A tire load identification method using LSTM-assisted extended Kalman filter

Abstract

This invention discloses a tire load state recognition method based on LSTM-assisted extended Kalman filter, comprising the following steps: S1. Establishing a 7-DOF vertical dynamic model of the vehicle; S2. Preparing observation data and state vector data as a training dataset; S3. Training the neural network based on the training dataset; S4. Acquiring observation data through sensors mounted on the vehicle body and axle; S5. Inputting the observation data and predicting the wheel vertical force through a neural network-assisted extended Kalman filter, wherein the Kalman gain coefficient is determined by LSTM. This method integrates extended Kalman filtering and LSTM networks, using the LSTM network to predict the Kalman gain coefficient, thereby reducing the need for an accurate dynamic model while improving the prediction accuracy of extended Kalman filtering.

Description

Technical Field

[0001] This invention relates to a tire load identification method using LSTM-assisted extended Kalman filtering, belonging to the field of vehicle signal processing. Background Technology

[0002] Tire load refers to the vertical force applied to the wheels of a vehicle, generated by the interaction between the vehicle and the road. Its magnitude and distribution are affected by factors such as the vehicle's total weight, axle layout, suspension design, and driving conditions. Tire load identification is of great significance for vehicle driving condition detection and structural design. Vertical force can most intuitively show the degree of rollover risk, and the suspension system also needs to be designed according to tire load to ensure vehicle stability and comfort.

[0003] The six-component force meter method is a current direct method for measuring tire load; however, it requires custom manufacturing based on the wheel hub, resulting in high cost and complex installation. Besides this, common methods include physical model-based monitoring methods and data-driven methods. The former attempts to construct a physical model to express the mapping relationship between observed and output quantities. This type of method requires a highly adaptable model and suffers from limitations such as low accuracy, high complexity, and high uncertainty. The latter typically uses neural networks to directly find the mathematical relationship between observed and output quantities, avoiding the exploration of complex physical structures. This type of method lacks mechanistic understanding and interpretability. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a tire load identification method based on LSTM-assisted extended Kalman filtering. Building upon the calculation of tire load using extended Kalman filtering, this method uses LSTM to predict the gain coefficients of the extended Kalman filter. While retaining the mechanistic aspects of physical modeling methods, it reduces the need for precise physical model adaptation.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] A tire load identification method using LSTM-assisted extended Kalman filtering includes the following steps:

[0007] S1. Establish a 7-DOF vertical dynamics model for the vehicle.

[0008] S2. Use the dynamic model to obtain the system observation y and state vector x data as the training dataset.

[0009] S3. Construct a neural network consisting of a three-layer LSTM network and train the neural network using the training set as input.

[0010] Step S3 includes the following sub-steps:

[0011] S3-1. Calculate the state vector estimate x t|t-1 Predicted value x t|t And calculate the Jacobian matrix of the predicted values ​​as The Jacobian matrix of the estimated value is used as

[0012] S3-2. Calculation As input to the neural network. The loss function uses L2 regularized mean squared error, based on x. t Calculate the loss.

[0013] S3-3. Input the three-layer LSTM network separately. Δy t and The covariance matrix Q of the predicted state noise and the covariance matrix of the predicted state vector are tracked respectively. Covariance matrix of predicted observations Output Kalman gain

[0014] S3-4. After training, the external recursive prediction model based on the extended Kalman filter outputs the predicted state vector.

[0015] S4. Acquire observations through sensors mounted on the vehicle body and axle axles.

[0016] S5. Input observation data is used to predict the vertical force of the wheel through a neural network-assisted extended Kalman filter prediction model, where the Kalman gain coefficient is determined by the neural network.

[0017] This invention has the following characteristics and beneficial effects:

[0018] By adopting the above technical solution, the present invention provides a tire load identification method based on LSTM-assisted extended Kalman filtering. It integrates extended Kalman filtering and LSTM network, and uses LSTM network to predict Kalman gain coefficient, thereby reducing the need for an accurate dynamic model while improving the prediction accuracy of extended Kalman filtering. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1This is a schematic diagram of the network structure of a tire load identification method using LSTM-assisted extended Kalman filtering according to an embodiment of the present invention.

[0021] Figure 2 This is a flowchart of a tire load identification method using LSTM-assisted extended Kalman filtering, according to an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the dynamic model of a tire load identification method using LSTM-assisted extended Kalman filtering according to an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0025] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0026] This invention provides a tire load identification method based on LSTM-assisted extended Kalman filtering, such as... Figure 2 As shown, it includes the following steps:

[0027] S1. Establish a 7-DOF vertical dynamics model for the vehicle.

[0028] like Figure 3 As shown, the state vector includes the vertical displacement z of the vehicle body. c ,speed acceleration Car body roll angle α c angular velocity angular acceleration Vehicle body pitch angle β c angular velocity angular acceleration Vertical displacement of the i-th axle bridge speed acceleration The i-th axle roll angle angular velocity angular acceleration Vertical force of the wheel to be identified

[0029] Then define the system's state vector. Then the Kalman filter state vector The system input vector is u = [g], and the system observations are... The extended Kalman filter prediction model is as follows:

[0030]

[0031] In the formula, Q and R are covariance matrices, e t v t There are two Gaussian white noise signals, f(·) and h(·) are two nonlinear functions, m is the dimension of x, which is 18, and n is the dimension of y, which is 7.

[0032] Understandable, y t The predicted observation is a priori estimated observation, while the actual observation is directly acquired. The role of Kalman gain is to adjust the prior estimate of the state vector based on the difference between the predicted and actual observations, thus obtaining the posterior estimate of the state vector.

[0033] S2. The system observation y and Kalman filter state vector x can be directly obtained by applying the dynamic model and used as the training dataset with a sampling frequency of 100Hz.

[0034] S3. Construct a neural network consisting of a three-layer LSTM network and train it based on the training set data.

[0035] Specifically, such as Figure 1 As shown, it includes the following steps:

[0036] S3-1. Calculate the state vector estimate x t|t-1 Predicted value x t|t And calculate the Jacobian matrix of the predicted values ​​as The Jacobian matrix of the estimated value is used as

[0037] Specifically, state vector estimate State vector prediction value Predicted Observations And calculate the Jacobian matrix of the predicted state vector values. Jacobian matrix of the state vector estimate

[0038] S3-2. Using the training set as input, apply the extended Kalman filter prediction model to calculate the observational differences. Observational residuals State vector prediction difference State increment And it is used as input to the neural network. The loss function uses L2 regularized mean squared error, based on x. t Calculate the loss.

[0039] S3-3. Inputting into three layers of LSTM respectively. Δy t and The covariance matrix Q of the predicted state noise and the covariance matrix of the predicted state vector are tracked respectively. Covariance matrix of predicted observations Output Kalman gain

[0040] S3-4. After training is complete, the external recursion outputs the predicted state vector based on the extended Kalman filter process.

[0041] Specifically, the neural network model refers to Figure 1 It consists of three LSTM layers and several fully connected layers. The first LSTM layer receives... Feature input, for Q t Perform tracking; the second LSTM layer receives the Q signal passed from the previous layer. t as well as Input, for Perform tracing; the third layer output receives the second layer output. and Δy t , Feature input, for Tracking will be conducted; finally... and Input to fully connected layer to calculate Kalman gain Feedback is relayed from the fully connected layer to the second layer for participation. The calculation.

[0042] Understandable, Kalman filter state vector Includes the vertical force of the wheel to be identified Therefore, the output predicted state vector It includes the predicted vertical force of the wheel.

[0043] S4. Acquire the actual observation y′ through sensors mounted on the vehicle body and axle, and sample y′ at 100Hz frequency intervals.

[0044] S5. Continuously input real observation data y′, and predict x′ using a neural network-assisted extended Kalman filter. The x′ vector contains the vertical force of the wheel. The Kalman gain coefficient of the filter is determined by a neural network.

[0045] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments, including components, without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A tire load identification method of LSTM-aided extended Kalman filtering, characterized in that, The method comprises the following steps: S1, establishing a dynamics model of a vehicle based on Simulink; S2, obtaining system observation y and state vector x data by applying the dynamics model, and constructing a training set, wherein the system observation y comprises vertical acceleration of a vehicle body, side roll angle acceleration of the vehicle body, nodding angle acceleration of the vehicle body, vertical acceleration of an ith axle bridge, side roll angle acceleration of the ith axle bridge; the state vector x comprises vertical displacement, speed and acceleration of the vehicle body, side roll angle, angular speed and angular acceleration of the vehicle body, nodding angle, angular speed and angular acceleration of the vehicle body, vertical displacement, speed and acceleration of the ith axle bridge, side roll angle, angular speed and angular acceleration of the ith axle bridge, and vertical force of a wheel to be identified, wherein i = 1, 2, 3, 4; First define the state vector of the dynamic model Then the Kalman filter state vector The system input vector is u = [g], and the system observation is Where z c is the vertical displacement of the vehicle body, is the vertical velocity of the vehicle body, is the vertical acceleration of the vehicle body, a c is the roll angle of the vehicle body, is the roll angular velocity of the vehicle body, is the roll angular acceleration of the vehicle body; b c is the pitch angle of the vehicle body, is the pitch angular velocity of the vehicle body, is the pitch angular acceleration of the vehicle body; is the vertical displacement of the ith axle, is the vertical velocity of the ith axle, is the vertical acceleration of the ith axle; is the roll angle of the ith axle, is the roll angular velocity of the ith axle, is the roll angular acceleration of the ith axle; is the vertical force of the wheel to be identified; S3, constructing a neural network composed of three layers of LSTM networks, and training the neural network by taking the training set as input; The training method of the neural network is as follows: S3-1. Apply the extended Kalman filter prediction model to calculate the state vector estimation value with the training set as input state vector prediction value predicted observation and calculate the Jacobian matrix of the state vector prediction value and the Jacobian matrix of the state vector estimate S3-2. Apply the extended Kalman filter prediction model to calculate the observation difference with the training set as input Observation residual State vector prediction difference State increment and as input to the neural network; S3-3. The three-layer LSTM network is respectively input Δy t and Covariance matrix Q of the state noise respectively tracked and predicted, covariance matrix Covariance matrix of the predicted observation And output Kalman gain S3-4. After the training, the external recursion outputs the predicted state vector based on the extended Kalman filter prediction model S4. Real observations are obtained by sensors mounted on the real vehicle body and axle bridge S5, inputting the observation data through the neural network to output a Kalman gain assisted extended Kalman filter prediction model to predict the vertical force of the wheel.

2. The tire load identification method of claim 1, wherein, In the step S3-2, the loss function uses L2 regularized mean squared error, based on x t Compute loss.

3. The tire load identification method of claim 1, wherein, The predicted state vector comprises a wheel vertical force.

4. The tire load identification method of claim 1, wherein, The extended Kalman filter prediction model is as follows: where Q, R are covariance matrices, e t , v t are two Gaussian white noise signals, f(·), h(·) are two nonlinear functions, m is the dimension of x, equal to 18, n is the dimension of y, equal to 7.

5. The tire load identification method of claim 1, wherein: The data in the step S2 is from a model state vector prepared in advance, and constitutes a data-driven method for realizing future load state prediction based on historical sample data.

6. The wheel load state recognition method based on LSTM auxiliary extended Kalman filtering according to claim 1, characterized in that: In the step S5, in each time's extended Kalman filtering process, the Kalman gain is determined by the neural network according to the input Δy t , .

Images