Multi-construction-method collaborative dust fall intelligent control method based on deep learning

Abstract

The invention provides a multi-construction-method collaborative dust fall intelligent control method based on deep learning. The method comprises the following steps: collecting mine field data of a mine field in real time; wherein the mine field data comprises mine field environment data, equipment operation data and topographic data; constructing a digital mine field model based on the mine field data; wherein the digital mine field model comprises a terrain model, an equipment model and a dust generation and diffusion model; based on the mine field data and the digital mine field model, constructing a reinforcement learning agent; based on the reinforcement learning agent, performing dust fall control on the mine field; the dust fall control is monitored on the basis of the pre-created human-computer interaction interface, intelligent and efficient control over the mine field dust fall process is achieved, the dust fall effect and the resource utilization efficiency are improved, and meanwhile the flexibility and reliability of the system are improved through human-computer cooperative control.

Description

Technical Field

[0001] The present invention relates to the technical field of dust reduction in rock mining, and in particular to a multi-method collaborative dust reduction intelligent control method based on deep learning. Background Art

[0002] During the mining, crushing and transportation of minerals and rocks, a large amount of dust will be generated. This dust not only causes serious pollution to the environment, but also may cause health problems such as respiratory diseases in the human body, posing a threat to the health of mine workers and surrounding residents.

[0003] Traditional dust reduction methods usually use a single dust reduction method, such as dry dust removal or wet spray, and lack coordinated control of various links and multiple dust reduction methods. These traditional methods are difficult to optimize dust reduction strategies in real time according to complex and changeable working conditions, resulting in poor dust reduction effects, waste of resources, and potential safety and environmental problems. For example, when the dust concentration is high, a single dust reduction method may not be able to effectively reduce the dust concentration; and when the equipment is operating in an unstable state, traditional methods may not be able to adjust the dust reduction strategy in time, thus affecting the dust reduction effect and the operating efficiency of the equipment. Summary of the invention

[0004] The present invention provides a multi-method coordinated dust reduction intelligent control method based on deep learning, which is used to solve the technical problem that the prior art cannot effectively reduce dust during mineral mining.

[0005] On the one hand, the present invention provides a multi-method collaborative dust reduction intelligent control method based on deep learning, comprising: Collecting mine data of the mine in real time; wherein the mine data includes environmental data of the mine, operation data of equipment and terrain data; Based on the mine data, a digital mine model is constructed; wherein the digital mine model includes a terrain model, an equipment model, and a dust generation and diffusion model; Based on the mine data and the digital mine model, construct a reinforcement learning agent; Based on the reinforcement learning agent, dust reduction control is performed on the mine; Monitoring the dust reduction control based on a pre-created human-computer interaction interface; Wherein, based on the mine data and the digital mine model, constructing a reinforcement learning agent includes: Constructing a state space of a reinforcement learning agent; wherein the state space includes information of a digital mine model, and the information of the digital mine model includes dust concentration values ​​of different particle sizes in each area, wind speed vector, humidity, temperature, current operating parameters of dust reduction equipment, and operating status of the equipment; Constructing an action space of a reinforcement learning agent; wherein the action space includes a variety of dust reduction operations, including turning on or off a preset dry dust collector, adjusting the electric field strength or filtering wind speed of the dry dust collector, starting or stopping a wet spray system in a preset area, changing the angle and coverage of the spray, switching between dry and wet combined dust reduction modes and allocating the working ratio of the two; The reinforcement learning agent is trained using a proximal policy optimization algorithm.

[0006] The multi-process collaborative dust reduction intelligent management and control method based on deep learning provided by the present invention constructs a digital mine model based on mine data; constructs a reinforcement learning intelligent agent based on mine data and the digital mine model; performs dust reduction control on the mine based on the reinforcement learning intelligent agent; and monitors dust reduction control based on a pre-created human-computer interaction interface, thereby realizing intelligent and efficient control of the dust reduction process in the mine, improving the dust reduction effect and resource utilization efficiency, and at the same time improving the flexibility and reliability of the system through human-computer collaborative control. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0008] Figure 1 It is a flowchart of a multi-process collaborative dust reduction intelligent control method based on deep learning provided by an embodiment of the present invention; Figure 2 It is a schematic diagram of the architecture of using a proximal strategy optimization algorithm to train a reinforcement learning agent provided by an embodiment of the present invention; Figure 3 It is a structural schematic diagram of a multi-method collaborative dust reduction intelligent control device based on deep learning provided by an embodiment of the present invention; Figure 4 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0009] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0010] Figure 1The figure is a flow chart of a multi-process collaborative dust reduction intelligent control method based on deep learning provided by an embodiment of the present invention. The execution subject of the method can be a computer, a tablet computer, a smart wearable device, etc.

[0011] See also Figure 1 The multi-method collaborative dust reduction intelligent control method based on deep learning can include the following steps.

[0012] Step 101: Collect mine data of the mine in real time; wherein the mine data includes environmental data of the mine, operation data of equipment and terrain data.

[0013] In this step, real-time collection of mine data of the mine may include: Sensors pre-deployed at preset locations in the mine collect environmental data in real time; the environmental data may include dust concentration, wind speed, humidity, and temperature, etc. The preset locations can generally be selected from key locations in the mine (such as mining areas, crushing areas, transportation areas, work areas, residential areas, etc.). Sensors may include dust concentration sensors (laser dust sensors can be used to accurately measure dust concentrations in different particle size ranges), wind speed sensors (such as three-cup anemometers, which have a wide measurement range and high accuracy), humidity sensors (capacitive humidity sensors, which have fast response and good stability), temperature sensors (thermocouple temperature sensors, which are resistant to high temperatures and have accurate measurements), etc., to collect environmental data in real time, and to transmit mine data to the data center through wireless transmission technology (Wi-Fi or 5G networks).

[0014] Based on the pre-deployed sensors of the mining equipment (such as rock drills, excavators, etc.), crushing equipment (such as jaw crushers, cone crushers, etc.), and transportation equipment (such as mine cars, conveyor belts, etc.) in the mine, the operation data of the equipment is collected; among them, the operation data may include the power of the mining equipment, the crushing frequency of the crushing equipment, and the speed and load of the transportation equipment. The operation data can also include the material feeding situation of the crushing equipment and whether there is material spillage on the conveyor belt. In addition to using the corresponding sensors, machine vision technology can also be used to install cameras around the equipment to collect images and videos of the equipment's operating status, so as to analyze the equipment's operation data through image analysis and video analysis.

[0015] Based on satellite remote sensing technology, geographic information system software and drone aerial photography technology, the terrain data of the mine is collected; the terrain data may include the location and shape of mountains, rivers, and mine pits, as well as the coordinates, area, altitude, etc. of each area within the mine. In general, satellite remote sensing technology can be used to obtain the topographic data of the mine, including the location and shape information of mountains, rivers, mine pits, etc.; geographic information system software is used to conduct detailed mapping of the mine to determine the coordinates, area, altitude and other information of each area within the mine; combined with drone aerial photography technology, high-resolution images of the mine are obtained to further supplement and refine the geographic information of the mine.

[0016] Step 102: construct a digital mine model based on the mine data; wherein the digital mine model includes a terrain model, an equipment model, and a dust generation and diffusion model.

[0017] In this step, constructing a digital mine model based on mine data may include: Using the collected terrain data, build a terrain model of the mine in 3D modeling software (such as 3ds Max). For example, accurately model mountains, mine pits, roads, etc., and add realistic texture materials to make them close to the appearance of real scenes.

[0018] According to the actual size, structural drawings and image data of the equipment in the mine, the mining equipment, crushing equipment and transportation equipment in the mine are 3D modeled to obtain the equipment model. The equipment model can accurately reflect the operating status and appearance characteristics of the equipment.

[0019] Based on the principles of fluid mechanics and particle dynamics, combined with the physical mechanism of dust generation during mining, crushing and transportation, a dust generation and diffusion model is constructed. The dust generation and diffusion model can simulate the amount of dust generated, initial velocity, diffusion path and concentration distribution in the environment under different working conditions (using computational fluid dynamics software Fluent to assist in modeling and solve related equations).

[0020] The terrain model, equipment model and dust generation and diffusion model can be integrated to obtain a digital mine model, ensuring that the coordinate systems of the terrain model, equipment model and dust generation and diffusion model are consistent and can interact with each other. Specifically, for example, first create a new project in Unity, import the terrain model and equipment model into the project resource folder in a suitable format (such as FBX, etc.), and use Unity's terrain editing tools to further optimize and adjust the imported terrain model in detail, such as adding textures, setting terrain height, etc. Then, add collision bodies, rigid bodies and other components to the equipment model for interactive simulation, and write scripts to control the animation and operation logic of the equipment model. Furthermore, the dust generation and diffusion model built based on the principles of fluid mechanics and particle dynamics is integrated into the Unity project in the form of code. By creating a particle system and combining it with script control to simulate the generation, movement and diffusion effects of dust, the variable associations related to environmental parameters (such as wind speed, humidity, etc.) and equipment operating status are set up so that the dust generation and diffusion model can be updated and displayed in real time according to various changes in the virtual mine, and finally the terrain model, equipment model and dust generation and diffusion model are seamlessly integrated in the Unity virtual reality development engine to build a unified virtual mine platform (i.e., digital mine model). The parameters in the virtual mine platform can also be continuously updated based on the real-time collected data to keep it highly consistent with the actual mine status, thus achieving the digital twin effect.

[0021] Step 103: Construct a reinforcement learning agent based on the mine data and the digital mine model.

[0022] In this step, building a reinforcement learning agent based on mine data and digital mine model includes: Construct the state space of the reinforcement learning agent; the state space includes the information of the digital mine model, which includes the dust concentration values ​​of different particle sizes in each area, wind speed vector, humidity, temperature, current operating parameters of dust reduction equipment (such as dry dust collector, wet spray system, dry-wet combined device, etc.) (such as the electric field strength of the dust collector, filtration wind speed, pressure, flow rate, spray angle of the spray system, etc.) and the operating status of the equipment (such as normal, faulty, maintained, etc.).

[0023] Construct an action space for a reinforcement learning agent; the action space includes a variety of dust reduction operations, including turning on or off a preset dry dust collector (such as a fan unit of a bag dust collector), adjusting the electric field strength or filtering wind speed of the dry dust collector, starting or stopping the wet spray system in a preset area, changing the angle and coverage of the spray, switching between dry and wet combined dust reduction modes and allocating the working ratio between the two.

[0024] Use the proximal policy optimization algorithm to train reinforcement learning agents.

[0025] Step 104: Based on the reinforcement learning agent, dust reduction control is performed on the mine.

[0026] Specifically, an example of dust reduction control in a mine based on reinforcement learning agents is as follows: Agent decision: The agent calculates the probability distribution of each dust reduction operation through the policy network based on the input status (such as the current dust concentration is 100μg / m³, the wind speed is 5m / s, the humidity is 60%, the temperature is 30°C, the equipment is operating normally, etc.). For example, the agent may calculate the probability distribution of the following actions: Open the dry dust collector in zone 1: 0.7; Adjust the spray angle of the wet spray system in zone 2: 0.2; Shut down the combined wet and dry dust suppression system in zone 3: 0.1.

[0027] Execute action: The agent selects an action based on the probability distribution, for example, "turn on the dry dust collector in area 1". Then, the agent sends a control instruction to the dry dust collector through the dust reduction equipment execution module to turn on the dust collector in this area. After receiving the instruction, the equipment starts to operate, thereby reducing the dust concentration in area 1.

[0028] Environmental feedback: After executing an action, the digital mine model can update the state according to the preset physical models (such as fluid dynamics model and particle diffusion model). For example, after turning on the dry dust collector, the dust concentration decreases from 100μg / m³ to 80μg / m³, while the energy consumption of the equipment increases from 100kW to 120kW. These updated state information will be fed back to the intelligent agent as the new state.

[0029] Reward calculation: The reward obtained by the agent is calculated according to the preset reward function. For example, the reward function may take into account factors such as the reduction of dust concentration, changes in energy consumption, and the operating status of the equipment. If the dust concentration is significantly reduced and the energy consumption is within an acceptable range, the agent will receive a higher reward value.

[0030] Step 105: Monitor dust reduction control based on a pre-created human-computer interaction interface.

[0031] Specifically, a specific example of the human-computer interaction interface monitoring dust reduction control is as follows: Real-time data display: The human-computer interaction interface displays information such as dust concentration distribution, equipment operation status, parameter settings of dust reduction equipment, and decision-making process of the intelligent agent in a real-time visual way in the virtual mine model (i.e., digital mine model). For example, the interface can display the heat map of dust concentration in different areas of the mine, as well as the current operating parameters of each dust reduction equipment (such as the electric field strength of the dry dust collector, the spray pressure of the wet spray system, etc.).

[0032] Early warning function: When key data (such as excessive dust concentration, equipment failure, etc.) are abnormal, the human-computer interaction interface will issue a real-time early warning. For example, if the dust concentration in a certain area exceeds the preset safety threshold (such as 150μg / m³), the interface will pop up a warning message and mark the area in red to remind the operator to pay attention in time.

[0033] Manual intervention: Operators can manually intervene in the operation of the dust suppression system through the human-computer interaction interface. For example, under special working conditions (such as sudden equipment failure, bad weather, etc.), operators can directly adjust the parameters of the dust suppression equipment (such as increasing the amount of water in the spray system) or switch the dust suppression method (such as switching from dry dust removal to wet spray). In addition, operators can also modify the training parameters of the reinforcement learning agent (such as the reward function weight coefficient, exploration rate, etc.) to adapt to different actual needs on site.

[0034] In this embodiment, a digital mine model is constructed based on mine data; a reinforcement learning agent is constructed based on the mine data and the digital mine model; dust reduction control is performed on the mine based on the reinforcement learning agent; and dust reduction control is monitored based on a pre-created human-computer interaction interface. This embodiment uses a multi-method to collaboratively control dust reduction, thereby achieving intelligent and efficient control of the dust reduction process in the mine, improving the dust reduction effect and resource utilization efficiency, and at the same time improving the flexibility and reliability of the system through human-computer collaborative control.

[0035] In one embodiment of this specification, Figure 2 Schematic diagram of the architecture of using the proximal strategy optimization algorithm to train a reinforcement learning agent provided by an embodiment of the present invention. Figure 2 , the training of reinforcement learning agents using proximal policy optimization algorithms can include initialization phase, interaction phase, network update phase and performance evaluation phase. Among them, the initialization phase can include building a policy network and a value network, creating a storage space as an experience buffer, and setting hyperparameters. The interaction phase can include action selection, executing actions and obtaining feedback, and calculating the estimated value of the advantage function. The network update phase can include sampling data, calculating losses, and updating parameters. If the performance evaluation is qualified, it ends. If it is not qualified, it can return to the initialization phase. The following is a detailed introduction to the training of reinforcement learning agents using proximal policy optimization algorithms.

[0036] Specifically, using the proximal policy optimization algorithm to train a reinforcement learning agent may include: Initialization phase: Use deep learning frameworks (such as TensorFlow or PyTorch) to build policy networks and value networks; the policy network is used to generate the probability distribution of each action based on the input mine data. The policy network includes an input layer, a hidden layer, and an output layer. The amount of neuron data in the input layer is determined by the dimension of the input data. The hidden layer includes two layers, one with 64 neurons and the other with 32 neurons. The output layer converts the output into a probability distribution vector with the same dimension as the action space through a softmax function; the value network is used to estimate the value of given mine data. The value network includes an input layer, a hidden layer, and an output layer. The amount of neuron data in the input layer is determined by the dimension of the input data. The hidden layer includes 1 layer with 64 neurons and the output layer has 1 neuron, which is used to output the value of given mine data.

[0037] Specifically, the policy network can be expressed as , s represents the given state (the given state is the input data), represents the action chosen by the agent in a given state s, Indicates in status s Next select action The number of neurons in the input layer of the policy network is determined according to the dimension of the state space. For example, if the state consists of the dust concentration in 5 areas of the mine, 3 environmental parameters (wind speed, humidity, temperature), 4 equipment operation status parameters, and 3 dust reduction equipment working status parameters, then the number of neurons in the input layer is 5+3+4+3=15. The hidden layer is set to 2 layers, with 64 neurons in the first layer and 32 neurons in the second layer. The activation function uses ReLU to enhance the nonlinear expression ability of the network. The output layer converts the output into a probability distribution vector with the same dimension as the action space through the softmax function, and each element represents the selection probability of the corresponding action.

[0038] For the policy network, assume that the input state s is linearly transformed by the first hidden layer weight matrix W1 (dimension is 64×15) and the bias vector b1 (dimension is 64×1), and the output of the first hidden layer is obtained by the ReLU activation function. Then h1 is transformed by the second hidden layer weight matrix W2 (dimension is 32×64) and the bias vector b2 (dimension is 32×1) to obtain the output of the second hidden layer. Finally, h2 undergoes a linear transformation through the output layer weight matrix W3 (dimension is n×32, n is the size of the action space) and the bias vector b3 (dimension is n×1), and the action probability distribution is obtained through the softmax function. The action probability distribution is shown in the following formula (1): (1); here represents the set of all learnable parameters of the policy network, e represents the base of the natural logarithm, Indicates all possible actions.

[0039] The value network can be expressed as , s represents a given state, Represents the value of a given state s. The input layer of the value network is the same as the policy network and is processed based on the same state information. The hidden layer structure can be set to one layer, which includes 64 neurons and also uses the ReLU activation function. The output layer is one neuron, which directly outputs the estimated value of state s. .

[0040] For the value network, the input state s is linearly transformed by the hidden layer weight matrix W4 (dimension is 64×15) and the bias vector b4 (dimension is 64×1), and the output of the hidden layer is obtained by the ReLU activation function. Then h3 is transformed linearly through the output layer weight matrix W5 (dimension is 1×64) and the bias vector b5 (dimension is 1×1) to obtain the state value, which is shown in the following formula (2): (2); among which, Represents the set of parameters of the value network.

[0041] Create a storage space as an experience buffer; the experience buffer is used to store the experience data during the interaction between the reinforcement learning agent and the digital mine model. For example, the structure of the experience data may include the current state s, the action taken, , the reward r obtained, the next state s', the probability of choosing this action under the current strategy And the advantage function estimate A. The size of the buffer is set according to the computing resources and training requirements. For example, it is initialized to a space that can store 5,000 pieces of experience data, and the initial state of the experience buffer is empty.

[0042] Set hyperparameters; hyperparameters include learning rate (which determines the step size of network weight adjustment each time the parameters are updated), discount factor (used to measure the importance of future rewards to current decisions), clipping parameter (this parameter is used to limit the difference between the new policy and the old policy when the policy is updated), mini-batch size (in each training iteration, N experience data are randomly drawn from the experience buffer to form a mini-batch) and number of training times (the number of times the mini-batch data is repeatedly trained in each training cycle).

[0043] Specifically, each hyperparameter can be explained as follows, and the learning rate is expressed as : can be set to 0.0003, which determines the step size of the network weight adjustment each time the parameters are updated. A smaller learning rate helps the network converge more stably during training, but may require more training time and iterations; a larger learning rate may cause the network to skip the optimal solution during training and make it difficult to converge to better performance.

[0044] The discount factor is expressed as : The value can be 0.99, which is used to measure the importance of future rewards to the current decision. When calculating the cumulative reward, the reward after t steps in the future will be discounted to This allows the agent to not only focus on the immediate rewards it receives, but also consider the long-term impact of current behavior on future states and rewards, thereby enabling it to learn better strategies to maximize long-term cumulative rewards. r Usually represents a reward, t represents the current time step, k is the step offset.

[0045] The cropping parameters are expressed as : can be set to 0.2. This parameter is used to limit the difference between the new strategy and the old strategy when the strategy is updated. By trimming the strategy ratio, the trimming formula is shown in the following formula (3): (3), Represent the ratio so that it is limited to arrive Within the range of , ensure that the strategy update will not be too radical, prevent instability during training, and ensure that the strategy can be gradually improved in a better direction. Represents the old parameter, which can also be understood as the previous parameter.

[0046] The mini-batch size is expressed as N: it can be 32. In each training iteration, N pieces of experience data are randomly extracted from the experience buffer to form a mini-batch. This can break the correlation between the experience data, make the samples of each training more diverse, improve the efficiency and stability of training, and also reduce the resource consumption of a single calculation, making the training process more stable and efficient.

[0047] The number of training times is represented by T: it can be set to 50. In each training cycle, the small batch data is repeatedly trained T times. By using small batch data for training multiple times, the network can better fit the rules in the data, further optimize the network parameters, improve the performance and generalization ability of the model, and enable it to more accurately cope with different mine working scenarios.

[0048] Interaction phase: The reinforcement learning agent interacts with the digital mine model to select actions, execute actions and obtain feedback, calculate rewards based on the reward function and store the experience data in the experience buffer; Calculate the advantage function estimate for each piece of experience data in the experience buffer.

[0049] For example, the action selection is as follows: At each training time step t, the agent first obtains the state s of the current environment t , and input it into the policy network The policy network outputs the probability distribution of each possible action based on the input state information. Then, the agent selects an action based on this probability distribution. When implementing it, you can use the random number generation function in the programming language combined with the probability distribution to implement the sampling operation. For example, in Python, you can use the numpy.random.choice function, use the action space as the selection range, and the action probability distribution output by the policy network as the selection probability, so as to randomly select an action. At the same time, record the probability of choosing this action , so that it can be used in the subsequent training process.

[0050] Execute actions and get feedback as follows: The agent performs the selected action in the virtual digital mine environment The virtual digital mine environment updates the current state according to its internal preset physical model and rules to generate the next state. For example, if the agent chooses to start the wet spray dust suppression equipment in a certain area, the virtual digital mine environment will calculate the change of dust concentration based on the parameters of the spray equipment (such as spray pressure, water volume, spray angle, etc.) and the current environmental conditions (such as wind speed, humidity, dust concentration distribution, etc.) through the corresponding physical models (such as fluid mechanics model, particle diffusion model, etc.), and update the operating status of the equipment (such as the operating status of the water pump, energy consumption, etc.). These comprehensive information constitute the next state . Calculate the reward obtained by the agent according to the pre-designed reward function . Reward function , where r is the reward value, C0 is the initial dust concentration, C t is the dust concentration after dust suppression operation, E0 is the estimated value of initial energy consumption, and E tis the actual energy consumption value, S is the equipment operation status score (S=1 when the equipment is operating normally, S<1 when a fault occurs), a1, a2, and a3 are weight coefficients corresponding to dust concentration, energy consumption, and equipment status, respectively, which are used to adjust the importance of different factors in the reward. They can be adjusted according to the key goals and needs of actual mine dust reduction. For example, if you are more concerned about the dust reduction effect, you can increase the a1 value appropriately; if you are more sensitive to energy consumption, increase the a2 value. This set of experience data is stored in the experience buffer to provide data support for subsequent training calculations.

[0051] The advantage function estimate is calculated as follows: For each experience in the experience buffer , i represents the i-th item, calculate the estimated value of the advantage function First, the state and Enter the value network separately , and obtain the corresponding value estimate and Then, according to the following formula (4) (4), calculate the estimated value of the advantage function. Indicates in status Take Action After that, you will get the reward immediately , plus the future state Value after discount , minus the current state Estimated value .if , indicating that in the state Take Action is better than the average behavior because it brings better returns than expected; conversely, if , it means that the action may not be the best choice and may result in a reward lower than the average level. Store the corresponding experience data position back in the experience buffer for use in subsequent policy network updates.

[0052] Network update phase: Sample mini-batches of data from the experience buffer, calculate policy loss and value loss, and update the parameters of the policy network and value network; Perform multiple training cycles until the policy network and value network converge or reach a predetermined training result.

[0053] Specifically, a small batch of data is sampled from the experience buffer. For example, 32 pieces of experience data are randomly sampled from the experience buffer using a random sampling function in a programming language to form a small batch of samples. For example, the random.sample function can be used to randomly select 32 elements (i.e., 32 pieces of experience data) from the experience buffer data set to ensure that the data obtained each time is random and independent, so that it can represent the distribution of the entire experience data and improve the generalization ability and stability of the training.

[0054] Calculate the policy loss and value loss. For example, the policy loss is: for each experience in the sampled mini-batch data , j represents the jth empirical data, and the calculation ratio , is the old policy network (with parameters ) in the state Next select action The probability of is the probability that the current policy network selects this action in the same state. Then, the policy loss is calculated by the following formula (5): (5).

[0055] Among them, the clip function is used to clip the ratio , limiting it to arrive Specifically, if ,but ;if ,but ;otherwise .

[0056] The purpose of designing the policy loss function is to optimize the parameters of the policy network so that it can select better actions while ensuring the stability of the policy update. Crop to prevent the strategy from updating too much and avoid instability during training. At the same time, use the advantage function to estimate the value To measure the pros and cons of each action, the policy network is more inclined to select actions with higher advantage values, thereby gradually improving the performance of the strategy.

[0057] Value loss: The value loss is calculated using the mean square error loss function, as shown in the following formula (6): (6).

[0058] in is the target value. For each sampled experience, the state value predicted by the value network is With target value For comparison, the parameters of the value network are adjusted by minimizing the mean squared error.

[0059] The goal of the value loss function is to enable the value network to estimate the value of the state more accurately. By comparing the predicted value With target value , and use the mean square error to measure this difference. The value network will gradually adjust its parameters during the training process to make the predicted value closer to the actual state value. The calculation method is based on the current reward and the discounted value of the next state This reflects the consideration of long-term value in reinforcement learning, that is, the value of the current state depends not only on the current reward, but also on the value of the future state.

[0060] Update the parameters of the policy network and the value network, for example, add the policy loss and the value loss to get the total loss , and then adjust the policy network parameters and value network parameters Find the gradient. In deep learning frameworks, you can use automatic differentiation functions (such as TensorFlow's GradientTape or PyTorch's autograd) to calculate the gradient of the loss function to the parameters. After obtaining the gradient, use the Adam optimizer to update the parameters based on the calculated gradient. For example, at each update, the policy network parameters The update formula is , value network parameters The update formula is ,in and They are the total loss versus policy network parameters and value network parameters The gradient vector of the network is . This process is repeated for 50 trainings to make full use of small batches of data for learning, so that the network parameters can better fit the data and improve the decision-making ability and value estimation accuracy of the intelligent agent.

[0061] Perform multiple training cycles, for example, repeat the steps of interacting with the agent and the environment, calculating the estimated value of the advantage function, and updating the policy network and the value network, and perform multiple training cycles (1000 to 10,000 cycles) until the policy network and the value network converge or achieve the predetermined training effect. There are many ways to determine whether the training end condition has been met, such as monitoring the change of network parameters. If the change of network parameters is less than a set threshold in multiple consecutive training cycles, it means that the network has converged; or evaluate the performance indicators of the agent on the validation set (such as the average dust concentration reduction rate, total energy consumption, equipment failure rate, etc.). When these indicators no longer have a significant improvement in multiple consecutive cycles, it can also be considered that the training has achieved a good effect; a maximum number of training steps can also be set. When this number of steps is reached, training is stopped regardless of whether the network is fully converged. After each training cycle, the policy network and the value network are evaluated once, and their performance on the validation set is recorded. The validation set is a part of the data separated from the pre-collected mine data. This data does not participate in the training process and is used to simulate different mine working conditions to test the generalization ability and performance of the agent on unseen data. By comparing the evaluation results of different training cycles, we can understand the performance improvement trend of the agent and determine whether it is necessary to further adjust the training parameters or improve the model structure.

[0062] Evaluation Phase: Evaluate the performance of reinforcement learning agents.

[0063] In this embodiment, by using the proximal policy optimization algorithm (PPO) to train the reinforcement learning agent, intelligent control of the dust reduction process in the mine is achieved. The dust reduction strategy can be dynamically adjusted to adapt to complex and changeable working conditions, optimize the dust concentration reduction effect, and at the same time consider energy consumption and equipment operation stability to reduce equipment failure rate.

[0064] In one embodiment of the present specification, evaluating the performance of a reinforcement learning agent includes: Evaluate the reinforcement learning agent according to a preset time period and record the evaluation results; the evaluation results include the average dust concentration reduction rate, total energy consumption and equipment failure rate; Adjust the reinforcement learning agent based on the evaluation results to optimize the reinforcement learning agent.

[0065] Specifically, the performance evaluation process, for example, during the training process, the performance of the agent is evaluated regularly (every 100 training cycles) using the validation set. The agent is deployed in the virtual mine environment corresponding to the validation set and runs for a certain period of time to simulate the actual mining-crushing-transportation process of ore and rock, observe whether the dust reduction method and parameter settings it selects are reasonable, and record relevant evaluation indicators.

[0066] Average dust concentration reduction rate: In each simulation scene of the validation set, record the initial dust concentration and the average dust concentration after the agent takes dust reduction measures, and calculate the dust concentration reduction rate. The average dust concentration reduction rate is obtained by averaging the dust concentration reduction rates of multiple scenes. This indicator directly reflects the dust reduction effect of the agent. If the average dust concentration reduction rate is high, it means that the agent can effectively select the appropriate dust reduction strategy to reduce the dust concentration in the mine and achieve a better dust reduction goal.

[0067] Total energy consumption: Monitor the energy consumption of various dust reduction equipment (such as power consumption of dry dust collectors, energy consumption of water pumps in wet spray systems, etc.) during the operation of the intelligent agent in the verification set, and calculate the total energy consumption value. A lower total energy consumption means that the intelligent agent can reasonably control energy use while achieving the dust reduction goal, improve the economy of the dust reduction process, and avoid excessive consumption of resources, which is of great significance for the actual operation cost control of the mine.

[0068] Equipment failure rate: Record the number and type of failures of various types of dust reduction equipment during the validation set operation. Combine the number of failures with the total equipment operation time or total number of times to calculate the equipment failure rate. A low equipment failure rate indicates that the agent's decision can not only effectively reduce dust, but also ensure the stable operation of dust reduction equipment, reduce equipment maintenance and downtime, and improve the reliability and stability of the entire dust reduction system, thereby ensuring the normal production order of the mine.

[0069] Hyperparameter adjustment and model improvement. For example, based on the evaluation results, if the performance of the agent is found to be unsatisfactory, such as poor dust reduction effect, excessive energy consumption, or high equipment failure rate, it is necessary to adjust the hyperparameters in the training process. For example, if the average dust concentration reduction rate is lower than the expected target, it may be necessary to appropriately increase the learning rate of the policy network to speed up the convergence of the network so that it can learn more effective dust reduction strategies more quickly; or adjust the weight coefficient in the reward function, such as increasing the coefficient a1 related to the dust reduction effect, to guide the agent to pay more attention to the improvement of the dust reduction effect. If the total energy consumption exceeds the acceptable range, you can try to reduce the feature weights related to energy consumption in the value network, or adjust the clipping parameters. , making the strategy update more conservative and avoiding unreasonable consumption of energy due to excessive pursuit of dust reduction effects.

[0070] If the equipment failure rate is high, it may be necessary to check whether the policy network has unreasonable situations when selecting dust reduction equipment operations, such as frequently and quickly switching equipment states. At this time, you can consider adding a reward and punishment mechanism for the stability of equipment operation, or adjusting the structure of the policy network, such as adding a memory unit for the equipment state, so that it can better consider the continuous operation of the equipment and avoid making decisions that lead to equipment failure. If the evaluation results show that the current structure of the policy network and the value network cannot effectively capture the complex relationship of the environment, resulting in poor decision-making results of the agent, you can consider improving the model structure. For example, increase the number of hidden layers or neurons in the policy network to improve its ability to fit complex state spaces; or introduce an attention mechanism in the value network so that the network can focus more on key state information that is closely related to dust reduction effects and equipment operation, so as to more accurately estimate the state value. After improving the model structure, retraining is required to optimize the performance of the agent so that it can better adapt to the complex and changeable working environment of the mine and achieve the goal of efficient dust reduction.

[0071] In this embodiment, the performance and adaptability of the system are further improved by regularly evaluating the reinforcement learning agent and optimizing and adjusting it according to the evaluation results. This evaluation mechanism can monitor the performance of the agent in dust reduction control in real time, such as key indicators such as dust concentration reduction rate, energy consumption and equipment failure rate, and dynamically adjust the training parameters or optimize the model structure accordingly. This not only ensures that the agent can maintain efficient dust reduction under different working conditions, but also improves the economy and reliability of the system, while reducing the maintenance cost and operation risk of the equipment.

[0072] In one embodiment of the present specification, adjusting the reinforcement learning agent according to the evaluation result to optimize the reinforcement learning agent includes at least one of the following: Adjust the learning rate, including: if the average dust concentration reduction rate is lower than the minimum value of the preset target interval, increase the learning rate of the strategy network; if the average dust concentration reduction rate is greater than the maximum value of the preset target interval, reduce the learning rate; Adjust the mini-batch size according to the resource consumption and training efficiency during the training process, including: if the computing resources are greater than the preset standard, increase the mini-batch size; if the computing resources are less than the preset standard, reduce the mini-batch size; Adjust the weight coefficients in the reward function, including: adjusting the weight coefficients related to dust concentration, energy consumption and equipment status in the reward function according to the actual dust reduction goals of the mine; Adjusting the clipping parameters includes: if the update amplitude during the policy update process is greater than a preset amplitude, reducing the clipping parameters to limit the amplitude of the policy update; Add a reward and penalty mechanism for equipment operation stability, including: if the equipment failure rate is greater than the preset failure threshold, add a reward and penalty mechanism for equipment operation stability in the reward function; According to the convergence situation during the training process, the number of training times in each training cycle is adjusted, including: if the network convergence speed is lower than the preset speed threshold, the number of training times is increased; If the evaluation result is lower than the preset evaluation target, improve the model structure, including: increasing the number of hidden layers or neurons in the policy network; or introducing an attention mechanism in the value network to enable the network to focus on status information related to dust reduction effects and equipment operation; Adjust the structure of the strategy network, including: if the intelligent agent controls the switching frequency of the device state to exceed the preset frequency when selecting the dust reduction device operation, add a memory unit for the device state to count the switching frequency of the device state so that the switching frequency is lower than the preset frequency.

[0073] In this embodiment, a flexible optimization mechanism is provided to respond to different evaluation results and working conditions through a variety of adjustment methods (such as learning rate adjustment, reward function weight modification, clipping parameter optimization, etc.). These adjustment methods can be optimized for specific problems, such as increasing the learning rate to speed up strategy learning when the dust concentration reduction effect is not ideal, or introducing a stability reward mechanism when the equipment failure rate is high. This flexible optimization method enables the system to better adapt to the complex and changing mine environment, and further improve the intelligence level and overall performance of dust reduction control.

[0074] In one embodiment of the present specification, monitoring dust reduction control based on a pre-created human-computer interaction interface includes: The dust concentration distribution, equipment operation status, parameter settings of dust reduction equipment, and decision-making process of the reinforcement learning agent in the digital mine model are displayed in real time through the human-computer interaction interface; Provide real-time warning for preset data types, including dust concentration and equipment failure.

[0075] For example, dust reduction can be performed through the dust reduction equipment execution module. Specifically, it receives the control instructions output by the reinforcement learning intelligent agent module, and accurately controls various types of dust reduction equipment through corresponding actuators such as electrical control loops and hydraulic control systems (such as motor speed regulation and electric field strength adjustment of dry dust collectors, water pump start and stop of wet spray systems, valve opening control to change spray pressure and flow, coordinated startup and parameter configuration of different subsystems in dry-wet combined devices, etc.) to perform corresponding dust reduction operations according to the instructions, thereby achieving precise regulation of the actual dust reduction process in the mine.

[0076] Monitoring is performed through the human-computer interaction and monitoring module. For example, the human-computer interaction and monitoring module includes a monitoring submodule and an interaction submodule. Monitoring submodule: The dust concentration distribution, equipment operation status, parameter settings of the dust reduction equipment, and the decision-making process of the intelligent agent in the virtual mine model are displayed in real time through a visual interface, so that operators can intuitively understand the operation of the dust reduction system of the entire mine. At the same time, real-time warnings are issued for key data (such as excessive dust concentration, equipment failure, etc.), reminding operators to pay attention in time and take corresponding measures.

[0077] Interaction submodule: Operators can manually intervene in the operation of the dust reduction system through this module. For example, under special working conditions (such as sudden equipment failure, bad weather, etc.), they can directly adjust the parameters of the dust reduction equipment or switch the dust reduction method. They can also modify the training parameters of the reinforcement learning agent (such as the reward function weight coefficient, exploration rate, etc.) to adapt to different actual needs on site, realize human-machine collaborative control, and improve the flexibility and reliability of the system.

[0078] In this embodiment, the real-time monitoring and early warning functions of the dust control process are realized through the human-computer interaction interface, which greatly enhances the operability and transparency of the system. Operators can intuitively understand the dust concentration distribution, equipment operation status and decision-making process of the intelligent agent in the mine through the interface, and manually intervene when necessary. At the same time, the real-time early warning function can promptly remind operators to pay attention to abnormal situations (such as excessive dust concentration or equipment failure), so that they can take quick measures to avoid potential safety and environmental risks. This human-computer collaborative control method not only improves the flexibility and reliability of the system, but also provides strong technical support for on-site operators.

[0079] In some other embodiments of the present invention, the real-time collection of mine data of a mine further includes: Multimodal data fusion: Combining image recognition technology, audio analysis technology and sensor data, more accurate monitoring of the dust source and diffusion path in the mine is carried out. Image recognition technology is used to identify the visual characteristics of dust, audio analysis technology is used to capture the sound characteristics of dust, and sensor data provides physical parameters such as dust concentration, wind speed, and humidity. Multimodal data is fused and processed through deep learning algorithms to generate a more comprehensive dust distribution feature map.

[0080] Dynamic data calibration: During the acquisition process, the sensor data is calibrated in real time through machine learning algorithms to compensate for errors caused by sensor aging, environmental interference, etc. For example, the Kalman filter algorithm is used to dynamically correct the sensor data to ensure the accuracy and reliability of the data.

[0081] Data enhancement and simulation: When data collection is insufficient, virtual data can be generated through data enhancement technology to enrich training samples. For example, the Generative Adversarial Network (GAN) is used to generate simulated dust concentration distribution and equipment operation status data to enhance the generalization ability of the model.

[0082] In some other embodiments of the present invention, the dust reduction control based on the reinforcement learning agent further includes: Environmental perception and dynamic strategy adjustment: Through the environmental perception sensor network deployed in the mine, the dynamic environmental changes of the mine are monitored in real time, including sudden weather changes (such as gusts, rainfall), surrounding construction activities or temporary equipment operation. The reinforcement learning agent can adjust the dust reduction strategy in real time based on this dynamic environmental information. For example, when a gust of wind is detected, the agent automatically adjusts the spray direction and intensity of the spray system to cope with the sudden spread of dust; when it rains, the agent reduces the operating power of the dry dust collector to optimize energy utilization.

[0083] Environmental change prediction and forward-looking strategy optimization: Combined with meteorological data and historical environmental change patterns, time series analysis and deep learning models (such as long short-term memory networks (LSTM)) are used to predict environmental changes in the mine in the future. Reinforcement learning agents adjust dust reduction strategies in advance based on the prediction results, such as starting the wet spray system in advance to deal with predicted dusty weather, or adjusting equipment parameters in advance before predicted rainfall to reduce energy consumption.

[0084] Multi-agent collaborative control: Multiple reinforcement learning agents are deployed in the mine, and each agent is responsible for dust reduction control in a specific area or specific equipment. By establishing a communication mechanism between agents, collaborative decision-making of multiple agents is achieved. For example, when the dust concentration in an area suddenly increases, the agent in that area can request the agent in the adjacent area to assist in adjusting the dust reduction strategy, and achieve a more efficient global dust reduction effect through collaborative work.

[0085] In some other embodiments of the present invention, the adopting the proximal strategy optimization algorithm to train the reinforcement learning agent further comprises: Transfer learning application: When training reinforcement learning agents, transfer learning technology is introduced to transfer the existing dust reduction experience of similar mines to the dust reduction control of the current mine. For example, the pre-trained policy network and value network are used as the initial model to quickly adapt to the specific working conditions of the current mine through fine-tuning, reducing training time and resource consumption.

[0086] Cross-scenario knowledge transfer: supports the transfer of dust control experience from other similar industrial scenarios (such as construction sites, port loading and unloading, etc.). Through feature extraction and alignment technology, the dust generation and diffusion characteristics in different scenarios are mapped, enabling the intelligent agent to use cross-scenario knowledge for more efficient dust control.

[0087] Incremental learning and continuous optimization: When the working conditions of the mine change (such as equipment update, terrain change, etc.), the agent is continuously optimized by incremental learning methods. For example, through online learning algorithms, the agent can update the policy network and value network in real time to adapt to the new working conditions without retraining the entire model, improving the adaptability and flexibility of the system.

[0088] In some other embodiments of the present invention, the dust reduction control method further includes: Edge computing and cloud computing collaborative architecture: Edge computing devices are deployed in the mine to process and analyze the mine data collected by sensors in real time, and quickly generate preliminary dust control instructions. At the same time, the data processed by the edge computing devices is uploaded to the cloud computing platform, which uses more powerful computing resources to conduct in-depth analysis and model training on the data, and optimize the policy network and value network of the reinforcement learning agent. Through the collaboration of edge computing and cloud computing, rapid response and global optimization of dust control can be achieved.

[0089] Dynamic resource allocation and load balancing: Dynamically allocate computing resources of edge computing devices and cloud computing platforms according to the real-time complexity and processing requirements of mine data. For example, when the dust concentration changes sharply, more resources are allocated to edge computing devices for rapid response; during the model training phase, more resources are allocated to the cloud computing platform to accelerate model optimization. At the same time, load balancing technology is used to ensure efficient use of computing resources to avoid resource waste or overload.

[0090] Security and privacy protection mechanism: In the process of edge computing and cloud computing collaboration, data encryption and access control technologies are used to ensure the safe transmission and storage of mine data. For example, end-to-end encryption technology is used to protect the privacy of sensor data during transmission, and strict access rights are set in the cloud computing platform to prevent unauthorized access and data leakage. In addition, blockchain technology is used to record the processing and use of data to enhance the traceability and credibility of data.

[0091] In some other embodiments of the present invention, the method for real-time collection of mine data of a mine further includes: Distributed sensor network architecture: A distributed sensor network is used to deploy sensors at multiple key locations in the mine, including mining faces, crushing areas, transportation routes, loading and unloading areas, and surrounding environmentally sensitive areas. The sensor network uses wireless communication technologies (such as LoRa, NB-IoT) or wired network connections to achieve real-time data transmission and synchronous collection. Each sensor node has data preprocessing capabilities and can perform preliminary processing on the collected raw data, such as filtering, normalization, and feature extraction, to reduce data transmission volume and improve data quality.

[0092] Adaptive data sampling and dynamic adjustment: The sensor network can adaptively adjust the data sampling frequency according to the actual working conditions of the mine and the changes in dust concentration. For example, when the dust concentration changes rapidly or the equipment is unstable, the sampling frequency is automatically increased to obtain more detailed dynamic data; when the dust concentration is stable, the sampling frequency is reduced to save energy and communication resources. The historical change trend of sensor data is analyzed through machine learning algorithms, and the sampling strategy is dynamically adjusted to ensure that high-quality data can be obtained under different working conditions.

[0093] Based on the same general inventive concept, the present invention also protects a multi-method collaborative dust reduction intelligent control device based on deep learning, such as Figure 3 Show, Figure 3 It is a structural schematic diagram of a multi-method collaborative dust reduction intelligent control device based on deep learning provided by an embodiment of the present invention. The multi-method collaborative dust reduction intelligent control device based on deep learning provided by the present invention is described below. The multi-method collaborative dust reduction intelligent control device based on deep learning described below and the multi-method collaborative dust reduction intelligent control method based on deep learning described above can be referred to each other.

[0094] The multi-method collaborative dust reduction intelligent management and control device based on deep learning includes a collection module 201, a model creation module 202, a reinforcement learning agent module 203, a dust reduction control module 204 and a human-computer interaction monitoring module 205.

[0095] The acquisition module 201 is used to collect mine data of the mine in real time; wherein the mine data includes environmental data of the mine, operation data of equipment and terrain data; The model creation module 202 is used to construct a digital mine model based on the mine data; wherein the digital mine model includes a terrain model, an equipment model, and a dust generation and diffusion model; The reinforcement learning agent module 203 is used to construct a reinforcement learning agent based on the mine data and the digital mine model; The dust control module 204 is used to perform dust control on the mine based on the reinforcement learning agent; The human-computer interaction monitoring module 205 is used to monitor the dust reduction control based on a pre-created human-computer interaction interface.

[0096] Figure 4 Schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Figure 4As shown, the electronic device may include: a processor 310, a communication interface 320, a memory 330 and a communication bus 340, wherein the processor 310, the communication interface 320 and the memory 330 communicate with each other through the communication bus 340. The processor 310 may call the logic instructions in the memory 330 to execute the multi-tasking collaborative dust reduction intelligent control method based on deep learning.

[0097] In addition, the logic instructions in the above-mentioned memory 330 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0098] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0099] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-method collaborative dust reduction intelligent control method based on deep learning, characterized in that: include: Collecting mine data of the mine in real time; wherein the mine data includes environmental data of the mine, operation data of equipment and terrain data; Based on the mine data, a digital mine model is constructed; wherein the digital mine model includes a terrain model, an equipment model, and a dust generation and diffusion model; Based on the mine data and the digital mine model, construct a reinforcement learning agent; Based on the reinforcement learning agent, dust reduction control is performed on the mine; Monitoring the dust reduction control based on a pre-created human-computer interaction interface; Wherein, based on the mine data and the digital mine model, constructing a reinforcement learning agent includes: Constructing a state space of a reinforcement learning agent; wherein the state space includes information of a digital mine model, and the information of the digital mine model includes dust concentration values ​​of different particle sizes in each area, wind speed vector, humidity, temperature, current operating parameters of dust reduction equipment, and operating status of the equipment; Constructing an action space of a reinforcement learning agent; wherein the action space includes a variety of dust reduction operations, including turning on or off a preset dry dust collector, adjusting the electric field strength or filtering wind speed of the dry dust collector, starting or stopping a wet spray system in a preset area, changing the angle and coverage of the spray, switching between dry and wet combined dust reduction modes and allocating the working ratio of the two; The reinforcement learning agent is trained using a proximal policy optimization algorithm.

2. According to the deep learning-based multi-tasking collaborative dust reduction intelligent control method of claim 1, it is characterized in that: The real-time collection of mine data includes: Based on sensors pre-deployed at preset locations of the mine, environmental data is collected in real time; wherein the environmental data includes dust concentration, wind speed, humidity and temperature; Based on the operation data of the sensors pre-deployed on the mining equipment, crushing equipment and transportation equipment of the mine, the operation data includes the power of the mining equipment, the crushing frequency of the crushing equipment, the speed and load of the transportation equipment; The terrain data of the mine is collected based on satellite remote sensing technology, geographic information system software and drone aerial photography technology; wherein the terrain data includes the location and shape of mountains, rivers, and mine pits, as well as the coordinates, area, and altitude of each area within the mine.

3. According to the deep learning-based multi-tasking collaborative dust reduction intelligent control method of claim 1, it is characterized in that: The constructing of a digital mine model based on the mine data comprises: Using the collected terrain data, a terrain model of the mine is constructed in 3D modeling software; According to the actual size, structural drawings and image data of the equipment in the mine, the mining equipment, crushing equipment and transportation equipment in the mine are 3D modeled to obtain the equipment model; Based on the principles of fluid mechanics and particle dynamics, and combined with the physical mechanism of dust generation during rock mining, crushing and transportation, a dust generation and diffusion model is constructed.

4. According to the deep learning-based multi-tasking collaborative dust reduction intelligent control method of claim 1, it is characterized in that: The adopting of the proximal strategy optimization algorithm to train the reinforcement learning agent comprises: Initialization phase: A policy network and a value network are constructed using a deep learning framework; wherein the policy network is used to generate the probability distribution of each action according to the input mine data, and the policy network includes an input layer, a hidden layer, and an output layer. The amount of neuron data in the input layer is determined by the dimension of the input data, and the hidden layer includes two layers, one layer has 64 neurons, and the other layer has 32 neurons. The output layer converts the output into a probability distribution vector with the same dimension as the action space through a softmax function; the value network is used to estimate the value of given mine data, and the value network includes an input layer, a hidden layer, and an output layer. The amount of neuron data in the input layer is determined by the dimension of the input data, and the hidden layer includes 1 layer with 64 neurons, and the output layer has 1 neuron, which is used to output the value of given mine data; Creating a storage space as an experience buffer; wherein the experience buffer is used to store experience data during the interaction between the reinforcement learning agent and the digital mine model; Setting hyperparameters; wherein the hyperparameters include learning rate, discount factor, clipping parameter, mini-batch size and number of training times; Interaction phase: The reinforcement learning agent interacts with the digital mine model to select actions, execute actions and obtain feedback, calculate rewards based on the reward function and store the experience data in the experience buffer; Calculate the advantage function estimate of each piece of experience data in the experience buffer; Network update phase: Sample mini-batches of data from the experience buffer, calculate policy loss and value loss, and update the parameters of the policy network and value network; Perform multiple training cycles until the policy network and value network converge or reach the predetermined training results; Evaluation Phase: Evaluate the performance of reinforcement learning agents.

5. According to the deep learning-based multi-tasking collaborative dust reduction intelligent control method of claim 4, it is characterized in that: The evaluation of the performance of the reinforcement learning agent includes: Evaluate the reinforcement learning agent according to a preset time period and record the evaluation results; wherein the evaluation results include the average dust concentration reduction rate, the total energy consumption and the equipment failure rate; The reinforcement learning agent is adjusted according to the evaluation results to optimize the reinforcement learning agent.

6. The multi-tasking collaborative dust reduction intelligent control method based on deep learning according to claim 5 is characterized in that: The step of adjusting the reinforcement learning agent according to the evaluation result to optimize the reinforcement learning agent includes at least one of the following: Adjusting the learning rate includes: if the average dust concentration reduction rate is lower than the minimum value of the preset target interval, increasing the learning rate of the strategy network; if the average dust concentration reduction rate is greater than the maximum value of the preset target interval, reducing the learning rate; Adjusting the mini-batch size according to resource consumption and training efficiency during training, including: if the computing resources are greater than a preset standard, increasing the mini-batch size; if the computing resources are less than the preset standard, reducing the mini-batch size; Adjust the weight coefficients in the reward function, including: adjusting the weight coefficients related to dust concentration, energy consumption and equipment status in the reward function according to the actual dust reduction goals of the mine; Adjusting the clipping parameters includes: if the update amplitude during the policy update process is greater than a preset amplitude, reducing the clipping parameters to limit the amplitude of the policy update; Add a reward and penalty mechanism for equipment operation stability, including: if the equipment failure rate is greater than the preset failure threshold, add a reward and penalty mechanism for equipment operation stability in the reward function; According to the convergence situation during the training process, the number of training times in each training cycle is adjusted, including: if the network convergence speed is lower than the preset speed threshold, the number of training times is increased; If the evaluation result is lower than the preset evaluation target, improve the model structure, including: increasing the number of hidden layers or neurons in the policy network; or introducing an attention mechanism in the value network to enable the network to focus on status information related to dust reduction effects and equipment operation; Adjusting the structure of the strategy network includes: if the intelligent agent controls the switching frequency of the device state to exceed the preset frequency when selecting the dust reduction device operation, adding a memory unit for the device state to count the switching frequency of the device state so that the switching frequency is lower than the preset frequency.

7. The multi-tasking collaborative dust reduction intelligent control method based on deep learning according to claim 1 is characterized in that: The dust reduction control based on the pre-created human-computer interaction interface includes: The dust concentration distribution, equipment operation status, parameter settings of dust reduction equipment, and decision-making process of the reinforcement learning agent in the digital mine model are displayed in real time through the human-computer interaction interface; Provide real-time warning for preset data types, wherein the preset data types include dust concentration and equipment failure.

Images