Wheeled mobile robot slip rate prediction method based on deep learning algorithm

Abstract

The invention belongs to the field of wheeled mobile robot state estimation, discloses a wheeled mobile robot slip rate prediction method based on deep learning, and solves the problem of insufficient prediction precision caused by incapability of prediction in advance in real-time measurement of a sensor and wheel-ground interaction complexity. The specific method comprises the following steps: firstly, constructing a wheel slip rate data set of related motion data of the wheeled mobile robot; secondly, optimizing and improving the number of layers and the structure of the CNN-LSTM network, and efficiently extracting information and capturing features; besides, the designed hidden state enhancement module enhances the updating of the hidden state, and the designed attention mechanism improvement module dynamically adjusts the input feature weight and highlights the most important feature. And finally, designing a composite loss function considering mechanical constraints to train the model, and performing verification and performance evaluation on the trained model. According to the invention, the wheeled mobile robot has good adaptability on different terrains, and the accuracy and stability of slip rate prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot application, and in particular relates to a slip rate prediction method for a wheeled mobile robot based on a deep learning algorithm. Background Art

[0002] Wheeled mobile robots are ideal for many fields, such as forestry, agriculture, and industry, due to their high reliability, simplicity of design, and inherent balance. With the advancement of science and technology, wheeled robots are deployed in more challenging environments to reduce risks or replace humans to complete specific tasks. However, navigation and control of wheeled mobile robots in outdoor environments and rugged terrains still face great challenges. When dealing with geometric hazards in various environments, special attention should be paid to the potential risk of wheel slippage in outdoor environments, which will significantly affect the maneuverability of wheeled mobile robots in off-road conditions. If this problem is not solved, wheeled robots may get stuck or lose control in soft or rugged terrain. Therefore, there is an urgent need for a reliable method to predict the slip rate of the wheel in advance, compensate for the slippage by controlling the speed or torque, thereby avoiding the negative impact caused by the terrain and achieving more precise control of the wheeled mobile robot.

[0003] From the perspective of prediction methods, they can be roughly divided into three categories: proprioceptive sensor-based, exteroceptive sensor-based, and model-based prediction methods. In the context of autonomous vehicles, proprioceptive sensors are on-board sensors such as inertial measurement units (IMUs), global positioning systems (GPS), and wheel encoders that measure vehicle status. The combination of inertial measurement units (IMUs) and wheel encoders will provide enough information to represent the behavior of wheeled robots for slip prediction. Although this method can directly measure the physical interaction between the wheels and the terrain surface, it cannot predict the slip in advance. On the other hand, exteroceptive sensor-based methods can predict the potential slip of wheeled robots at target locations in the future in real time. However, existing exteroceptive sensor-based methods usually rely on visual sensors based only on the characteristics of the terrain surface, which may not be sufficient for accurate slip prediction. Wheel-ground interaction is inherently nonlinear and highly complex, so deep learning algorithms are potentially more suitable solutions. In addition, wheeled mobile robots are dynamic systems, and the slip of the robot depends on the current and past states of the robot, and the estimation problem of time series data also needs to be considered. Therefore, combining the above factors, the present invention proposes an improved CNN-LSTM network model. The core of the model is to design and add a hidden state enhancement module and an improved attention mechanism module, as well as to design and add a composite loss function with mechanical constraints during model training. This model can more efficiently and accurately extract spatiotemporal feature information and capture time series features, fully consider the impact of different input variables on prediction, accelerate the convergence speed of model errors, and improve the generalization ability and robustness of the model. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention provides a slip rate prediction method for a wheeled mobile robot based on a deep learning algorithm, which mainly includes the following steps:

[0005] Step 1: Construct a wheel slip rate dataset for a wheeled mobile robot, which contains the driving data of the mobile robot under four different terrains (ice, road, mud, and sand). The input variables of the dataset are: terrain category, motor torque value T, linear velocity v in the x direction (forward) and z direction (longitudinal), angular velocity ω, and slope value θ, and the output variable is slip rate s;

[0006] Step 2: An improved CNN-LSTM network model is proposed. The core is to design and add a hidden state enhancement module and an improved attention mechanism module. The improved attention mechanism module is a dynamic weight channel attention mechanism (Dynamic Weighting SENet, DW-SENet). The model predicts the slip rate of a wheeled mobile robot by extracting spatiotemporal feature information and capturing time series features. Feature information is efficiently extracted by optimizing the combination and connection structure of convolutional layers and pooling layers. The consistency of input and output dimensions is maintained in each layer to ensure that the model fully utilizes input features at different levels, avoids information loss due to dimensional changes, and improves the efficiency of feature utilization. The hidden state enhancement module designed and added learns the information in the hidden state by adding a linear layer to further optimize the update of the hidden state. The dynamic weight channel attention mechanism module (DW-SENet) designed and added generates a dynamic weight matrix for each channel by adding a convolutional layer on the basis of SENet, fully considering the impact of different input variables on the prediction to adjust and allocate weights;

[0007] Step 3: Design a composite loss function that takes mechanical constraints into account. By controlling the weight of the mechanical constraints in the composite loss function, the error convergence speed during model training is faster, fewer rounds of training are required, and the generalization ability and robustness of the model are enhanced. According to the data set produced in step 1, a training set with a ratio of 70% is used to train the model, and a verification set with a ratio of 30% is input into the network model described after training for verification. After the verification is completed, its performance indicators are evaluated;

[0008] Beneficial effects of the present invention: The method for predicting the slip rate of a wheeled mobile robot based on a deep learning algorithm proposes an improved CNN-LSTM network model. By designing and adding a hidden state enhancement module, the information in the hidden state is learned, and the capture of dependency relationships can be effectively enhanced. This improvement optimizes the update mechanism of the hidden state, which can improve the accuracy and stability of the prediction and enhance the sensitivity to the change of slip rate on different terrains. By designing and adding an improved attention mechanism module, a dynamic weight matrix is ​​generated for each channel, and the influence of different input variables on the prediction is fully considered to adjust and allocate weights, highlight the most important features for slip rate prediction, and suppress secondary or noise features. In this way, the feature representation ability of the network is enhanced. By designing and adding a composite loss function that considers mechanical constraints, the error convergence speed during model training is faster, the number of training rounds required is reduced, and the generalization ability and robustness of the model are enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 Flow chart of the technical route implemented by the present invention

[0010] Figure 2 Schematic diagram of the improved CNN-LSTM network model structure

[0011] Figure 3 Schematic diagram of the improved attention mechanism module structure

[0012] Figure 4 Schematic diagram of the hidden state enhancement module structure designed DETAILED DESCRIPTION

[0013] The following is a further detailed description of the embodiments of the present invention in conjunction with the accompanying drawings and examples. The embodiment of the present invention provides a method for predicting the slip rate of a wheeled mobile robot based on a deep learning algorithm, and the content shown includes the following steps:

[0014] Step 1: Construct a wheel slip rate dataset for a wheeled mobile robot, which contains the driving data of the mobile robot under four different terrains (ice, road, mud, and sand). The input variables of the dataset are: terrain category, motor torque value T, linear velocity v in the x direction (forward) and z direction (longitudinal), angular velocity ω, and slope value θ, and the output variable is slip rate.

[0015] The calculation method of the parameters contained in the data set is as follows: Slope angle calculation, using the difference between the vertical position of the front wheel (Fz) and the vertical position of the rear wheel (Rz) and the length between the front and rear wheels (L) to calculate the slope value at each moment, the specific formula is: The specific formula for calculating the longitudinal slip rate is: The slip value is between 0 and 1, where s represents the slip value, r is the wheel radius, v is the actual speed, and ω is the angular velocity. The calculated longitudinal slip rate is used as the true value of the data set.

[0016] Step 2: An improved CNN-LSTM network model is proposed. The core is to design and add a hidden state enhancement module and an improved attention mechanism module. The improved attention mechanism module is a dynamic weighting channel attention mechanism (Dynamic Weighting SENet, DW-SENet). The model predicts the slip rate of the wheeled mobile robot by extracting spatiotemporal feature information and capturing time series features.

[0017] The improved CNN-LSTM network model efficiently extracts feature information by optimizing the combination and connection structure of convolutional layers and pooling layers. The consistency of input and output dimensions is maintained in each layer, ensuring that the model fully utilizes input features at different levels, avoiding information loss due to dimensional changes, and improving feature utilization efficiency.

[0018] Extract spatiotemporal feature information, the feature information is: F d (i) = {f i1 ,f i 2 ,f i 3 ,f i 4 ,f i 5}, F g (i) = {f i 6}, i = 1, 2, 3...N, where Fd represents the five features extracted from the motion data, Fg represents the terrain category, and N is the total number of instances in the dataset.

[0019] The input data is fed into a convolutional layer with 4 layers of 3×1 kernels. After stacking the convolutions layer by layer, a deep feature map representation is formed. After each convolution operation, the ReLU activation function is applied in the activation layer to introduce nonlinear features, normalize and filter out negative values ​​to enhance the network's expressiveness.

[0020] The feature information after convolution and activation operations is input into 4 pooling layers with a pooling window size and a step size of 2. The maximum value of every 2 elements is taken to halve the length of the feature map layer by layer, reducing the amount of data and retaining the main features.

[0021] To capture time series features, a three-layer module stacking structure is used. The input dimension of the first layer is directly taken from the number of output features of the feature extraction part, ensuring that the convolutional features can be fully utilized. The input dimension of each layer depends on the hidden state dimension of the previous layer, and features can be extracted and abstracted layer by layer. In the forward propagation of each layer, the dropout parameter is set to 0.3 to prevent overfitting of the model.

[0022] The hidden state enhancement module designed and added further optimizes the update of the hidden state by adding a linear layer to learn the information in the hidden state.

[0023] The hidden state enhancement module is deployed in the improved LSTM network. The specific steps are: adding a linear layer to extract enhanced information on the hidden state, and controlling the inflow of enhanced information through a sigmoid function. The expression of the sigmoid function is: The specific formula design of the hidden state enhancement module is: output t =sigmoid(W m ·h t +b m ) Among them, W m is the weight matrix, h t is the hidden state of the output, b mis the bias term, output t is the additional information extracted from the hidden state.

[0024] The extracted additional information is combined with the output hidden state h of the model t Combined to form a reinforced hidden state h t ', the specific formula is designed as: h t '=output t +h t This enhanced hidden state can better preserve and strengthen important memory information for use in subsequent time steps.

[0025] The dynamic weight channel attention mechanism module (DW-SENet) designed and added generates a dynamic weight matrix for each channel by adding a convolutional layer on the basis of SENet, fully considering the impact of different input variables on the prediction to adjust and allocate weights;

[0026] The dynamic weight channel attention mechanism module (DW-SENet) has the following specific steps: design a lightweight 1×1 convolutional layer to process the input features of each channel, thereby generating the channel weight. The specific formula is expressed as: W c =σ(Conv 1×1 (x c )) Among them, Conv 1×1 is a lightweight 1×1 convolutional layer, σ is the sigmoid function, x c are the input features of different channels, W c is the dynamic weight matrix.

[0027] Multiply the generated dynamic weight matrix and the input feature map, and then sum them up by time step to get the dynamic feature description of each channel. The specific formula is: in, is a dynamic feature description, W(x c )(t) is the dynamic weight of channel c at the tth time step, x c (t) is the feature of that time step.

[0028] Global average pooling (GAP) is used to generate global feature descriptions for each channel. The specific formula is expressed as: The dynamic feature description is combined with the global feature description through a balance parameter α to generate a global description vector for each channel. The specific formula is designed as follows: Among them, α is a learnable parameter, and the initialization value of α is set to 0.4. During the training process, the model will automatically adjust the value of α according to the loss function and gradient descent algorithm to adapt to different input features.

[0029] The output of the weighted pooling is passed to the subsequent fully connected layer to calculate the final weight S for each channel c , the specific formula is expressed as: Among them, W1 and W2 are the weight matrices of the two fully connected layers.

[0030] The final weight of each channel generated is multiplied by the convolutional feature map and then used as the input feature information of the improved LSTM network.

[0031] Step 3: Design a composite loss function that takes mechanical constraints into account. By controlling the weight of the mechanical constraints in the composite loss function, the error converges faster during model training, fewer rounds of training are required, and the generalization ability and robustness of the model are enhanced.

[0032] Design a composite loss function that takes mechanical constraints into account. The specific steps are as follows: the mechanical constraint equation is designed as: Where N is the number of samples, T is the torque value, R is the wheel radius, μ max is the maximum friction coefficient, s critical is the critical slip rate, s i is the true value of the slip rate, m is the mass of the wheeled mobile robot, and θ is the slope angle. max The specific setting is 0.2, s critical The specific setting is 0.9.

[0033] The specific expression of the standard MSE loss function is: in, is the predicted value of the slip rate. Considering the weight of the mechanical constraint in the composite loss function, the above mechanical constraint equation is added with the MSE loss function, and the composite loss function can be obtained as follows: L=L m +λL p , Among them, L is the total loss function, L m is the MSE loss function, λ is the weight of the control mechanical constraint in the composite loss function, the initial value of λ is set to 0.1, and a more appropriate value is gradually adjusted and learned during the training process, and a regularization constraint is imposed to make it change within a reasonable range. p is the loss function considering mechanical constraints.

[0034] Therefore, the specific expression of the composite loss function considering mechanical constraints is:

[0035] According to the data set in step 1, 70% of the data is divided into a training set to train the model, and 30% of the data is divided into a validation set and input into the network model after training for verification. After the verification is completed, its performance indicators are evaluated.

[0036] The optimizer is Adam. By iterating the training data, the model parameters are continuously adjusted. In each iteration, the output predicted by the model is compared with the actual label, the loss function value is calculated, and then the model weight is updated to reduce the loss. The training round epoch is set to 100 and the learning rate is 0.0003. The final version of the model and the version with the best performance on the test set can be saved.

[0037] Set the random seed to make the experimental results repeatable and define the mean square error Root mean square error Mean absolute error These three errors are used as indicators to evaluate the performance of the model.

Claims

1. A method for predicting slip rate of a wheeled mobile robot based on a deep learning algorithm, characterized in that: The following steps are involved: Step 1: Construct a wheel slip rate dataset for a wheeled mobile robot, which contains the driving data of the mobile robot under four different terrains (ice, road, mud, and sand). The input variables of the dataset are: terrain category, motor torque value T, linear velocity v in the x direction (forward) and z direction (longitudinal), angular velocity ω, and slope value θ, and the output variable is slip rate s; Step 2: An improved CNN-LSTM network model is proposed. The core is to design and add a hidden state enhancement module and an improved attention mechanism module. The improved attention mechanism module is the dynamic weight channel attention mechanism (DynamicWeighting SENet, DW-SENet). The model predicts the slip rate of the wheeled mobile robot by extracting spatiotemporal feature information and capturing time series features. By optimizing the combination and connection structure of the convolution layer and the pooling layer, feature information can be efficiently extracted. The consistency of the input and output dimensions is maintained in each layer to ensure that the model fully utilizes the input features at different levels, avoids information loss due to dimensional changes, and improves the efficiency of feature utilization. The hidden state enhancement module designed and added learns the information in the hidden state by adding a linear layer to further optimize the update of the hidden state. The dynamic weight channel attention mechanism module (DW-SENet) designed and added generates a dynamic weight matrix for each channel by adding a convolution layer on the basis of SENet, fully considering the impact of different input variables on the prediction to adjust and allocate weights; Step 3: Design a composite loss function that takes mechanical constraints into account. By controlling the weight of the mechanical constraints in the composite loss function, the error convergence speed during model training is faster, fewer rounds of training are required, and the generalization ability and robustness of the model are enhanced. According to the data set in step 1, a training set with a ratio of 70% is used to train the model, and a validation set with a ratio of 30% is input into the network model described after training for verification. After the verification is completed, its performance indicators are evaluated.

2. The method for predicting slip rate of a wheeled mobile robot based on a deep learning algorithm according to claim 1, characterized in that: The calculation method of the parameters contained in the data set in step 1 is as follows: Slope angle calculation, using the difference between the vertical position of the front wheel (Fz) and the vertical position of the rear wheel (Rz) and the length between the front and rear wheels (L) to calculate the slope value at each moment, the specific formula is: The specific formula for calculating the longitudinal slip rate is: The slip value is between 0 and 1, where s represents the slip value, r is the wheel radius, v is the actual speed, and ω is the angular velocity. The calculated longitudinal slip rate is used as the true value of the data set.

3. The method for predicting slip rate of a wheeled mobile robot based on a deep learning algorithm according to claim 1, characterized in that: The specific steps in step 2 are: Step 3.1, extract spatiotemporal feature information, the feature information is: F d (i) = {f i 1 ,f i 2 ,f i 3 ,f i 4 ,f i 5 }, F g (i) = {f i 6 }, i = 1, 2, 3...N, where Fd represents the five features extracted from the motion data, Fg represents the terrain category, and N is the total number of instances in the dataset. The feature information is input into a convolutional layer containing 4 layers of 3×1 kernels. After the convolutions are stacked layer by layer, a deep feature representation is formed. After each convolution operation, the ReLU activation function is applied in the activation layer to introduce nonlinear features, normalize and filter out negative values ​​to enhance the network's expressive power. The feature information after convolution and activation operations is input into a 4-layer pooling layer with a pooling window size and a step size of 2. The maximum value of every 2 elements is taken to halve the length of the feature map, reduce the amount of data and retain the main features; Step 3.2, to capture time series features, a three-layer module stacking structure is used. The input dimension of the first layer is directly taken from the number of output features of the feature extraction part, ensuring that the convolutional features can be fully utilized. The input dimension of each layer depends on the hidden state dimension of the previous layer, and features can be extracted and abstracted layer by layer. In the forward propagation of each layer, the dropout parameter is set to prevent overfitting of the model. Step 3.3, according to the method for predicting the slip rate of a wheeled mobile robot based on a deep learning algorithm according to claim 1, it is characterized in that the hidden state enhancement module in step 2 has the following specific steps: adding a linear layer on the hidden state to enhance the extraction of information, and controlling the inflow of enhanced information through a sigmoid function, and the specific formula is designed as follows: output t =sigmoid(W m ·h t +b m ) Among them, W m is the weight matrix, h t is the hidden state of the output, b m is the bias term, output t is the additional information extracted from the hidden state. The extracted additional information is combined with the output hidden state h of the model t Combined to form a reinforced hidden state h t ', this enhanced hidden state can better preserve and strengthen important memory information for use in subsequent time steps. h t '=output t +h t Step 3.4, according to a method for predicting slip rate of a wheeled mobile robot based on a deep learning algorithm as described in claim 1, characterized in that the improved attention mechanism module in step 2 has the following specific steps: design a lightweight 1×1 convolutional layer to generate a dynamic weight matrix for each channel, multiply the generated dynamic weight matrix by the input feature map, and then sum it by time step to obtain the dynamic feature description of each channel. The dynamic feature description is combined with the global feature description through a balance parameter α to generate a global description vector for each channel, and the specific formula is expressed as: Among them, Z c dynamic is a dynamic feature description, It is a global feature description. Finally, the output of weighted pooling is passed to the subsequent fully connected layer to calculate the weight of each channel.

4. The method for predicting slip rate of a wheeled mobile robot based on a deep learning algorithm according to claim 1, characterized in that: In step 3, a composite loss function considering mechanical constraints is designed, and the specific steps are as follows: the mechanical constraint equation is designed as: Where N is the number of samples, T is the torque value, R is the wheel radius, μ max is the maximum friction coefficient, s critical is the critical slip rate, m is the mass of the wheeled mobile robot, and θ is the slope angle. By adding the above mechanical constraint equation to the MSE loss function, we can get the complete composite loss function L = L m +λL p , its specific expression is: Among them, s i is the true value of the slip rate, is the predicted value of the slip rate, and λ is the weight of the control mechanical constraint in the composite loss function. The optimizer is Adam. The model parameters are adjusted by iterating the training data. In each iteration, the output predicted by the model is compared with the actual label, the loss function value is calculated, and then the model weight is updated to reduce the loss. The training round epoch is set to 100 and the learning rate is 0.0003. The final version of the model and the version with the best performance on the test set can be saved. The random seed is set to make the experimental results reproducible and the mean square error is defined. Root mean square error Mean absolute error These three errors are used as indicators to evaluate the performance of the model.

Images