection method. Its advantages are that the technology is very mature and the detection results are highly accurate. However, it has disadvantages such as the operation process takes a long time, the results are slow to come out, and the detection cost is high.

[0004] X-ray fluorescence spectrometry (XRF): This is a fast, non-destructive detection method that can determine multiple elements at the same time. It determines the type and content of elements in the sample by measuring the fluorescence response of the sample to X-rays. However, the equipment cost is high and the particle size and uniformity of the sample are required to be high.

[0005] Neutron activation analysis: This is a highly sensitive method that can detect a wide range of elements, but it requires specialized equipment and radioactive isotopes and has limited use.

[0006] Laser-induced fluorescence spectroscopy: Suitable for trace analysis of high-purity materials, but may be affected by the physical state of the sample.

[0007] Laser-induced breakdown spectroscopy (LIBS): This is a method capable of remote analysis and is suitable for rapid on-site detection. It can perform rapid and accurate elemental analysis on solid samples without the need for sample pretreatment. However, the volume of iron ore piled up at the port is large, and the manual use of laser-induced breakdown spectroscopy has great limitations in sampling.

[0008] Therefore, these current detection methods have limitations. In order to improve detection efficiency and accuracy, a new and intelligent method is needed for monitoring. As an emerging intelligent technology, drones equipped with laser induced breakdown spectroscopy (LIBS) have broad application prospects in port iron ore grade detection. Summary of the invention

[0009] The purpose of the present invention is to provide a method for detecting the iron content of iron ore, so as to solve the problems mentioned in the background technology.

[0010] In order to achieve the above technical objectives, the technical solution adopted by the present invention is as follows:

[0011] A method for detecting the iron content of iron ore, characterized in that it specifically comprises the following steps:

[0012] Step 1: Use drones to plan inspection routes and communicate with IoT data in the area to reach the designated cargo location;

[0013] Step 2: Acquire spectral data and process the data;

[0014] Step 3: Use machine learning models to identify and classify spectral lines;

[0015] Step 4: Improve the learning model.

[0016] The specific steps of step 1 are: the drone takes off from its predetermined starting position, defines the path coverage by drawing the path of the drone on the map, and communicates with the IoT devices along the way to avoid restricted areas or possible obstacles to reach the designated cargo location area, wherein the drone model is DJI Mavic 3Pro drone, and the path planning method used is:

[0017] (1) Establish a model for drone takeoff and landing, no-fly zones, and obstacle-marking wireless networks:

[0018] (2) Constructing the POMDP decision model:

[0019] ① Establish arrays and general reward functions;

[0020] ②Use dual deep Q networks DDQNs to solve the remaining problems of POMDP;

[0021] ③Use the RNN model to realize information transmission between different time steps;

[0022] ④Use attention LSTM to overcome the limitations of the original encoding and decoding model on fixed vectors;

[0023] ⑤Construct loss function.

[0024] The specific steps of step 2 are:

[0025] ① Using a DJI Mavic 3Pro drone equipped with an American SENTERA 6X multispectral imager, the spectrometer wavelength range is 20nm to 840-nm for spectral data collection;

[0026] ② Construct sample set partitioning: randomly divide the spectral sample data set into training set and test set. The training set sample data is used for model training. The model parameters are optimized through iteration to obtain the model with the best prediction accuracy. The test set sample data is used to evaluate the model. The standard spectral data comes from the atomic spectral database provided by the National Institute of Standards and Technology of the United States.

[0027] ③Preprocess the spectral data

[0028] Ⅰ. Perform baseline correction;

[0029] Ⅱ. Perform data smoothing and denoising.

[0030] The specific steps of step three are:

[0031] (1) Construct a spectral data conversion model;

[0032] (2) Build a network model;

[0033] ①Construct the optimization function of convolutional neural network;

[0034] ②Build a channel attention mechanism;

[0035] ③Weight each channel to enhance important features.

[0036] The specific steps of step 4 are:

[0037] Atrous spatial pyramid pooling and Ghost module are used to optimize the model. The Ghost module is used to reduce the amount of calculation and parameters while retaining important feature information. Then, the atrous spatial pyramid pooling operation is applied to the feature map output by the Ghost module to extract multi-scale contextual information. Finally, the feature vector obtained by atrous spatial pyramid pooling is used for classification tasks. Such a fusion can make full use of the contextual information extraction capability of atrous spatial pyramid pooling while keeping the model lightweight, thereby improving the performance of the model in image classification.

[0038] Compared with the prior art, the present invention has at least the following advantages:

[0039] Using the method of the present invention to test iron ore can, on the one hand, improve efficiency and reduce manual time, and on the other hand, can learn through machine models to improve the accuracy and efficiency of ore testing, thereby better meeting the needs of commodity trading and improving the efficiency of production activities. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention is further illustrated by means of the non-limiting examples given in the accompanying drawings.

[0041] Figure 1 Schematic diagram of the calculation process of step one of the present invention.

[0042] Figure 2 Schematic diagram of the calculation process of step three of the present invention.

[0043] Figure 3 FIG. 4 is another schematic diagram of the calculation flow of step three of the present invention. DETAILED DESCRIPTION

[0044] In order to enable those skilled in the art to better understand the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0045] like Figure 1-3 As shown, a method for detecting the iron content of iron ore is characterized in that it specifically comprises the following steps:

[0046] Step 1: Use drones to plan inspection routes and communicate with IoT data in the area to reach the designated cargo location;

[0047] Step 2: Acquire spectral data and process the data;

[0048] Step 3: Use machine learning models to identify and classify spectral lines;

[0049] Step 4: Improve the learning model.

[0050] The specific steps of step 1 are: the drone takes off from its predetermined starting position, defines the path coverage by drawing the path of the drone on the map, and communicates with the IoT devices along the way to avoid restricted areas or possible obstacles to reach the designated cargo location area, wherein the drone model is DJI Mavic 3Pro drone, and the path planning method used is:

[0051] (1) Establish a model for drone takeoff and landing, no-fly zones, and obstacle-marking wireless networks:

[0052] Define and set up a grid-based environment model, treating the environment as a square grid of size G×G∈N 2 , each cell size is c, g and N represent the set of available potential locations and the set of natural numbers respectively, and the annotated environment may include specified no-fly zones, landing points and obstacles:

[0053] The specified starting / landing position coordinates of the L environment are:

[0054]

[0055] The Z position coordinates of the drone’s no-fly zone are expressed as:

[0056]

[0057] The coordinates of the B obstacles that block the wireless network are:

[0058]

[0059] (2) Constructing the POMDP decision model:

[0060] ① Establish arrays and general reward functions;

[0061] Array (S p , A, Z, R, T, O, γ) is used to describe POMDP, where S p represents the state space, A represents the action space, Z:S p ×A×S p →R maps the probability function of the transition. The reward function is R×S p ×A×S p →R, which also provides a real-value reward for each state, action, and subsequent state triple combination. p →T represents the observation function, and T defines the observation space. The discount factor (γ) determines the importance of long-term and short-term returns, γ∈[0,1].

[0062] The state formula for drone coverage path planning is:

[0063] S p =B G×G×3 ×R G×G ×N 2 ×N

[0064] B G×G×3 For a given environment map, R G×G is the target area map, N 2 is the position of the UAV, and N is the flight time of the UAV.

[0065] State space s(t)∈S p The elements are:

[0066] s(t)=(G,D(t),P(t),E b (t))

[0067] The four elements of the state space element are as follows:

[0068] G is the environment map, which contains all information such as takeoff, landing, obstacles and no-fly zones.

[0069] D(t) serves as a target map, showing the remaining information of the device locations or cells that will be exposed at time t.

[0070] P(t) represents the position of the drone at a specific time t

[0071] E b (t) is the battery energy level of the drone at time t.

[0072] The action space of the drone a(t)∈A f ,A f =[N, E, S, W, H, L].

[0073] Here the four directions are N, E, S, and W. H and L are the hovering direction and landing action of the drone. The following arrays constitute the general reward function R(s(t), a(t), s(t+1)):

[0074] r c (+ve): The difference between s(t+1) and s(t) in terms of the number of newly covered target areas or the area coverage that could have been achieved with the acquired data or the data collection reward.

[0075] r s c(-ve): The safety controller will be penalized if the drone must be prevented from hitting an obstacle or entering the NFZ.

[0076] r m ov(-ve): Every action of the drone results in a continuous loss of maneuverability, and these actions do not lead to mission accomplishment.

[0077] r cras h(-ve): If the drone does not land safely in the landing zone before the remaining flight time expires, a penalty will be imposed.

[0078] ②Use dual deep Q networks DDQNs to solve the remaining problems of POMDP;

[0079] This step uses dual deep Q networks (DDQNs) to solve the above POMDP problem. The Q value of each state-action combination is expressed as:

[0080]

[0081] Next, the optimal Q value is achieved. The first Q network is obtained by reducing two parameters φ and The Q network is updated by the loss caused by the experience replay memory.

[0082] Loss(φ)=ε s，a，s′ ~D[(Q φ (s, a)-γ(s, a, s′)) 2 ]

[0083] The target value can be expressed as:

[0084]

[0085] The second network parameter can be updated by the following equation:

[0086]

[0087] ③ Use the RNN model to realize information transmission between different time steps. This step uses the RNN network model. The basic unit of RNN contains one or more recursive connections, which allows the network to transmit information between different time steps. At each time step, RNN receives a data input, and combines the current input data and the hidden state at the previous time point to refresh its internal hidden state, and then generates the corresponding output. This process will be recursively performed in the entire sequence.

[0088] In order to solve the problems that occur in the standard RNN training process, this step introduces the following improvements: LSTM, GRU, LSTM-based Bi-LSTM, and Bi-GRU networks.

[0089] ④Use attention LSTM to overcome the limitations of the original encoding and decoding model on fixed vectors;

[0090] LSTM consists of three key gate units: input gate, forget gate and output gate, which together determine the state of the memory unit c t The update and final hidden state output of these gating units are generated by the sigmoid function, which generates values ​​∈ (0, 1), representing the degree of acceptance of new information, the degree of forgetting old information, and the degree of exposure of the memory unit state to the output.

[0091] At time point t, based on the input X t and the previous output h t-1 , candidate unit status and the actual cell state c t The calculation method is as follows:

[0092] f t =σ(W f x t +U f h t-1 +b f )

[0093] i t =σ(W i x t +U i h t-1 +b i )

[0094]

[0095] where x=(x1,...,x n ) is the input vector, where x t (for t=1,...,n) is the data point at time t in a sequence of length n. t is the candidate cell state, h tis the output of the unit, also called the hidden state. t ,i t , o t are the values ​​of the forget, input, and output gates, respectively. f , W i , W o , is the weight matrix associated with the input x. f , U i , U o , is the same as the output h t The associated weight matrix. b f , b i , b o , is the bias vector. Finally, the output h t :

[0096] h t =o t tanh(c t )

[0097] This step uses the attention method to overcome the limitations of the original encoding and decoding model on fixed vectors. The main steps of the attention mechanism are as follows: the output of LSTM [h1, h2, h3, ..., h n ] is obtained through nonlinear transformation [u1, u2, u3, …, u n ]. More influential parameters should be given more weight. The attention mechanism generates an attention weight matrix [1, 2, 3, …, n], which represents the importance of each intermediate condition. Finally, the vector V is obtained by performing a weighted sum of the input parameters and the weights. Decoding the vector V outputs y as follows:

[0098] u k =tanh(W k h k +b k )

[0099]

[0100] where u k is the unnormalized attention score of the kth head. k h k is the weighted key-query product of the kth head. b k is the bias vector of the kth head. k is the attention weight of the k-th head. is the unnormalized self-attention score of the k-th head. i is the unnormalized self-attention score of the ith head. K is the number of heads. h kis the hidden state of the kth head, i.e. the output of LSTM. V is the attention output.

[0101] ⑤Construct loss function.

[0102] The loss function of this step adopts the concept of DDQN (Double Deep Q-Network). It consists of two steps: First, use the main network or main network parameters (φ i ), determine the action that maximizes the Q value, denoted by max a Q(S,a′,φ i ).

[0103] Then, the corresponding action is determined in the target network. This Q value is used to calculate Q target value.

[0104] The mathematical expression of the loss function used by ARDDQN (Advanced Double Deep Q-Network) is:

[0105] Loss = ε{R + γ·Q target (s′, arg max Q main (s′, a)-Q main (s,a)) 2}

[0106] The specific steps of step 2 are:

[0107] ① Using a DJI Mavic 3Pro drone equipped with an American SENTERA 6X multispectral imager, the spectrometer wavelength range is 20nm to 840-nm for spectral data collection;

[0108] ② Construct sample set partitioning: randomly divide the spectral sample data set into training set and test set. The training set sample data is used for model training. The model parameters are optimized through iteration to obtain the model with the best prediction accuracy. The test set sample data is used to evaluate the model. The standard spectral data comes from the atomic spectral database provided by the National Institute of Standards and Technology of the United States.

[0109] ③Preprocess the spectral data

[0110] Ⅰ. Baseline correction: This step uses a correction baseline based on polynomial fitting. First, all the minimum points in the multispectral data are defined as the main minimum points, and the data is divided into multiple bands. The minimum point in each band is determined as the characteristic value point, and then the characteristic value site of each level is fitted by spline difference and connected. The specific implementation method is as follows:

[0111] The first step is to divide the spectral data into N parts according to the wavelength range, where the wavelength range of the spectral data is recorded as wavelength, and number is the number of points after being divided into N parts;

[0112]

[0113] In each segmented part, the lowest point of the spectral intensity is identified and defined as the main minimum point as the characteristic value point of the band:

[0114] min{A(α i )} j , j = 1, 2, 3 ... n

[0115] The characteristic value points of each band are strung together, and then the curves generated by these nodes are smoothed by spline interpolation technology, thus the baseline function f can be obtained. baseline Finally, by converting the original spectral data spec org With baseline data f baseline Perform differential calculation to obtain the corrected spectral data spec correct , specifically expressed as:

[0116] spec correct =spec org -f baseline

[0117] II. Data smoothing and denoising: In this step, the db2 wavelet function is used to decompose the laser-induced breakdown spectroscopy data, and the soft threshold method is used to filter out high-frequency noise signals. The soft threshold function expression used is:

[0118]

[0119] Where: λ is the given threshold; ω i，j is the wavelet coefficient; is the new wavelet coefficient obtained after soft threshold processing;

[0120] The two-dimensional wavelet is decomposed in the two-dimensional plane to obtain three detail components. For example, the original data S is decomposed into three detail components, including the horizontal component Vertical Component Diagonal component The process of two-dimensional wavelet denoising is as follows:

[0121] a. Input 2D raw LIBS data;

[0122] b. Confirm the wavelet basis function, decomposition level, and threshold;

[0123] c. Decompose the original data into four components, namely, approximate component, horizontal component, vertical component, and diagonal component;

[0124] d. Further decompose the approximate components obtained in the previous step into 4 components, and so on, until the predetermined number of decomposition levels is reached;

[0125] e. Use the threshold to denoise the coefficients obtained from each layer of decomposition;

[0126] f. Reconstruct the denoised wavelet components into new two-dimensional spectral data;

[0127] In this step, the db2 wavelet function is used as the wavelet basis function. During the operation, the optimal decomposition layer number is determined to be 3, and the coefficients obtained after decomposition are subjected to soft threshold processing. The wdencmp function provided by the MATLAB software is used to select the threshold to obtain the spectral data after noise reduction.

[0128] The specific steps of step three are:

[0129] (1) Construct a spectral data conversion model;

[0130] Construct a spectral information matrix and convert the spectral data vector of each sample into a two-dimensional spectral information matrix. The specific processing method is as follows:

[0131] Let x be the spectral data vector, and the two-dimensional spectral information matrix is:

[0132] S=xx T

[0133] Assume x is the three-dimensional spectral data column vector, and the two-dimensional spectral information matrix is:

[0134]

[0135] (2) Build a network model, where the steps use the Inception-ResNet v2 algorithm;

[0136] ①Construct the optimization function of convolutional neural network;

[0137] First, we construct the activation function of the convolutional neural network. The activation function can usually increase the nonlinear mapping ability of the network. The CNN model with the activation function can better solve some difficult problems. This paper uses the hyperbolic tangent (hereinafter referred to as tanh) function.

[0138] The expression of tanh function is:

[0139]

[0140] In the process of error back propagation, the gradient descent optimization algorithm is used to update the weights and other parameters. m ,y m )}, for each sample x, the predicted value y can be obtained by y = p(x, θ), where p(·) is the decision function and θ is the parameter of the model. Therefore, f i (θ) represents the loss function of the i-th sample with respect to parameter θ, f i (θ) is the sum of p(·) and y i The constructed function.

[0141] Therefore, the average loss function of the training set samples is:

[0142]

[0143] The purpose of model optimization is to minimize the average loss function, so it is necessary to find a suitable θ.

[0144] This step uses the Adam optimizer, which requires setting three parameters: the initial learning rate α, which determines the update amplitude in the gradient descent step; the first-order moment attenuation coefficient β1, and the second-order moment attenuation coefficient β2, which are used to control the exponential decay rate of the first-order moment estimate of the gradient and the exponential decay rate of the second-order moment estimate of the gradient, respectively. Calculate the first moment m of the gradient g(t) t and the second moment v t . m t The calculation method is:

[0145] m t =β1·m t-1 +(1-β1)·g t

[0146] v t The calculation formula is:

[0147]

[0148] For m t and v t To correct the deviation,

[0149]

[0150] Adam's parameter update formula is:

[0151]

[0152] ②Build a channel attention mechanism;

[0153] In LIBS spectral analysis, spectral data usually contains rich chemical composition information, but these data are also accompanied by noise and complex background interference. The spectra of some elements may not be important, while the spectra of some key elements such as iron are more important. In order to solve the above problems, the channel attention mechanism can be used. It is proposed to integrate SENet (Squeeze-and-Excitation Networks) into the model to form the Inception-ResNet-C module.

[0154] The Inception-ResNet-C module consists of three branches:

[0155] The first branch: This branch is directly output without processing, usually to implement residual connection, after the connection, the input can be added to the output of the subsequent layer. This can introduce short connections in the network, help the gradient flow, and reduce the problem of gradient disappearance.

[0156] Second branch: This branch undergoes a 1x1 convolution, and the number of output channels is 192. 1x1 convolution is usually used for dimensionality reduction and overfitting control while reducing the amount of computation.

[0157] The third branch: 1x1 convolution to reduce the dimension, 3x3 convolution, and then another 3x3 convolution, the number of output channels are 192, 224, and 256 respectively.

[0158] Afterwards, the output feature maps of the three branches of Inception-ResNet-C will be sent to the SENet module for processing. The SENet module performs adaptive channel weighting on these feature maps through the process of compression, excitation, and double calibration, further improving the expressiveness of features and achieving more effective feature selection and recalibration.

[0159] ③Weight each channel to enhance important features.

[0160] To obtain attention in the channel dimension, the feature map X is taken as input and passed through F tr The output after the convolution structure is U, F sq It represents the compression operation, which compresses the spatial feature dimension to 1×1×C and is represented by the symbol Z.

[0161] Next, proceed to F ex The excitation operation uses a fully connected layer and a nonlinear activation function to establish connections between channels, and then obtains normalized weights through a Sigmoid activation function. This operation aims to capture the interdependence between channels. The channel weights are then multiplied by the corresponding channels of the feature map one by one, thereby weighting each channel of the original feature map and strengthening important features. This can be expressed as the following formula:

[0162]

[0163] s=F ex (z, W)=σ(g(z, W))=σ(W2δ(W1z))

[0164]

[0165] The specific steps of step 4 are:

[0166] Atrous spatial pyramid pooling and Ghost module are used to optimize the model. The Ghost module is used to reduce the amount of calculation and parameters while retaining important feature information. Then, the atrous spatial pyramid pooling operation is applied to the feature map output by the Ghost module to extract multi-scale contextual information. Finally, the feature vector obtained by atrous spatial pyramid pooling is used for classification tasks. Such a fusion can make full use of the contextual information extraction capability of atrous spatial pyramid pooling while keeping the model lightweight, thereby improving the performance of the model in image classification.

[0167] Based on this, this step integrates the introduced Ghost module with the atrous spatial pyramid pooling to obtain the GP module (i.e., the module integrated with the Ghost module and the atrous spatial pyramid pooling (ASPP)).

[0168] The Ghost module is implemented by the following steps:

[0169] Generate intrinsic feature maps: Use 1x1 convolution to generate intrinsic feature maps. These feature maps are small in number but contain the basic feature information of the input data.

[0170] Generating ghost feature maps: Using depthwise separable convolutions or other linear transformations, additional ghost feature maps are generated from the intrinsic feature maps. These feature maps increase the receptive field of the network with relatively low computational cost.

[0171] Splicing feature map: Splice the intrinsic feature map and the ghost feature map together to form the final output feature map. Through this feature map fusion strategy, the key information of the input data can be captured and expressed more comprehensively.

[0172] Then, the rich feature map output by the Ghost module is fed into the ASPP module.

[0173] The ASPP module is implemented through the following steps:

[0174] Multi-scale dilated convolution: Dilated convolution kernels with different dilation rates are used in parallel to capture contextual information of different scales.

[0175] Global average pooling: Extract global context information through the global average pooling layer and the 1x1 convolution layer.

[0176] Feature fusion: The dilated convolution outputs with different dilation rates in the ASPP module and the results of global average pooling are concatenated in the channel dimension, and feature fusion is performed through a 1x1 convolution layer to form the final multi-scale feature representation.

[0177] By using the Ghost module and ASPP, the model's ability to recognize key spectral features can be enhanced, and its robustness to noise and interference can be improved, thereby improving recognition accuracy. Specifically, the Ghost module can increase the model's sensitivity to spectral nuances by adding additional feature maps, while ASPP can improve the model's ability to recognize objects of different sizes and shapes through multi-scale feature fusion. The application of these two technologies can help the LIBS spectral recognition model better handle complex situations in practical applications and improve the accuracy and efficiency of analysis.

[0178] Finally, the fused feature map is passed through a global average pooling layer to convert it into a one-dimensional feature vector. This feature vector is then sent to a fully connected layer for final classification to predict the iron content in the iron ore sample.

[0179] Example:

[0180] Step 1

[0181] (1) Define and set up a grid-based environment model. Consider the environment as a square grid of size G×G∈N. 2 , each cell size is c, g and N represent the set of available potential locations and the set of natural numbers respectively. The annotated environment may include specified no-fly zones, landing points and obstacles.

[0182] (2) Adopting partially observable Markov decision process.

[0183] ① Initialize, then use the epsilon-greedy strategy based on DDQN and LSTM prediction to select actions, execute actions, observe the next state s', reward r, and completion flag done. Store (s, a, r, s', done) in the experience replay buffer D. Randomly extract a small batch of samples from D: minibatch.

[0184] ② Use DDQN and LSTM target network to calculate the target Q value.

[0185] ③ Apply the attention mechanism to combine the outputs of DDQN and LSTM.

[0186] ④Calculate DDQN and LSTM losses.

[0187] ⑤ Use back propagation to update the DDQN and LSTM networks and update the DDQN and LSTM network parameters.

[0188] ⑥ Update the target network regularly: If the current cycle has not ended, continue to the next time step. When all cycles are completed, the training is completed.

[0189] ⑦Deploy the tested system onto drone hardware and operate it in a real environment.

[0190] Example: Step 1:

[0191] (1) Define and set up a grid-based environment model. Consider the environment as a square grid of size G×G∈N. 2 , each cell size is c, g and N represent the set of available potential locations and the set of natural numbers respectively. The annotated environment may include specified no-fly zones, landing points and obstacles.

[0192] (2) Adopting partially observable Markov decision process.

[0193] ① Initialize, then use the epsilon-greedy strategy based on DDQN and LSTM prediction to select actions, execute actions, observe the next state s', reward r, and completion flag done. Store (s, a, r, s', done) in the experience replay buffer D. Randomly extract a small batch of samples from D: minibatch.

[0194] ② Use DDQN and LSTM target network to calculate the target Q value.

[0195] ③ Apply the attention mechanism to combine the outputs of DDQN and LSTM.

[0196] ④Calculate DDQN and LSTM losses.

[0197] ⑤ Use back propagation to update the DDQN and LSTM networks and update the DDQN and LSTM network parameters.

[0198] ⑥ Update the target network regularly: If the current cycle has not ended, continue to the next time step. When all cycles are completed, the training is completed.

[0199] ⑦Deploy the tested system onto drone hardware and operate it in a real environment.

[0200] Step 2:

[0201] (1) Data acquisition: A DJI Mavic 3Pro drone equipped with a US SENTERA 6X multispectral imager was used to capture multispectral images at the iron ore pile sampling point at the port.

[0202] (2) Data processing:

[0203] ①Baseline correction

[0204] First, the spectrum data is evenly divided into N parts according to the wavelength range, where the wavelength range of the spectrum data is recorded as wavelength, and number is the number of points after being divided into N parts. In each divided part, the lowest point of the spectrum intensity is identified and defined as the main minimum point as the characteristic value point of the band. The characteristic value points of each band are strung together, and then the curves generated by these nodes are smoothed by spline interpolation technology to obtain the baseline function f baseline Finally, by converting the original spectral data spec org With baseline data f baseline Perform differential operation to obtain the corrected spectral data spec correct .

[0205] ② Smoothing denoising

[0206] The db2 wavelet function is used to decompose the LIBS data, and the soft threshold method is used to filter out high-frequency noise signals. The two-dimensional wavelet is decomposed in the two-dimensional plane to obtain three detail components, such as Figure 1 As shown: The original data S is decomposed into three detail components, including the horizontal component Vertical Component Diagonal component

[0207] The process of two-dimensional wavelet denoising is as follows:

[0208] i. Input 2D raw LIBS data;

[0209] ii. Confirm the wavelet basis function, decomposition level, and threshold;

[0210] iii. Decompose the original data into four components, namely, approximate component, horizontal component, vertical component, and diagonal component;

[0211] iv. Further decompose the approximate components obtained in the previous step into 4 components, and so on, until the predetermined number of decomposition levels is reached;

[0212] v. Use the threshold to denoise the coefficients obtained from each layer of decomposition;

[0213] vi. Reconstruct the denoised wavelet components into new two-dimensional spectral data.

[0214] In this step, the db2 wavelet function is used as the wavelet basis function. During the operation, the optimal decomposition layer number is determined to be 3, and the coefficients obtained after decomposition are subjected to soft threshold processing. The wdencmp function provided by the MATLAB software is used to select the threshold to obtain the spectral data after noise reduction.

[0215] Step 3( Figure 2 and Figure 3 ):

[0216] (1) Convert the spectral data vector into a two-dimensional spectral information matrix.

[0217] (2) Apply the Inception-ResNet v2 algorithm to the two-dimensional spectral information matrix

[0218] ①Stem part: Use a series of convolutional layers and pooling layers as the entrance of the network to extract preliminary features.

[0219] ②Inception-A, B, C modules: According to the structure of Inception-ResNet v2 (such as Figure 2 As shown in the figure, Inception-A, B, and C modules are stacked, and each module consists of multiple convolution kernels of different scales to capture features of different scales.

[0220] ③Reduction-A, B modules: Reduction-A and Reduction-B modules are used where it is necessary to reduce the dimension or change the size of the feature map.

[0221] ④Inception-ResNet-C module: Apply SENet (such as Figure 3 As shown in the figure), channel attention weights are obtained through global average pooling and fully connected layers to enhance feature expression.

[0222] Branch design:

[0223] The first branch: directly output the input.

[0224] Second branch: 1x1 convolution, the number of output channels is 192.

[0225] The third branch: 1x1 convolution to reduce the dimension, 3x3 convolution, and then another 3x3 convolution, with the output channels of 192, 224, and 256 respectively.

[0226] To obtain attention in the channel dimension, the feature map X is taken as input and passed through F tr The output after the convolution structure is U, F sqIt represents the compression operation, which compresses the spatial feature dimension to 1×1×C and is represented by the symbol Z.

[0227] Next, proceed to F ex In the excitation operation, the fully connected layer and the nonlinear activation function are used to establish the connection between channels, and then the normalized weights are obtained through the Sigmoid activation function. Then, the channel weights are multiplied by the corresponding channels of the feature map one by one, thereby achieving the weighting of each channel of the original feature map and strengthening important features.

[0228] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the technology may modify or change the above embodiments without violating the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A method for detecting the iron content of iron ore, characterized in that: The specific steps include: Step 1: Use drones to plan inspection routes and communicate with IoT data in the area to reach the designated cargo location; Step 2: Acquire spectral data and process the data; Step 3: Use machine learning models to identify and classify spectral lines; Step 4: Improve the learning model.

2. The method for detecting the iron content of iron ore according to claim 1, characterized in that: The specific steps of step 1 are: the drone takes off from its predetermined starting position, defines the path coverage by drawing the path of the drone on the map, and communicates with the IoT devices along the way to avoid restricted areas or possible obstacles to reach the designated cargo location area, wherein the drone model is DJI Mavic 3Pro drone, and the path planning method used is: (1) Establish a model for drone takeoff and landing, no-fly zones, and obstacle-marking wireless networks: (2) Constructing the POMDP decision model: ① Establish arrays and general reward functions; ②Use dual deep Q networks DDQNs to solve the remaining problems of POMDP; ③Use the RNN model to realize information transmission between different time steps; ④Use attention LSTM to overcome the limitations of the original encoding and decoding model on fixed vectors; ⑤Construct loss function.

3. The method for detecting the iron content of iron ore according to claim 2, characterized in that: The specific steps of step 2 are: ① Using a DJI Mavic 3Pro drone equipped with an American SENTERA 6X multispectral imager, the spectrometer wavelength range is 20nm to 840-nm for spectral data collection; ② Construct sample set partitioning: randomly divide the spectral sample data set into training set and test set. The training set sample data is used for model training. The model parameters are optimized through iteration to obtain the model with the best prediction accuracy. The test set sample data is used to evaluate the model. The standard spectral data comes from the atomic spectral database provided by the National Institute of Standards and Technology of the United States. ③Preprocess the spectral data Ⅰ. Baseline correction: This step uses a correction baseline based on polynomial fitting. First, all the minimum points in the multispectral data are defined as the main minimum points, and the data is divided into multiple bands. The minimum point in each band is determined as the characteristic value point, and then the characteristic value site of each level is fitted by spline difference and connected. The specific implementation method is as follows: The first step is to divide the spectral data into N parts according to the wavelength range, where the wavelength range of the spectral data is recorded as wavelength, and number is the number of points after being divided into N parts; In each segmented part, the lowest point of the spectral intensity is identified and defined as the main minimum point as the characteristic value point of the band: min{A(a i )} j′ j=1,2,3……n The characteristic value points of each band are strung together, and then the curves generated by these nodes are smoothed by spline interpolation technology, thus the baseline function f can be obtained. baseline Finally, by converting the original spectral data spec org With baseline data f baseline Perform differential calculation to obtain the corrected spectral data spec correct , specifically expressed as: spec correct =spec org -f baseline II. Data smoothing and denoising: In this step, the db2 wavelet function is used to decompose the laser-induced breakdown spectroscopy data, and the soft threshold method is used to filter out high-frequency noise signals. The soft threshold function expression used is: Where: λ is the given threshold; ω i，j is the wavelet coefficient; is the new wavelet coefficient obtained after soft threshold processing; The two-dimensional wavelet is decomposed in the two-dimensional plane to obtain three detail components. For example, the original data S is decomposed into three detail components, including the horizontal component Vertical Component Diagonal component The process of two-dimensional wavelet denoising is as follows: a. Input 2D raw LIBS data; b. Confirm the wavelet basis function, decomposition level, and threshold; c. Decompose the original data into four components, namely, approximate component, horizontal component, vertical component, and diagonal component; d. Further decompose the approximate components obtained in the previous step into 4 components, and so on, until the predetermined number of decomposition levels is reached; e. Use the threshold to denoise the coefficients obtained from each layer of decomposition; f. Reconstruct the denoised wavelet components into new two-dimensional spectral data; In this step, the db2 wavelet function is used as the wavelet basis function. During the operation, the optimal decomposition layer number is determined to be 3, and the coefficients obtained after decomposition are subjected to soft threshold processing. The wdencmp function provided by the MATLAB software is used to select the threshold to obtain the spectral data after noise reduction.

4. The method for detecting the iron content of iron ore according to claim 3, characterized in that: The specific steps of step three are: (1) Construct a spectral data conversion model; (2) Build a network model; ①Construct the optimization function of convolutional neural network; ②Build a channel attention mechanism; ③Weight each channel to enhance important features.

5. The method for detecting the iron content of iron ore according to claim 4, characterized in that: The specific steps of step 4 are: Atrous spatial pyramid pooling and Ghost module are used to optimize the model. The Ghost module is used to reduce the amount of calculation and parameters while retaining important feature information. Then, the atrous spatial pyramid pooling operation is applied to the feature map output by the Ghost module to extract multi-scale contextual information. Finally, the feature vector obtained by atrous spatial pyramid pooling is used for classification tasks. Such a fusion can make full use of the contextual information extraction capability of atrous spatial pyramid pooling while keeping the model lightweight, thereby improving the performance of the model in image classification.

Images