A deep learning-based motion control optimization method for large-scale bulk material equipment

Abstract

The present invention discloses a deep learning-based motion control optimization method for large-scale bulk material equipment, which relates to the fields of computer systems based on specific computational models, deep learning, intelligent control, and industrial automation. To address the problems of incomplete environmental perception and poor adaptability to dynamic environments in bulk material equipment control systems, the present method first collects and preprocesses data through multi-source sensors; secondly, a cross-modal attention fusion model is constructed to achieve feature alignment; and finally, an LSTM-PPO hybrid network architecture is designed, in which the LSTM layer processes temporal state features and the PPO algorithm optimizes the control strategy. Compared with the existing technology, the present invention can improve the accuracy and controllability of traditional large-scale bulk material equipment motion control, and more easily improve the system's operating efficiency and robustness. The present method can be widely used in logistics and bulk cargo handling, industrial intelligent manufacturing, and production.

Description

Technical Field

[0001] The present invention relates to the fields of machine learning, deep learning, intelligent manufacturing and industrial automation, and in particular to a motion control optimization method for large-scale bulk material equipment based on deep learning. Background Art

[0002] In modern industrial production, large-scale bulk material handling equipment is widely used in locations such as mines, ports, and power plants for loading, unloading, and stacking large quantities of materials. Due to the complex and ever-changing operating environment, traditional single-sensor sensing methods struggle to fully capture the material state and equipment operating conditions. This can lead to control deviations, impacting operational efficiency and even causing equipment failure. Therefore, intelligent perception and decision-making based on multi-source information fusion are crucial for improving the automation level of bulk material handling equipment.

[0003] In the field of large-scale bulk material handling equipment, precise control of the material conveying process is a critical component of industrial production. To ensure operational efficiency and safety standards, industry standards have established strict technical specifications for parameters such as equipment motion trajectory and operating speed. Due to the strong coupling between various control parameters in the system, the equipment's operating state is easily affected by material characteristics and environmental factors, leading to control imbalances and posing a severe challenge to the equipment's efficient and stable operation. Therefore, achieving intelligent control of large-scale bulk material handling equipment has become a key issue that needs to be urgently addressed to ensure production efficiency and operational safety. This type of motion control optimization is not only a complex nonlinear control problem, but also a systems engineering project involving multi-sensor fusion and multi-objective optimization. Research on precise control and stable operation has become a key research direction in the field of industrial automation. Traditional control methods generally rely on single-modality sensor data and empirical models. However, in practice, vision-based methods are susceptible to interference from lighting and weather, and lidar performance degrades in dusty environments. On the one hand, the traditional control methods based on physical mechanism modeling in the existing control technology system have poor adaptability to dynamically changing working conditions. On the other hand, advanced control algorithms based on policy gradients generally have technical bottlenecks such as insufficient stability and poor convergence during the training process.

[0004] To address the shortcomings of the aforementioned existing technologies, a cross-modal attention LSTM-PPO (Long Short-Term Memory-Proximal Policy Optimization) decision network that integrates multi-source data is proposed for motion control optimization. This approach enables precise perception and control of complex working conditions, leveraging the complementary strengths of multi-source features to build a more complete environmental perception system. Building on the Proximal Policy Optimization (PPO) algorithm, this approach introduces a cross-modal attention mechanism and a Long Short-Term Memory (LSTM) temporal modeling module. By extracting temporal features, this approach enhances the system's adaptability to dynamic working conditions, resulting in more stable training and a more optimized control strategy. Summary of the Invention

[0005] The present invention proposes a motion control optimization method for large-scale bulk material equipment based on deep learning. The motion control optimization of large-scale bulk material equipment is achieved through a cross-modal attention mechanism and an LSTM-PPO network model. The cross-modal attention mechanism can effectively fuse multi-source data such as lidar point clouds and visual images, and the PPO algorithm can automatically learn the weight distribution of different sensor data, significantly improving the feature fusion effect. In addition, the LSTM network can accurately capture the temporal dependency of the equipment's operating status and address the impact of historical operating status on current control decisions. This method can stabilize the equipment's operating parameters within the optimal range and improve its dynamic response characteristics, thereby enhancing the system's adaptability and ensuring long-term operational reliability.

[0006] To achieve the above objectives, the following technical solutions are adopted:

[0007] Step 1: Data is collected and preprocessed through multi-source sensors. Point cloud and image features are extracted based on curvature features and the improved ORB (Oriented FAST and Rotated BRIEF) algorithm, respectively, to achieve collaborative processing of multi-source data.

[0008] Step 1.1: LiDAR scans the material yard environment, acquiring real-time 3D point cloud data of the material pile, including its shape, height, and surface profile. An RGB industrial camera collects surface image data, extracting information about the material's color distribution, texture characteristics, and particle size, assisting in identifying material type and stacking status. Inertial sensors monitor the equipment's motion in real time, including spatial position, acceleration, and angular velocity, to compensate for measurement errors caused by equipment vibration. A laser flow meter monitors material flow data in real time, ensuring its real-time and accuracy.

[0009] Step 1.2: Use the curvature-based feature point detection method for the point cloud data obtained by the lidar. , construct its local neighborhood point set , by calculating the covariance matrix of the neighborhood point To characterize local geometric features:

[0010]

[0011]

[0012] in is a point in the neighborhood, is the centroid of the neighborhood point set, is the transpose. The eigenvalues are obtained according to the covariance matrix , then calculate the point Curvature .

[0013]

[0014]

[0015] Retention satisfaction point, is the threshold, The key features of the stockpile contour are screened for local point cloud density.

[0016] Step 1.3: Use the improved ORB feature detector to extract visual features from the image captured by the RGB industrial camera. Use the FAST feature detection algorithm with adaptive threshold to calculate the feature point response value:

[0017]

[0018] in Indicates The detection window is centered. is the pixel intensity value, is the center pixel intensity value. At the same time, the feature points are dynamically controlled according to the overall contrast of the image. quantity:

[0019] N f = N max [ 1 − exp ( − k f ) ]

[0020] in is the maximum feature point, represents the image contrast measure, is the adjustment coefficient. This ensures that stable image feature points can be extracted under different lighting conditions.

[0021] Step 2: Construct a cross-modal attention feature fusion model to enhance the correlation between 3D point cloud and 2D visual features, and realize the complementary advantages and collaborative perception of multimodal data.

[0022] Step 2.1: The cross-modal attention fusion model uses a multi-head scaled dot product attention mechanism to establish geometric consistency constraints on the spatial alignment module, stably extracting common features and complementary information across modalities in complex environments. A learnable linear transformation is used to map them into a unified feature space:

[0023]

[0024] in is the characteristic of active query, is the matched feature, For the actual information delivered, is the point cloud feature matrix, is the image feature matrix, 、 、 is the learnable projection matrix.

[0025] Using Hamming distance Measuring feature similarity:

[0026]

[0027]

[0028] in 、 are the descriptors of feature points A and B respectively, is the dynamic Hamming distance threshold, is the basic threshold, is the threshold adjustment amplitude, Normalize the contrast of the image and keep The matching pairs are regarded as candidate corresponding points. The coarse matching stage can quickly screen out potential matching relationships.

[0029] Step 2.2: Then perform feature matching and construct an optimization objective function that includes geometric error and motion prior :

[0030] E ( T ) = ∑ [ r ( ‖ π ( T · p i ) − q i ‖ 2 ) ] + l ‖ log ( T · T previous − 1 ) ‖ F 2

[0031]

[0032] in is the transformation matrix that describes the projection relationship from point cloud to image, is the projected 2D image point, is the Huber robust kernel function, is the camera projection model, is the regularization coefficient, Represents the transformation matrix of the previous frame, is the input error, is the threshold parameter. The objective function is solved iteratively using the Gauss-Newton algorithm until convergence.

[0033] Step 2.3: Perform nonlinear mapping on the optimized matching results obtained by optimizing the objective function, and use the multi-head attention mechanism to capture the dependencies between features from multiple dimensions. Attention weight The calculation is as follows:

[0034]

[0035] in is the association strength between each point in the point cloud and each area of the image, is the scaling factor, This is row-by-row normalization.

[0036] The multi-head scaling dot product attention mechanism is used to expand the projected features into Each attention head is calculated independently, and the multi-head attention expansion calculation is as follows:

[0037]

[0038]

[0039] in is the single-head attention output, is the output projection matrix. The cross-modal attention weight mechanism is designed to be calculated as follows:

[0040]

[0041] Finally, the output environment state representation :

[0042]

[0043] in is the fusion feature of the cross-modal attention output, is a feedforward neural network, It is layer normalization processing.

[0044] Step 3: Construct an LSTM-PPO hybrid decision network and train it to obtain the optimization results.

[0045] Step 3.1: The LSTM-PPO decision network architecture, based on the LSTM network and the PPO algorithm, enables precise control of the movement of large bulk material handling equipment. The network model consists of an input layer, an LSTM layer, a fully connected layer, an Actor (policy) network, and a Critic (value) network.

[0046] The input layer receives multi-source data after feature extraction and matching, including environmental state representation, equipment operating status, and material flow data; the LSTM layer uses a gating mechanism to capture the time series dependencies in the data and explore the long-term and short-term changes in the equipment's motion state; the fully connected layer integrates the features output by the LSTM layer; the Actor generates an equipment motion control strategy based on the integrated features, and the Critic evaluates the value of the strategy and provides feedback for optimization.

[0047] The LSTM layer has 128 hidden units and 2 layers. The bidirectional structure enhances the perception of historical and future information, ensuring that the network can fully learn the dynamic changes during the movement of equipment. Output short-term memory as follows:

[0048]

[0049] in is the unit state, For output gating, is the activation function.

[0050] Step 3.2: Fusion of LSTM-encoded temporal features , construct an Actor-Critic framework, in which the Actor network is responsible for strategy generation and the Critic network performs value evaluation. The PPO algorithm is integrated into the decision network, and the iterative update of the control strategy is achieved by alternately optimizing the policy network and the value network.

[0051] Actor output equipment continuous control action :

[0052] a t = tanh [ W a ⋅ Re LU ( W 1 ⋅ h t + b 1 ) + b a ]

[0053] in , is the fully connected layer parameter, is the activation function. , is the output layer parameter, Constrain the action to [ − 1 , 1 ] , and then linearly mapped to the actual control range.

[0054] Critic predicts the long-term operating benefits of the current state of the equipment, guides the direction of strategy optimization, and estimates Momentary status Starting discount accumulation reward The calculation is as follows:

[0055]

[0056] in , is the fully connected layer parameter, , are the output layer parameters.

[0057] Step 3.3: Based on the PPO algorithm, an end-to-end training method is used, with the training batch size set to 64 and the number of optimization iterations set to 10. The cumulative discounted return is used as the optimization target to limit the magnitude of the policy update and avoid policy crashes caused by excessive single updates. Advantage function The calculation is as follows:

[0058]

[0059]

[0060] The discount factor Set to 0.95, smoothing coefficient Set to 0.9, For instant rewards, for Time series difference error, set the objective function:

[0061] L CLIP ( i ) = E t { minutes [ r t ( i ) A t , clip ( r t ( i ) , 1 − e , 1 + e ) A t ] }

[0062]

[0063] in is the comparison between the new and old strategies, The clipping threshold is set to 0.2, is a random strategy. The training objective function and the total loss function are calculated as follows:

[0064]

[0065] The critic loss weight Set to 0.5, The policy entropy weight is set to 0.02, is the value loss function, is the policy entropy.

[0066] Real-time monitoring of the loss function and performance index changes during network training, timely adjustment of parameters, ensuring the generalization ability of the network, and obtaining the final optimized control strategy.

[0067] The beneficial effects of the present invention are as follows:

[0068] 1. This motion control optimization method for large-scale bulk material equipment first adopts a collaborative extraction method based on curvature features and an improved ORB algorithm for multi-source data to solve the problems of inconsistent feature scales and low matching accuracy between point cloud and image data. It can improve the pose estimation accuracy of large-scale bulk material equipment and reduce control deviations caused by sensor data mismatch.

[0069] 2. By constructing a cross-modal attention feature fusion model, the problem of weak correlation in cross-modal feature fusion is solved, and high-precision alignment and complementary enhancement of multimodal data are achieved, which can improve the target recognition and tracking capabilities of large bulk equipment in complex environments.

[0070] 3. Adopting the LSTM-PPO hybrid decision network, the long-term dependency modeling capability of LSTM is utilized to extract state features, and the policy gradient optimization of the PPO algorithm is combined to realize adaptive decision-making in dynamic environments, solve the gradient vanishing problem in long-term dependency modeling, and enable large-scale bulk material equipment to have better environmental adaptability and control accuracy.

[0071] The present invention will be described in further detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is a schematic diagram of the process of the present invention for optimizing motion control of large-scale bulk material equipment;

[0073] Figure 2 This is the structural diagram of the cross-modal attention fusion model. DETAILED DESCRIPTION

[0074] The detailed embodiments of the present invention are described below in conjunction with the accompanying drawings.

[0075] A deep learning-based motion control optimization method for large bulk material equipment, see the attached Figure 1 , including the following steps:

[0076] Step 1: Collect material field environmental data through multi-source sensors, extract robust features of point cloud and image respectively based on curvature features and improved ORB algorithm, and use spatiotemporal alignment and normalization processing to realize collaborative preprocessing of multi-source data, providing highly consistent input for cross-modal fusion.

[0077] Step 2: Construct a cross-modal attention fusion model to enhance the semantic relevance between point clouds and visual features through feature matching optimization and dynamic weight allocation under geometric constraints, and output a multimodal joint feature expression with complementary advantages.

[0078] Step 3: Construct an LSTM-PPO hybrid decision network, use the long-term dependency modeling capability of LSTM to extract state features, and combine the policy gradient optimization of the PPO algorithm to achieve adaptive decision-making in a dynamic environment to obtain the final optimization strategy.

[0079] Implementation step 1: Data acquisition and preprocessing are performed through multi-source sensors, and point cloud and image features are extracted based on curvature features and the improved ORB algorithm respectively.

[0080] Step 1.1: LiDAR scans the material yard environment, acquiring real-time 3D point cloud data of the material pile, including its shape, height, and surface profile. An RGB industrial camera collects surface image data, extracting information about the material's color distribution, texture characteristics, and particle size, assisting in identifying material type and stacking status. Inertial sensors monitor the equipment's motion in real time, including spatial position, acceleration, and angular velocity, to compensate for measurement errors caused by equipment vibration. A laser flow meter monitors material flow data in real time, ensuring its real-time and accuracy.

[0081] Step 1.2: Use the curvature-based feature point detection method for the point cloud data obtained by the lidar. , construct its local neighborhood point set , by calculating the covariance matrix of the neighborhood point To characterize local geometric features:

[0082]

[0083]

[0084] in is a point in the neighborhood, is the centroid of the neighborhood point set, is the transpose. The eigenvalues are obtained according to the covariance matrix , then calculate the point Curvature .

[0085]

[0086]

[0087] Retention satisfaction point, is the threshold, The key features of the stockpile contour are screened for local point cloud density.

[0088] Step 1.3: Use the improved ORB feature detector to extract visual features from the image captured by the RGB industrial camera. Use the FAST feature detection algorithm with adaptive threshold to calculate the feature point response value:

[0089]

[0090] in Indicates The detection window is centered. is the pixel intensity value, is the center pixel intensity value. At the same time, the feature points are dynamically controlled according to the overall contrast of the image. quantity:

[0091] N f = N max [ 1 − exp ( − k f ) ]

[0092] in is the maximum feature point, represents the image contrast measure, is the adjustment coefficient. This ensures that stable image feature points can be extracted under different lighting conditions.

[0093] Implementation step 2: Based on the features extracted in implementation step 1, see the attached Figure 2 , construct a cross-modal attention fusion model, enhance the semantic correlation between point cloud and visual features through feature matching optimization and dynamic weight allocation, and output a multimodal environment state representation with complementary advantages.

[0094] Step 2.1: The cross-modal attention fusion model uses a multi-head scaled dot product attention mechanism to establish geometric consistency constraints on the spatial alignment module, stably extracting common features and complementary information across modalities in complex environments. A learnable linear transformation is used to map them into a unified feature space:

[0095]

[0096] in is the characteristic of active query, is the matched feature, For the actual information delivered, is the point cloud feature matrix, is the image feature matrix, 、 、 is the learnable projection matrix.

[0097] Using Hamming distance Measuring feature similarity:

[0098]

[0099]

[0100] in 、 are the descriptors of feature points A and B respectively, is the dynamic Hamming distance threshold, is the basic threshold, is the threshold adjustment amplitude, Normalize the contrast of the image and keep The matching pairs are regarded as candidate corresponding points. The coarse matching stage can quickly screen out potential matching relationships.

[0101] Step 2.2: Then perform feature matching and construct an optimization objective function that includes geometric error and motion prior :

[0102] E ( T ) = ∑ [ r ( ‖ π ( T · p i ) − q i ‖ 2 ) ] + l ‖ log ( T · T previous − 1 ) ‖ F 2

[0103]

[0104] in is the transformation matrix that describes the projection relationship from point cloud to image, is the projected 2D image point, is the Huber robust kernel function, is the camera projection model, is the regularization coefficient, Represents the transformation matrix of the previous frame, is the input error, is the threshold parameter. The objective function is solved iteratively using the Gauss-Newton algorithm until convergence.

[0105] Step 2.3: Perform nonlinear mapping on the optimized matching results obtained by optimizing the objective function, and use the multi-head attention mechanism to capture the dependencies between features from multiple dimensions. Attention weight The calculation is as follows:

[0106]

[0107] in is the association strength between each point in the point cloud and each area of the image, is the scaling factor, This is row-by-row normalization.

[0108] The multi-head scaling dot product attention mechanism is used to expand the projected features into Each attention head is calculated independently, and the multi-head attention expansion calculation is as follows:

[0109]

[0110]

[0111] in is the single-head attention output, is the output projection matrix. The cross-modal attention weight mechanism is designed to be calculated as follows:

[0112]

[0113] Finally, the output environment state representation :

[0114]

[0115] in is the fusion feature of the cross-modal attention output, is a feedforward neural network, It is layer normalization processing.

[0116] Implementation step 3: Construct an LSTM-PPO hybrid decision network, using the fused features generated in implementation step 2 as input. Utilize the long-term dependency modeling capability of LSTM to extract state features, and combine the policy gradient optimization of the PPO algorithm to achieve adaptive decision-making in a dynamic environment and obtain the final optimization strategy.

[0117] Step 3.1: The LSTM-PPO decision network architecture, based on the LSTM network and the PPO algorithm, enables precise control of the movement of large bulk material equipment. The network model consists of an input layer, an LSTM layer, a fully connected layer, an actor network, and a critic network.

[0118] The input layer receives multi-source data after feature extraction and matching, including environmental state representation, equipment operating status, and material flow data; the LSTM layer uses a gating mechanism to capture the time series dependencies in the data and explore the long-term and short-term changes in the equipment's motion state; the fully connected layer integrates the features output by the LSTM layer; the Actor generates an equipment motion control strategy based on the integrated features, and the Critic evaluates the value of the strategy and provides feedback for optimization.

[0119] The LSTM layer has 128 hidden units and 2 layers. The bidirectional structure enhances the perception of historical and future information, ensuring that the network can fully learn the dynamic changes during the movement of equipment. Output short-term memory as follows:

[0120]

[0121] in is the unit state, For output gating, is the activation function.

[0122] Step 3.2: Fusion of LSTM-encoded temporal features , construct an Actor-Critic framework, in which the Actor network is responsible for strategy generation and the Critic network performs value evaluation. The PPO algorithm is integrated into the decision network, and the iterative update of the control strategy is achieved by alternately optimizing the policy network and the value network.

[0123] Actor output equipment continuous control action :

[0124] a t = tanh [ W a ⋅ Re LU ( W 1 ⋅ h t + b 1 ) + b a ]

[0125] in , is the fully connected layer parameter, is the activation function. , is the output layer parameter, Constrain the action to [ − 1 , 1 ] , and then linearly mapped to the actual control range.

[0126] Critic predicts the long-term operating benefits of the current state of the equipment, guides the direction of strategy optimization, and estimates Momentary status Starting discount accumulation reward The calculation is as follows:

[0127]

[0128] in , is the fully connected layer parameter, , are the output layer parameters.

[0129] Step 3.3: Based on the PPO algorithm, an end-to-end training method is used, with the training batch size set to 64 and the number of optimization iterations set to 10. The cumulative discounted return is used as the optimization target to limit the magnitude of the policy update and avoid policy crashes caused by excessive single updates. Advantage function The calculation is as follows:

[0130]

[0131]

[0132] r t = [ 0 . 6 α + 0 . 7 β ]

[0133] The discount factor Set to 0.95, smoothing coefficient Set to 0.9, For instant rewards, The time required to achieve the control target, is the spatial offset between the target position and the actual position, for Time series difference error, set the objective function:

[0134] L CLIP ( i ) = E t { minutes [ r t ( i ) A t , clip ( r t ( i ) , 1 − e , 1 + e ) A t ] }

[0135]

[0136] in is the comparison between the new and old strategies, The clipping threshold is set to 0.2, is a random strategy. The training objective function and the total loss function are calculated as follows:

[0137]

[0138] The critic loss weight Set to 0.5, The policy entropy weight is set to 0.02, is the value loss function, is the policy entropy.

[0139] By monitoring the changes in loss functions and performance indicators during network training, parameters are adjusted in a timely manner to ensure network training convergence and obtain the final optimized control strategy.

Claims

1. A motion control optimization method for large bulk material equipment based on deep learning, characterized in that: The following steps are involved: Step 1: Collect material field environmental data through multi-source sensors, extract robust features of point clouds and images based on curvature features and the improved ORB algorithm, and implement collaborative preprocessing of multi-source data through spatiotemporal alignment and normalization. Step 2: Build a cross-modal attention fusion model. Through feature matching optimization and dynamic weight allocation under geometric constraints, output a multimodal joint feature expression with complementary advantages. Specifically, it includes the following steps: Step 2.1: The cross-modal attention fusion model uses a multi-head scaled dot-product attention mechanism to establish geometric consistency constraints on the spatial alignment module. It stably extracts common features and complementary information across modalities in complex environments and maps them to a unified feature space using a learnable linear transformation: Where Q is the active query feature, K is the matched feature, V is the actual transmitted information, and F pc is the point cloud feature matrix, F img is the image feature matrix, W Q 、W K 、W V For the learnable projection matrix, the Hamming distance D is used H Measuring feature similarity: where desc A 、desc B are the descriptors of feature points A and B respectively, is the dynamic Hamming distance threshold, is the basic threshold, is the threshold adjustment amplitude, Normalize the contrast of the image and keep The matching pairs are taken as candidate corresponding points; Step 2.2: Then perform feature matching and construct the optimization objective function E(T) including geometric error and motion prior: Where T is the transformation matrix describing the projection relationship from point cloud to image, q i is the projected 2D image point, p i is any point in the point cloud, ρ is the Huber robust kernel function, π is the camera projection model, λ is the regularization coefficient, T prev represents the transformation matrix of the previous frame, e is the input error, δ is the threshold parameter, and the objective function is iteratively solved by the Gauss-Newton algorithm until convergence; Step 2.3: Perform nonlinear mapping on the optimized matching results obtained by optimizing the objective function, and use the multi-head attention mechanism to capture the dependencies between features from multiple dimensions. The attention weight Attention(Q,K,V) is calculated as follows: Among them QK T is the association strength between each point in the point cloud and each area of the image, is the scaling factor, and softmax is the row-by-row normalization process; The multi-head scaling dot product attention mechanism is used to expand the projected features into h attention heads. Each attention head is calculated independently. The multi-head attention expansion calculation is as follows: MultiHead(Q,K,V)=Concat(head1,...,head h )W Q Among them head1 to head h For single-head attention output, a cross-modal attention weight mechanism is designed and calculated as follows: Finally, the output environment state representation F out : F out =LayerNorm(F fusion +FFN(F fusion )) Among them F fusion is the fusion feature of the cross-modal attention output, FFN is the feedforward neural network, and LayerNorm is the layer normalization processing; Step 3: Build an LSTM-PPO hybrid decision network, use the long-term dependency modeling capability of LSTM to extract state features, and combine the policy gradient optimization of the PPO algorithm to achieve adaptive decision-making in a dynamic environment to obtain the final optimization strategy. The specific steps include: Step 3.1: The LSTM-PPO decision network architecture is based on the LSTM network and the PPO algorithm to achieve the control of the movement of large bulk material equipment. The network model consists of an input layer, an LSTM layer, a fully connected layer, an Actor network, and a Critic network. The input layer receives multi-source data after feature extraction and matching, including environmental status representation, equipment operation status and material flow data; the number of hidden units in the LSTM layer is 128 and the number of layers is 2; Step 3.2: Fusion of LSTM-encoded temporal features h t , build an Actor-Critic framework, realize the iterative update of control strategy by alternately optimizing the strategy network and the value network, and the Actor outputs the continuous control action a t : a t =tanh[W a ·ReLU(W1·h t +b1)+b a ] Where W1, b1 are the parameters of the fully connected layer, ReLU is the activation function, W a , b a As the output layer parameter, tanh constrains the action to [-1,1] and then linearly maps it to the actual control range; Critic estimates the discounted cumulative reward V(s) starting from state s at time t t ) is calculated as follows: V(s t )=W v ·ReLU(W2·h t +b2)+b v Where W2, b2 are the parameters of the fully connected layer, W v , b v is the output layer parameter; Step 3.3: Based on the PPO algorithm, an end-to-end training method is used, the training batch size is set to 64, the number of optimization iterations is set to 10, the cumulative discounted return is used as the optimization target, the amplitude of the policy update is limited to avoid the policy crash caused by a single update being too large, and the advantage function A t The calculation is as follows: δ t =r t +γV(s t+1 )-V(s t ) r t ＝[0.6α+0.7β] The discount factor γ is set to 0.95, the smoothing coefficient λ is set to 0.9, and r t is the immediate reward, α is the time required to reach the control target, β is the spatial offset between the target position and the actual position, δ t For the time series difference error at time t, set the objective function: L CLIP (θ)=E t {min[r t (i)A t ,clip(r t (θ),1-ε,1+ε)A t ]} where r t (θ) is the ratio of the new and old strategies, ε is the clipping threshold set to 0.2, and π θ For a random strategy, the training objective function and the total loss function are calculated as follows: L total =-L CLIP (θ)+c1L VF -c2H(π θ ) The critic loss weight c1 is set to 0.5, c2 is the policy entropy weight set to 0.02, and L VF is the value loss function, H(π θ ) is the policy entropy.

2. The method for optimizing motion control of large-scale bulk material equipment based on deep learning according to claim 1, characterized in that: In step 1, the material field environment data is collected through multi-source sensors, and robust features of point clouds and images are extracted based on curvature features and the improved ORB algorithm. The collaborative preprocessing of multi-source data is achieved by using spatiotemporal alignment and normalization. The specific steps are as follows: Step 1.1: Use LiDAR to acquire real-time 3D point cloud data of the pile, including its shape, height, and surface profile. RGB industrial cameras collect surface image data of the pile, extracting its color distribution, texture characteristics, and particle size information to assist in identifying its type and stacking state. Inertial sensors acquire real-time motion status of the equipment, including spatial position, acceleration, and angular velocity. Laser flow meters acquire real-time material flow data. Step 1.2: Use the curvature-based feature point detection method for the point cloud data obtained by the lidar. For any point p in the point cloud, i , construct its local neighborhood point set N(p i ), the local geometric features are characterized by calculating the covariance matrix C of the neighborhood points: where p j is a point in the neighborhood, μ is the centroid of the neighborhood point set, T is the transpose, and the eigenvalue λ0≤λ1≤λ2 is obtained according to the covariance matrix, and then the point p is calculated. i The curvature K; τ=0.04e -0.3ρ Retention satisfaction point, is the threshold, ρ is the local point cloud density, and the key features of the stockpile contour are screened out; Step 1.3: Use the improved ORB feature detector to extract visual features from the image collected by the RGB industrial camera, and use the FAST feature detection algorithm with adaptive threshold to calculate the feature point response value: Where Ω represents the detection window centered at (x, y), I(u, v) is the pixel intensity value, I(x, y) is the center pixel intensity value, and the feature point N is dynamically controlled according to the overall image contrast. f quantity: where N max is the maximum feature point, Represents the image contrast measurement value, and k is the adjustment coefficient.

Images