Machine learning based autonomous loader safety control method and system

Abstract

The application discloses a safety control method and system for unmanned loader based on machine learning, and relates to the technical field of safety control.The method comprises the following steps: processing multi-source sensor data by using an anti-vibration feature extraction network to generate a safety evaluation coefficient; analyzing a safety state based on a variational autoencoder model to generate structured safety warning information; applying reinforcement learning for multi-objective optimization to output a safe driving trajectory; and finally converting the trajectory into a control instruction through a deep learning controller and adjusting the control parameters in real time based on an online learning mechanism.The application realizes intelligent safety control of the unmanned loader under complex working conditions by constructing a whole-process machine learning processing chain from data acquisition to control execution, and solves the problems of poor safety control adaptability, insufficient cooperation between control modules and vibration interference affecting the accuracy of judgment in the prior art.

Description

Technical Field

[0001] This invention relates to the field of safety control technology, and in particular to a safety control method and system for unmanned loaders based on machine learning. Background Technology

[0002] With the rapid development of smart mines and intelligent construction, the operational safety of unmanned loaders under complex conditions has become a key focus of the industry. In typical operating environments such as open-pit mines and construction sites, loaders face complex scenarios such as high vibration interference, frequent changes in operating conditions, and mixed human-machine operations, which places higher demands on safety control systems.

[0003] In related technologies, safety control of unmanned loaders mainly relies on preset rules and fixed thresholds, such as performing emergency braking when an obstacle is detected, or implementing anti-tipping control based on a preset tilt angle threshold. However, these methods suffer from poor adaptability in safety control: on the one hand, fixed thresholds are difficult to adapt to dynamically changing operating environments, and are prone to misjudgment when material humidity changes or road conditions change; on the other hand, each safety control module operates independently, lacking coordinated optimization, resulting in delayed control response under complex working conditions, affecting operational safety and efficiency. Furthermore, traditional methods are insufficient in handling the vibration interference unique to loaders, affecting sensor data quality and further reducing the accuracy of safety judgments. Summary of the Invention

[0004] This invention provides a machine learning-based method and system for the safety control of unmanned loaders, which solves the problems of poor adaptability of safety control, insufficient coordination among control modules, and the impact of vibration interference on the accuracy of judgment in the prior art, and realizes intelligent safety control of unmanned loaders under complex working conditions.

[0005] This invention provides a machine learning-based safety control method for unmanned loaders, which is applied to a machine learning-based safety control system for unmanned loaders, and includes:

[0006] Collect multi-source sensor data from the unmanned loader, including vibration sensor data, position sensor data, and environmental sensor data. Align the three types of data by timestamp and normalize them to form a unified three-dimensional feature vector.

[0007] The three-dimensional feature vector is input into the vibration resistance feature extraction network, which uses a temporal convolutional layer to separate vibration frequency band signals, uses a self-attention mechanism to enhance effective features, and generates safety assessment coefficients through weighted fusion.

[0008] The safety assessment coefficients are input into the variational autoencoder model, the latent variable distribution is generated by the encoder, the safety assessment coefficients are reconstructed by the decoder, the reconstruction error is calculated, and the safety status is classified as risky or non-risk based on the Bayesian posterior distribution and adaptive threshold mechanism, generating structured safety early warning information containing risk level, location information and timestamp.

[0009] The structured safety warning information is input into the reinforcement learning agent, a preliminary trajectory is generated through the proximal policy optimization algorithm, the expectation-maximization algorithm is applied to iteratively optimize the path planner parameters and the evaluation model parameters, and a safe driving trajectory is output.

[0010] The safe driving trajectory is input into the deep learning controller, which generates control commands through the feedforward network. Based on real-time control effect feedback, the controller parameters are adjusted through an online learning mechanism to control the operation of the unmanned loader.

[0011] Furthermore, the processing steps of the vibration-damping feature extraction network include:

[0012] The input data is processed using a temporal convolutional layer, and the kernel parameters of the temporal convolutional layer are optimized using a formula. It is confirmed that, among them, This represents the optimized filter parameters. This represents the convolution operation. Indicates input data, This represents the mean squared error loss function. express Regularization term, This represents the first balance coefficient. Indicates data distribution, Indicates filter parameters;

[0013] The output features of the temporal convolutional layer Input a self-attention mechanism, calculate the attention weight matrix, and obtain the enhanced features. ,in Indicates the sequence length. , Indicates feature dimension;

[0014] Output features and enhanced features Mapping to a unified dimension via a fully connected layer: , ,in , Represents the weight matrix. , Indicates the bias term;

[0015] Weighted fusion based on time steps: ,in This represents the safety assessment coefficient at time t. This represents the convolutional features at time t. This represents the attention characteristics at time t. This represents the weight coefficients learned through a neural network.

[0016] Furthermore, the steps for processing the input data by the temporal convolutional layer include:

[0017] Initialize filter parameters Set the first balance coefficient The search scope;

[0018] From the distribution of training data Medium-sampled input data Where i is the index of the sampled data and N is the total number of sampled data, calculate the convolution output. ;

[0019] Calculate the mean square error loss and regularization term ;

[0020] The filter parameters are updated iteratively using the gradient descent method. This continues until the loss function converges, yielding the optimized filter parameters. .

[0021] Furthermore, the step of generating structured security early warning information using the variational autoencoder model includes:

[0022] Safety assessment coefficient Input encoder network, output the mean of the latent variable distribution and variance ;

[0023] From the distribution using reparameterization techniques Mid-sampled latent variables ;

[0024] latent variables Input the decoder network to reconstruct the security evaluation coefficients. ;

[0025] Calculating the lower bound of evidence ,in Indicates the likelihood of reconstruction. Indicates the standard Gaussian prior;

[0026] Based on the mean of the latent variable distribution Calculate the adaptive threshold ,in This is the preset sensitivity coefficient;

[0027] Compare reconstruction errors With threshold If reconstruction error If the value exceeds the threshold, it is classified as a risky state; otherwise, it is classified as a non-risky state.

[0028] Generate structured security warning information that includes risk status, risk level, location coordinates, and timestamp.

[0029] Furthermore, the training process of the variational autoencoder model includes:

[0030] Initialize the encoder and decoder network parameters;

[0031] Security evaluation coefficients sampled from the training set Calculate the distribution parameters of the latent variables , ;

[0032] Latent variables are obtained through reparameterized sampling. ,in, , This is the convergence threshold;

[0033] Decoder Reconstruction Security Assessment Coefficient Calculate reconstruction loss and KL divergence ;

[0034] Total loss ,in This is the second balance coefficient;

[0035] Backpropagation updates the network parameters; this process is repeated until convergence.

[0036] Furthermore, the step of the reinforcement learning agent outputting a safe driving trajectory includes:

[0037] Initialize the policy network and value network of the reinforcement learning agent;

[0038] Structured security early warning information is used as state input, and the policy network outputs the action value distribution.

[0039] The policy parameters are updated using a near-end policy optimization algorithm to generate a preliminary trajectory.

[0040] In the E-step of the expectation-maximization algorithm, the posterior distribution of the latent variables is calculated. ,in Indicates a cooperative state. This indicates a warning message;

[0041] In step M, update the path planner parameters. and evaluation model parameters : , ;

[0042] in, For the updated path planner parameters, These are the parameters from the previous evaluation model. Here are the updated evaluation model parameters, where p is the probability. For expectations;

[0043] Repeat the E-step and M-step until the parameters converge, and output the optimized safe driving trajectory.

[0044] Furthermore, the specific steps of the expectation-maximization algorithm include:

[0045] Define hidden variables The discrete state space is {well coordinated, partially coordinated, and poorly coordinated};

[0046] E-step calculation: ,in Let be the prior probability of the state. Let the likelihood probability be denoted as '(').

[0047] M-step update:

[0048] , ,in , The learning rate;

[0049] Convergence criterion is and ,in This is the convergence threshold.

[0050] Furthermore, the step of the deep learning controller generating execution instructions includes:

[0051] The safe driving trajectory is input into the feedforward network of the deep learning controller. The structure of the feedforward network consists of alternating fully connected layers and activation functions.

[0052] The forward network outputs preliminary control commands, including steering angle, throttle, and brake control values;

[0053] Execute initial control commands and collect control effect feedback, including trajectory tracking error. Stability indicators and safety indicators ;

[0054] Calculate the loss function ,in , , These are the weighting coefficients;

[0055] Update the parameters of the deep learning controller using gradient descent: ,in The learning rate;

[0056] Repeat the process until the control effect meets the requirements.

[0057] This invention provides a machine learning-based safety control system for unmanned loaders, which implements a machine learning-based safety control method for unmanned loaders, including a data acquisition module, a feature extraction module, a safety assessment module, a trajectory planning module, and a control execution module.

[0058] The data acquisition module is used to collect multi-source sensor data from the unmanned loader, including vibration sensor data, position sensor data and environmental sensor data. The three types of data are aligned by timestamp and normalized to form a unified three-dimensional feature vector.

[0059] The feature extraction module is used to input the three-dimensional feature vector into the vibration-resistant feature extraction network. The vibration-resistant feature extraction network uses a temporal convolutional layer to separate vibration frequency band signals, uses a self-attention mechanism to enhance effective features, and generates safety assessment coefficients through weighted fusion.

[0060] The security assessment module is used to input security assessment coefficients into a variational autoencoder model, generate a latent variable distribution through the encoder, reconstruct the security assessment coefficients through the decoder, calculate the reconstruction error, classify the security status as risky or non-risk based on Bayesian posterior distribution and adaptive threshold mechanism, and generate structured security early warning information containing risk level, location information and timestamp.

[0061] The trajectory planning module is used to input structured safety warning information into the reinforcement learning agent, generate a preliminary trajectory through the near-end policy optimization algorithm, apply the expectation-maximization algorithm to iteratively optimize the path planner parameters and evaluation model parameters, and output a safe driving trajectory.

[0062] The control execution module is used to input the safe driving trajectory into the deep learning controller, generate control commands through the feedforward network, and adjust the controller parameters through an online learning mechanism based on real-time control effect feedback to control the operation of the unmanned loader.

[0063] Compared with the prior art, the present invention has at least the following technical effects:

[0064] The present invention provides a machine learning-based safety control method and system for unmanned loaders. This method constructs a three-dimensional feature vector through multi-source sensor data acquisition and fusion; it employs an anti-vibration feature extraction network to process sensor data, uses a temporal convolutional layer to separate vibration interference, and a self-attention mechanism to enhance effective features, generating safety assessment coefficients; it analyzes the safety status based on a variational autoencoder model, and generates structured safety warning information through Bayesian posterior distribution and adaptive threshold mechanisms; it applies reinforcement learning for multi-objective optimization, combining the expectation-maximization algorithm to achieve synergistic optimization of path planning and safety assessment, outputting a safe driving trajectory; finally, it uses a deep learning controller to convert the trajectory into control commands and adjusts control parameters in real time based on an online learning mechanism.

[0065] This invention effectively separates vibration interference signals from actual operating signals through multi-level processing of an anti-vibration feature extraction network, improving sensor data quality. Furthermore, in the safety status identification stage, a variational autoencoder combined with a dynamic calculation method based on Bayesian posterior distribution is employed to achieve adaptive threshold adjustment, avoiding misjudgments caused by fixed thresholds in dynamic environments and improving the accuracy and adaptability of safety status identification. Further, through the synergistic optimization of reinforcement learning and expectation-maximization algorithms, path planning and safety assessment are established, optimizing operational efficiency while ensuring safety, achieving a balanced improvement in safety control and operational effectiveness. Finally, through the online learning mechanism of a deep learning controller, control parameters are continuously optimized to adapt to different operating conditions, improving the overall robustness and control accuracy of the system. Attached Figure Description

[0066] Figure 1 A flowchart of a machine learning-based safety control method for unmanned loaders provided in an embodiment of the present invention;

[0067] Figure 2 This is a schematic diagram of the structure of a machine learning-based unmanned loader safety control system provided in an embodiment of the present invention. Detailed Implementation

[0068] This invention provides a machine learning-based safety control method and system for unmanned loaders, which solves the problems of poor adaptability of safety control, insufficient coordination among control modules, and the impact of vibration interference on the accuracy of judgment in the prior art. By constructing a full-process machine learning processing chain from data acquisition to control execution, it realizes intelligent safety control of unmanned loaders under complex working conditions.

[0069] In related technologies, safety control methods based on fixed thresholds do not consider the dynamic characteristics of the loader's operating environment. Factors such as changes in material pile shape and fluctuations in road surface adhesion coefficient can cause changes in the distribution characteristics of sensor data, making it difficult for fixed thresholds to adapt to these dynamic changes. Independent decision-making by each safety control module, lacking information sharing and collaborative optimization, leads to delays or conflicts in control response. Furthermore, strong vibration interference during loader operation affects the quality of sensor data; traditional filtering methods struggle to effectively separate vibration noise from valid signals, further impacting the accuracy of safety assessments.

[0070] To address the aforementioned technical challenges, this invention provides a machine learning-based safety control method and system for unmanned loaders. This system achieves adaptive safety control through machine learning. It employs a vibration-resistant feature extraction network to enhance signal processing capabilities under vibration conditions, utilizes a variational autoencoder to achieve data-driven dynamic assessment of safety status, applies reinforcement learning and expectation-maximization algorithms to achieve multi-module collaborative optimization, and ultimately continuously improves the adaptability and robustness of the control system under complex operating conditions through an online learning mechanism.

[0071] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0072] Example 1

[0073] like Figure 1 As shown, this embodiment of the invention provides a machine learning-based safety control method for unmanned loaders. This method is applied to a machine learning-based safety control system for unmanned loaders and includes:

[0074] Collect multi-source sensor data from the unmanned loader, including vibration sensor data, position sensor data, and environmental sensor data. Align the three types of data by timestamp and normalize them to form a unified three-dimensional feature vector.

[0075] The three-dimensional feature vector is input into the vibration-resistant feature extraction network, which uses a temporal convolutional layer to separate vibration frequency band signals, uses a self-attention mechanism to enhance effective features, and generates safety assessment coefficients through weighted fusion.

[0076] The safety assessment coefficients are input into the variational autoencoder model, the latent variable distribution is generated by the encoder, the safety assessment coefficients are reconstructed by the decoder, the reconstruction error is calculated, and the safety status is classified as risky or non-risk based on the Bayesian posterior distribution and adaptive threshold mechanism, generating structured safety early warning information containing risk level, location information and timestamp.

[0077] The structured safety warning information is input into the reinforcement learning agent, a preliminary trajectory is generated through the proximal policy optimization algorithm, the expectation-maximization algorithm is applied to iteratively optimize the path planner parameters and the evaluation model parameters, and a safe driving trajectory is output.

[0078] The safe driving trajectory is input into the deep learning controller, which generates control commands through the feedforward network. Based on real-time control effect feedback, the controller parameters are adjusted through an online learning mechanism to control the operation of the unmanned loader.

[0079] In this embodiment, multi-source sensor data is preprocessed and input into the vibration-resistant network. The output safety assessment coefficients reflect the system's safety status. After classification by a variational autoencoder, structured early warning information containing fields such as risk level and location is generated. The reinforcement learning module generates a trajectory based on this information, which is ultimately converted into control commands by the controller. The normalization processing of the three-dimensional feature vectors adopts the max-min normalization method, scaling the data of each sensor to the [0,1] interval to eliminate the differences in dimensions between different sensors. The temporal convolutional layer adopts a one-dimensional causal convolutional structure to ensure that the output depends only on the current and historical inputs, avoiding future information leakage. During the calculation of the self-attention mechanism, lower weights are assigned to time steps with higher vibration intensity to achieve feature enhancement under vibration conditions. The reconstruction error calculation of the variational autoencoder uses mean square error, and the safety status classification threshold is dynamically adjusted according to the statistical characteristics of historical data. The state space of the reinforcement learning agent includes the current position, velocity, attitude, and early warning information, while the action space includes steering, throttle, and braking control quantities. The online learning of the deep learning controller adopts a sliding window mechanism, maintaining the control effect data of the most recent N time steps for parameter updates.

[0080] Furthermore, the processing steps of the vibration-damping feature extraction network include:

[0081] The input data is processed using a temporal convolutional layer, and the kernel parameters of the temporal convolutional layer are optimized using a formula. It is confirmed that, among them, This represents the optimized filter parameters. This represents the convolution operation. Indicates input data, This represents the mean squared error loss function. express Regularization term, This represents the first balance coefficient. Indicates data distribution, Indicates filter parameters;

[0082] The output features of the temporal convolutional layer Input a self-attention mechanism, calculate the attention weight matrix, and obtain the enhanced features. ,in Indicates the sequence length. , Indicates feature dimension;

[0083] Output features and enhanced features Mapping to a unified dimension via a fully connected layer: , ,in , Represents the weight matrix. , Indicates the bias term;

[0084] Weighted fusion based on time steps: ,in This represents the safety assessment coefficient at time t. This represents the convolutional features at time t. This represents the attention characteristics at time t. This represents the weight coefficients learned through a neural network.

[0085] In this embodiment, the first balance coefficient in the filter parameter optimization formula Cross-validation was used to determine the optimal method for testing on the training set. Considering the performance of [0.1, 0.2, ..., 1.0], select the one with the minimum loss on the validation set. value.

[0086] The weight matrix , and bias terms , Learning is achieved through backpropagation, and initialization is performed using the Xavier initialization method.

[0087] The adaptive weight coefficient α is learned through a single-layer neural network. The input is a concatenation of temporal convolutional features and attention features, and the output is a weight value in the interval [0,1].

[0088] The temporal convolutional layer outputs features Let d represent the convolutional feature vector after feature mapping at time t, with dimension d.

[0089] The attention features This represents the attention feature vector after feature mapping at time t, with dimensions equal to 1. same.

[0090] Furthermore, the steps for processing the input data by the temporal convolutional layer include:

[0091] Initialize filter parameters Set the first balance coefficient The search scope;

[0092] From the distribution of training data Medium-sampled input data Where i is the index of the sampled data and N is the total number of sampled data, calculate the convolution output. ;

[0093] Calculate the mean square error loss and regularization term ;

[0094] The filter parameters are updated iteratively using the gradient descent method. This continues until the loss function converges, yielding the optimized filter parameters. .

[0095] In this embodiment, the filter parameter P is initialized using the He initialization method to ensure stability during the initial training phase.

[0096] The gradient descent method uses the Adam optimizer with a learning rate of 0.001.

[0097] The training data distribution D comes from historically collected loader operation data, including sensor readings under different operating conditions.

[0098] When calculating the mean square error loss, higher weights are given to periods with stronger vibration signals to improve separation accuracy under vibration conditions.

[0099] Furthermore, the step of generating structured security early warning information using the variational autoencoder model includes:

[0100] Safety assessment coefficient Input encoder network, output the mean of the latent variable distribution and variance ;

[0101] From the distribution using reparameterization techniques Mid-sampled latent variables ;

[0102] latent variables Input the decoder network to reconstruct the security evaluation coefficients. ;

[0103] Calculating the lower bound of evidence ,in Indicates the likelihood of reconstruction. Indicates the standard Gaussian prior;

[0104] Based on the mean of the latent variable distribution Calculate the adaptive threshold ,in This is the preset sensitivity coefficient;

[0105] Compare reconstruction errors With threshold If reconstruction error If the value exceeds the threshold, it is classified as a risky state; otherwise, it is classified as a non-risky state.

[0106] Generate structured security warning information that includes risk status, risk level, location coordinates, and timestamp.

[0107] In this embodiment, the encoder network structure is a three-layer fully connected layer with 128, 64 and 32 neurons respectively, and the activation function used is ReLU.

[0108] The decoder network structure is symmetrical to the encoder, and the Sigmoid activation function is used to ensure that the output is in the range [0,1].

[0109] The sensitivity coefficient k is set according to the safety requirements of the application scenario. When the safety requirements are high, a smaller value is taken to improve the detection sensitivity.

[0110] The structured security warning information is in JSON format.

[0111] Furthermore, the training process of the variational autoencoder model includes:

[0112] Initialize the encoder and decoder network parameters;

[0113] Security evaluation coefficients sampled from the training set Calculate the distribution parameters of the latent variables , ;

[0114] Latent variables are obtained through reparameterized sampling. ,in, , This is the convergence threshold;

[0115] Decoder Reconstruction Security Assessment Coefficient Calculate reconstruction loss and KL divergence ;

[0116] Total loss ,in This is the second balance coefficient;

[0117] Backpropagation updates the network parameters; this process is repeated until convergence.

[0118] In this embodiment, the second balance coefficient β adopts an annealing scheduling strategy, with β=0 in the early stage of training and gradually increasing to β=1, to avoid the KL divergence term dominating the training process.

[0119] When calculating the reconstruction loss, different weights are assigned to different dimensions of the safety assessment coefficient, with more important dimensions having higher weights.

[0120] The training set includes safety assessment coefficients under normal and abnormal operating conditions, ensuring that the model learns the differences between normal and abnormal modes.

[0121] Furthermore, the step of the reinforcement learning agent outputting a safe driving trajectory includes:

[0122] Initialize the policy network and value network of the reinforcement learning agent;

[0123] Structured security early warning information is used as state input, and the policy network outputs the action value distribution.

[0124] A preliminary trajectory is generated by updating the policy network parameters using a proximal policy optimization algorithm and utilizing a value network.

[0125] In the E-step of the expectation-maximization algorithm, the posterior distribution of the latent variables is calculated. ,in Indicates a cooperative state. This indicates a warning message;

[0126] In step M, update the path planner parameters. and evaluation model parameters : , ;

[0127] in, For the updated path planner parameters, These are the parameters from the previous evaluation model. Here are the updated evaluation model parameters, where p is the probability. For expectations;

[0128] Repeat the E-step and M-step until the parameters converge, and output the optimized safe driving trajectory.

[0129] In this embodiment, both the policy network and the value network adopt a two-layer fully connected network with a hidden layer dimension of 256 and the activation function is Tanh.

[0130] The cooperative state of the latent variable h is defined based on path tracking error and control stability index: good coordination: tracking error < 0.1m and control fluctuation < 5%; partial coordination: tracking error ∈ [0.1, 0.3]m or control fluctuation ∈ [5%, 10%]; insufficient coordination: tracking error > 0.3m or control fluctuation > 10%.

[0131] The parameter convergence criterion is that the parameter change in three consecutive iterations is less than the threshold 1e-4.

[0132] Furthermore, the specific steps of the expectation-maximization algorithm include:

[0133] Define hidden variables The discrete state space is {well coordinated, partially coordinated, and poorly coordinated};

[0134] E-step calculation:

[0135] ,in Let be the prior probability of the state. Let the likelihood probability be denoted as '(').

[0136] M-step update:

[0137] , ,in , The learning rate;

[0138] Convergence criterion is and ,in This is the convergence threshold.

[0139] In this embodiment, the prior state probability Initially set to uniform distribution, it is dynamically adjusted according to the actual frequency of cooperative states during training.

[0140] The learning rate , An adaptive adjustment strategy is adopted, with an initial value of 0.01, which is dynamically adjusted according to the gradient magnitude.

[0141] The convergence threshold The value is set according to the accuracy requirements of the actual application, with a typical value of 1e-4.

[0142] Furthermore, the step of the deep learning controller generating execution instructions includes:

[0143] The safe driving trajectory is input into the feedforward network of the deep learning controller. The structure of the feedforward network consists of alternating fully connected layers and activation functions.

[0144] The forward network outputs preliminary control commands, including steering angle, throttle, and brake control values;

[0145] Execute initial control commands and collect control effect feedback, including trajectory tracking error. Stability indicators and safety indicators ;

[0146] Calculate the loss function ,in , , These are the weighting coefficients;

[0147] Update the parameters of the deep learning controller using gradient descent: ,in The learning rate;

[0148] Repeat the process until the control effect meets the requirements.

[0149] In this embodiment, the feedforward network contains four fully connected layers with 256, 128, 64, and 3 neurons, respectively, corresponding to three control outputs.

[0150] The trajectory tracking error The mean square error between the actual trajectory and the expected trajectory is calculated.

[0151] The stability index The calculation is the sum of squares of the rate of change of the control quantity.

[0152] The security indicators The calculation is a weighted sum of the degrees of violation of safety constraints.

[0153] The weighting coefficient , , Based on expert experience, the typical value is set as follows: =0.6, =0.3, =0.1.

[0154] like Figure 2 As shown, this embodiment of the invention provides a machine learning-based unmanned loader safety control system for implementing the machine learning-based unmanned loader safety control method, including: a data acquisition module, a feature extraction module, a safety assessment module, a trajectory planning module, and a control execution module;

[0155] The data acquisition module is used to collect multi-source sensor data from the unmanned loader, including vibration sensor data, position sensor data and environmental sensor data. The three types of data are aligned by timestamp and normalized to form a unified three-dimensional feature vector.

[0156] The feature extraction module is used to input the three-dimensional feature vector into the vibration-resistant feature extraction network. The vibration-resistant feature extraction network uses a temporal convolutional layer to separate vibration frequency band signals, uses a self-attention mechanism to enhance effective features, and generates safety assessment coefficients through weighted fusion.

[0157] The security assessment module is used to input security assessment coefficients into a variational autoencoder model, generate a latent variable distribution through the encoder, reconstruct the security assessment coefficients through the decoder, calculate the reconstruction error, classify the security status as risky or non-risk based on Bayesian posterior distribution and adaptive threshold mechanism, and generate structured security early warning information containing risk level, location information and timestamp.

[0158] The trajectory planning module is used to input structured safety warning information into the reinforcement learning agent, generate a preliminary trajectory through the near-end policy optimization algorithm, apply the expectation-maximization algorithm to iteratively optimize the path planner parameters and evaluation model parameters, and output a safe driving trajectory.

[0159] The control execution module is used to input the safe driving trajectory into the deep learning controller, generate control commands through the feedforward network, and adjust the controller parameters through an online learning mechanism based on real-time control effect feedback to control the operation of the unmanned loader.

[0160] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0161] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0162] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0163] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0164] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0165] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based safety control method for unmanned loaders, characterized in that, Includes the following steps: Collect multi-source sensor data from the unmanned loader, including vibration sensor data, position sensor data, and environmental sensor data. Align the three types of data by timestamp and normalize them to form a unified three-dimensional feature vector. The three-dimensional feature vector is input into the vibration resistance feature extraction network, which uses a temporal convolutional layer to separate vibration frequency band signals, uses a self-attention mechanism to enhance effective features, and generates safety assessment coefficients through weighted fusion. The safety assessment coefficients are input into the variational autoencoder model, the latent variable distribution is generated by the encoder, the safety assessment coefficients are reconstructed by the decoder, the reconstruction error is calculated, and the safety status is classified as risky or non-risk based on the Bayesian posterior distribution and adaptive threshold mechanism, generating structured safety early warning information containing risk level, location information and timestamp. The structured safety warning information is input into the reinforcement learning agent, a preliminary trajectory is generated through the proximal policy optimization algorithm, the expectation-maximization algorithm is applied to iteratively optimize the path planner parameters and the evaluation model parameters, and a safe driving trajectory is output. The safe driving trajectory is input into the deep learning controller, which generates control commands through the feedforward network. Based on real-time control effect feedback, the controller parameters are adjusted through an online learning mechanism to control the operation of the unmanned loader.

2. The machine learning-based safety control method for unmanned loaders as described in claim 1, characterized in that, The processing steps of the vibration-damping feature extraction network include: The input data is processed using a temporal convolutional layer, and the kernel parameters of the temporal convolutional layer are optimized using a formula. It is confirmed that, among them, This represents the optimized filter parameters. This represents the convolution operation. Indicates input data, This represents the mean squared error loss function. express Regularization term, This represents the first balance coefficient. Indicates data distribution, Indicates filter parameters; The output features of the temporal convolutional layer Input a self-attention mechanism, calculate the attention weight matrix, and obtain the enhanced features. ,in Indicates the sequence length. , Indicates feature dimension; Output features and enhanced features Mapping to a unified dimension via a fully connected layer: , ,in , Represents the weight matrix. , Indicates the bias term; Weighted fusion based on time steps: ,in This represents the safety assessment coefficient at time t. This represents the convolutional features at time t. This represents the attention characteristics at time t. This represents the weight coefficients obtained through neural network learning.

3. The machine learning-based safety control method for unmanned loaders as described in claim 2, characterized in that, The steps of the temporal convolutional layer in processing the input data include: Initialize filter parameters Set the first balance coefficient The search scope; From the distribution of training data Medium-sampled input data Where i is the index of the sampled data and N is the total number of sampled data, calculate the convolution output. ; Calculate the mean square error loss and regularization term ; The filter parameters are updated iteratively using the gradient descent method. This continues until the loss function converges, yielding the optimized filter parameters. .

4. The machine learning-based safety control method for unmanned loaders as described in claim 1, characterized in that, The steps for generating structured security early warning information using the variational autoencoder model include: Safety assessment coefficient Input encoder network, output the mean of the latent variable distribution and variance ; From the distribution using reparameterization techniques Medium-sampled latent variables ; latent variables Input the decoder network to reconstruct the security evaluation coefficients. ; Calculating the lower bound of evidence ,in Indicates the likelihood of reconstruction. This represents the standard Gaussian prior; Based on the mean of the latent variable distribution Calculate the adaptive threshold ,in This is the preset sensitivity coefficient; Compare reconstruction errors With threshold If reconstruction error Greater than the threshold If it is classified as a risky state, then it is classified as a non-risky state. Generate structured security warning information that includes risk status, risk level, location coordinates, and timestamp.

5. The machine learning-based safety control method for unmanned loaders as described in claim 4, characterized in that, The training process of the variational autoencoder model includes: Initialize the encoder and decoder network parameters; Security evaluation coefficients sampled from the training set Calculate the distribution parameters of the latent variables , ; Latent variables are obtained through reparameterized sampling. ,in, , This is the convergence threshold; Decoder Reconstruction Security Assessment Coefficient Calculate reconstruction loss and KL divergence ; Total loss ,in This is the second balance coefficient; Backpropagation updates the network parameters; this process is repeated until convergence.

6. The machine learning-based safety control method for unmanned loaders as described in claim 1, characterized in that, The steps for the reinforcement learning agent to output a safe driving trajectory include: Initialize the policy network and value network of the reinforcement learning agent; Structured security early warning information is used as state input, and the policy network outputs the action value distribution. A preliminary trajectory is generated by updating the policy network parameters using a proximal policy optimization algorithm and utilizing a value network. In the E-step of the expectation-maximization algorithm, the posterior distribution of the latent variables is calculated. ,in Indicates a cooperative state. This indicates a warning message; In step M, update the path planner parameters. and evaluation model parameters : , ; in, For the updated path planner parameters, These are the parameters from the previous evaluation model. Here are the updated evaluation model parameters, where p is the probability. For expectations; Repeat the E-step and M-step until the parameters converge, and output the optimized safe driving trajectory.

7. The machine learning-based safety control method for unmanned loaders as described in claim 6, characterized in that, The specific steps of the expectation maximization algorithm include: Define hidden variables The discrete state space is {well coordinated, partially coordinated, and poorly coordinated}; E-step calculation: ,in Let be the prior probability of the state. Let the likelihood probability be denoted as '('). M-step update: , ,in , The learning rate; Convergence criterion is and ,in This is the convergence threshold.

8. The machine learning-based safety control method for unmanned loaders as described in claim 1, characterized in that, The steps for the deep learning controller to generate execution instructions include: The safe driving trajectory is input into the feedforward network of the deep learning controller. The structure of the feedforward network consists of alternating fully connected layers and activation functions. The forward network outputs preliminary control commands, including steering angle, throttle, and brake control values; Execute initial control commands and collect control effect feedback, including trajectory tracking error. Stability indicators and safety indicators ; Calculate the loss function ,in , , These are the weighting coefficients; Update the parameters of the deep learning controller using gradient descent: ,in The learning rate; Repeat the process until the control effect meets the requirements.

9. A machine learning-based safety control system for unmanned loaders, used to implement the machine learning-based safety control method for unmanned loaders as described in any one of claims 1-8, characterized in that, include: Data acquisition module, feature extraction module, safety assessment module, trajectory planning module, control execution module; The data acquisition module is used to collect multi-source sensor data from the unmanned loader, including vibration sensor data, position sensor data, and environmental sensor data. The three types of data are aligned by timestamp and normalized to form a unified three-dimensional feature vector. The feature extraction module is used to input the three-dimensional feature vector into the vibration-resistant feature extraction network. The vibration-resistant feature extraction network uses a temporal convolutional layer to separate vibration frequency band signals, uses a self-attention mechanism to enhance effective features, and generates safety assessment coefficients through weighted fusion. The security assessment module is used to input security assessment coefficients into a variational autoencoder model, generate a latent variable distribution through the encoder, reconstruct the security assessment coefficients through the decoder, calculate the reconstruction error, classify the security status as risky or non-risk based on Bayesian posterior distribution and adaptive threshold mechanism, and generate structured security early warning information containing risk level, location information and timestamp. The trajectory planning module is used to input structured safety warning information into the reinforcement learning agent, generate a preliminary trajectory through the near-end policy optimization algorithm, apply the expectation-maximization algorithm to iteratively optimize the path planner parameters and evaluation model parameters, and output a safe driving trajectory. The control execution module is used to input the safe driving trajectory into the deep learning controller, generate control commands through the feedforward network, and adjust the controller parameters through an online learning mechanism based on real-time control effect feedback to control the operation of the unmanned loader.

Images