A method for building an attention mechanism-based LSTM neural network vehicle suspension system state observer

Abstract

The application relates to a kind of establishment methods of vehicle suspension system state observer based on attention mechanism LSTM neural network, to observe unknown state information, belong to the field of automobile chassis suspension. The specific steps are as follows: a half-car suspension system dynamics model is built on a simulation software, and the suspension system state variables and road information are obtained by simulation; data preprocessing; design an encoder-decoder neural network model based on attention mechanism; randomly initialize parameters; select hyperparameters using Bayesian optimization; optimize the state observer model using mini-batch stochastic gradient descent and momentum optimization; use dropout regularization and early stopping methods to optimize the model to prevent overfitting of the state observer model; suspension state observer model training and saving; model accuracy verification. Based on the observer, the observation problem of unknown state variables, road information, displacement input spectrum and force input spectrum of the nonlinear half-car suspension system is solved, with higher precision and faster calculation speed advantage.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automobile chassis suspension, and creates a vehicle suspension system and a road information observer based on an LSTM neural network with an attention mechanism, observes unknown state information and road information, improves the control accuracy of suspension dynamics, and facilitates the evaluation of vehicle ride comfort. BACKGROUND

[0002] The automobile suspension system is an important part of the half-car system, which not only determines the ride comfort of the vehicle, but also affects the steering stability of the vehicle. The research on suspension system control has always been a research hotspot for scholars at home and abroad. In order to effectively control the automobile suspension, the state variables of the suspension system and the road information must be known. Although sensors applied to vehicles can measure part of the suspension system state variables, for the road state information and non-sprung mass displacement and other difficult-to-measure state variables, currently only state observers can be created for estimation.

[0003] The traditional method of building a state observer is mostly based on a model, and the Kalman filter algorithm is the most representative method, but its working principle has some inherent shortcomings. First, the previous research mostly adopts a linear quarter-suspension model, ignoring the strong nonlinearity caused by the internal friction of the suspension system structure and the viscoelasticity of the rubber material, and the model is not accurate enough. Secondly, the pitch and roll movements of the vehicle are ignored. In addition, the parameters identified by the Kalman filter under different driving conditions are different, which affects the observation accuracy. Some scholars have proposed an adaptive Kalman filter algorithm to address the above shortcomings, but the algorithm has a large amount of calculation, which limits its implementation in real vehicles.

[0004] Road spectrum collection is necessary in vehicle dynamics control and vehicle chassis component fatigue life analysis. The currently widely used road spectrum collection method is sensor measurement, but due to the limited accuracy of sensors and the influence of test conditions, a high-precision and easy-to-implement road spectrum collection method has always been the research direction of scholars at home and abroad. This paper proposes a vehicle suspension system state observer that can effectively and reliably collect road spectrum, including road displacement input spectrum and road force input spectrum.

[0005] Tire dynamic load is the main reason for the damage of tires to the road and one of the factors affecting the safety and ride comfort of vehicles. In vehicle vertical dynamics control, tire dynamic load is always the first consideration. However, tire dynamic load is difficult to measure directly by sensors and can only be predicted by an observer. The state observer proposed in this paper can observe tire dynamic load in real time with high precision, which is beneficial to the design of vehicle dynamics control algorithm.

[0006] The LSTM neural network based on the attention mechanism is widely used in natural language processing and high-precision image recognition. The training object is a complex corresponding relationship between sequences. The key feature information of the input sequence is extracted through the encoder network. The attention mechanism can filter out the information strongly related to the target output. The feature information is converted into the output through the decoder network. Through continuous iteration and optimization of data, a deep corresponding relationship between the input sequence and the output sequence is established. For the vehicle state observer, the input sequence is the time sequence of the measurable suspension system state variables, and the output is the time sequence of the unknown suspension system state variables and the road information. The LSTM neural network based on the attention mechanism is applied to the observation of the suspension system state variables, a state observer is established, and parameter optimization is performed, which is described as follows: SUMMARY

[0007] The purpose of the present application is to use the LSTM neural network based on the attention mechanism to build a vehicle suspension system state variable and road information observer, especially the LSTM vehicle suspension system state and road information observer based on the attention mechanism, to estimate unknown states and road information in real time using known state variables. According to the state variable information collected by simulation, the training set and the test set are divided, the long short-term memory neural network model is trained, and a high-precision, fast-computing vehicle suspension system state observer based on data is obtained.

[0008] The encoder-decoder recurrent neural network is used to train the deep relationship between the input sequence and the output sequence, and to improve the accuracy of the state observer.

[0009] The attention mechanism is used to record the similarity between the output of the encoder and the previous hidden state of the decoder, and to obtain the maximum weight of each encoder output. The neural network model is optimized to solve the memory failure problem of long time sequence and to speed up the learning efficiency.

[0010] Dropout regularization and early stopping are used to optimize the model to solve the problem of model overfitting.

[0011] Bayesian optimization is used to select hyperparameters, and Glorot is used to initialize hyperparameters. The optimal model is obtained while greatly reducing the learning time.

[0012] Momentum optimization and small batch stochastic gradient descent method are used to train and optimize the model, which speeds up the convergence speed and effect of the model.

[0013] The suspension system state observer established by the LSTM neural network model based on the test simulation data has the advantages of high observation accuracy and fast calculation rate compared with the traditional state observer.

[0014] The application provides a building method of an attention mechanism-based LSTM neural network vehicle suspension system state observer, and comprises the following steps:

[0015] Step 1: a half-vehicle suspension system dynamics model is built on a simulation software, and suspension system state variables and road surface information are obtained through simulation;

[0016] Step 2 comprises the following steps: 2.1, data preprocessing: 2.1.1, the state variable data is divided into input and output, the state variable measured by a sensor is input, and the state variable and road surface information difficult to measure are output; 2.1.2, the state variable is converted into a time sequence with a fixed time step, and the data is normalized; 2.1.3, all data is divided into a training set, a verification set and a test set by random sampling, and the proportion is 10:2:2;

[0017] The data in the training set is used for training of the state observer model, the data in the verification set is used for cross-validation of the accuracy of the model during the training process, and the model is iteratively optimized; the data in the test set is used for final testing of the generalization ability of the model, and the estimation accuracy of the state observation model for unknown state variables is tested;

[0018] 2.2, designing a state observer model: an encoder-decoder neural network model is built, and an attention mechanism layer is built between the encoder layer and the decoder layer; the encoder layer processes the information in the input sequence, each neuron receives the input at the time step and the output at the previous time step, and the sequence information is memorized and retained; the attention mechanism layer combines the output of the encoder layer with the decoder output at the previous step, and evaluates the corresponding relationship, and can focus on learning the data closely related to the target output; the decoder layer takes the output of the attention mechanism layer as input, and outputs the state variable information corresponding to each time step;

[0019] Step 3: Glorot is used to initialize hyperparameters: the distribution range of the random initialization parameters is determined according to the number of inputs and outputs of each layer;

[0020] Step 4: Bayesian optimization is used to select hyperparameters: learning rate, number of neurons in each layer; first, a target function is determined, then a distribution domain space of each hyperparameter is determined, and an algorithm is designed to optimize the state observer model, and an optimal model hyperparameter combination is obtained;

[0021] Step 5: small batch random descent and momentum optimization are used to optimize the state observer model;

[0022] Step 6: dropout regularization and early stopping are used to optimize the model to prevent overfitting of the state observer model;

[0023] Step 7, model training and saving: the body pitch angle θ, the body acceleration Front suspension relative displacement Rear suspension relative displacement Front suspension unsprung mass acceleration Rear suspension unsprung mass acceleration The time series of state variables as input; the vehicle front suspension unsprung mass displacement Rear suspension unsprung mass displacement State variables and front wheel road displacement Rear wheel road displacement Front wheel tire dynamic load Rear wheel tire dynamic load Front wheel road force input F l f Rear wheel road force input The time series of the output. Import data for state observer model training, and save the observer model.

[0024] Step 8, model accuracy verification: compare the predicted unknown state variables with the output in the verification set, and obtain the model accuracy of 96.35%, which is more accurate than the traditional modeling method.

[0025] The technical effects of the present application are as follows:

[0026] The LSTM neural network vehicle suspension system state observer based on attention mechanism not only solves the observation problem of unknown state variables and road information of the nonlinear half-car suspension system, but also has higher accuracy and faster calculation speed than the traditional state observer. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 Half-car suspension model;

[0028] Figure 2 LSTM neural unit;

[0029] Figure 3 Encoder-decoder neural network structure diagram based on attention mechanism;

[0030] Figure 4 LeakyReLU activation function;

[0031] Figure 5 Suspension state observer model schematic diagram;

[0032] Figure 6 Suspension state observer model building flowchart;

[0033] Figure 7a Front wheel road information observation and actual comparison diagram;

[0034] Figure 7b Rear wheel road surface information observation and actual comparison chart;

[0035] Figure 7c Front suspension unsprung displacement observation and actual comparison chart;

[0036] Figure 7d Rear suspension unsprung displacement observation and actual comparison chart;

[0037] Figure 7e Front wheel tire dynamic load observation and actual comparison chart;

[0038] Figure 7f Rear wheel tire dynamic load observation and actual comparison chart;

[0039] Figure 7g Front road surface force input observation and actual comparison chart;

[0040] Figure 7h Rear road surface force input observation and actual comparison chart; DETAILED DESCRIPTION

[0041] 1) Half-car suspension system dynamics modeling and simulation:

[0042] Half-car suspension model reference Figure 1

[0043] The half-car suspension model used in this paper is an air spring suspension model with high modeling accuracy. The spring stiffness is obtained by polynomial fitting, and the damper damping force is obtained by neural network model learning, both of which are nonlinear models. The state observer can solve the problem of nonlinear calculation difficulty. However, the state observer described in this paper is not limited to the suspension model referred to in this paper.

[0044] The half-car suspension system dynamics equation can be expressed as:

[0045]

[0046]

[0047]

[0048]

[0049] In the formula:

[0050] m s is the vehicle body mass

[0051] I s is the vehicle body moment of inertia is the front and rear wheel non-sprung mass

[0052] x sBody displacement

[0053] θ Body pitch angle u f, u r Road displacement of front and rear wheels z f, z r Suspension displacement of front and rear wheels z f, z r Suspension displacement of front and rear wheels k f, k r Tire stiffness of front and rear wheels c f, c r Nonlinear air spring force of front and rear suspensions c f, c r Nonlinear air spring force of front and rear suspensions a Distance from body mass center to front axle

[0054] b Distance from body mass center to rear axle

[0055] wherein:

[0056]

[0057]

[0058] The fitted nonlinear air spring force is:

[0059]

[0060] The time-domain expression of filtered white noise is used as the road input model, and the front and rear wheel road input equations are respectively:

[0061]

[0062]

[0063] f0Lower cutoff frequency

[0064] G0Road roughness coefficient

[0065] u c Vehicle speed

[0066] w1, w2 Random white noise of front and rear road inputs

[0067] The half-car suspension model parameters are imported into the matlab software, and the simulink software is used to create a half-car suspension dynamics model for simulation. The simulation data of the suspension state variables and road displacement are recorded.

[0068] 2) Build an LSTM neural network state observer model based on attention mechanism:

[0069] LSTM neural unit reference Figure 2: LSTM units are the most popular of the long short-term memory units in neural networks. It works like a normal basic neural unit, but trains faster and detects long-term dependencies in data. It feeds the current input vector x(t) and the previous short-term state h(t-1) into four different fully connected layers. The main layer is the layer that outputs g(t), which stores the most important parts of the input in the long-term state and discards the rest. The other three layers are gate controllers, including: the forget gate, the input gate, and the output gate. The forget gate controls which parts of the long-term state should be deleted; the input gate controls which parts of g(t) should be added to the long-term state; and the output gate controls which parts of the long-term state should be read at this time step and merged into h(t) and the output y(t) for the next time step.

[0070] LSTM computation equation:

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077] In this equation:

[0078] W xi , W xf , W xo , W xg is the weight matrix W (t) , W hi , W hf , W ho , W hg is the weight matrix W (t-1) , W i , W f , W o , W g is the bias term b

[0079] 2.1 Data pre-processing

[0080] 2.1.1 Splitting sequential data into training, validation, and test sets: the first 80% of the data is used as the training set, the middle 10% is used as the validation set, and the last 10% is used as the test set.

[0081] 2.1.2 Normalize the data. Normalize the data to the same order of magnitude to prevent learning errors due to the difference in the size of the values.

[0082] Scaling function:

[0083] F' is the normalized data.

[0084] F text is the data before normalization.

[0085] F max is the maximum value among all input data.

[0086] F min is the maximum value among all input data.

[0087] 2.1.3 Cut the sequential data set into multiple time series: Determine the time step of each state vector, and then take the staggered data collection method to cut all data into sequences of the same time step.

[0088] 2.1.4 Randomly shuffle the state vector sequence: Shuffle the order to make the learned model more universal.

[0089] 2.1.5 Determine the input vector sequence and the output vector sequence: Take the state variables that are easy to measure with sensors as the training input of the model, including: vehicle body pitch angle θ, vehicle body acceleration Front suspension relative displacement Rear suspension relative displacement Front suspension unsprung mass acceleration Rear suspension unsprung mass acceleration Take the state variables and road information that are difficult to measure as the training output of the model, including: front suspension unsprung mass displacement Rear suspension unsprung mass displacement State variables and front wheel road displacement Rear wheel road displacement Front wheel tire dynamic load Rear wheel tire dynamic load Front wheel road force input F l f Rear wheel road force input

[0090] 2.2 Design an encoder-decoder network based on attention mechanism

[0091] The structure of the encoder-decoder neural network based on attention mechanism is shown in Figure 3 .

[0092] 2.2.1 Encoder part: The basic unit of the encoder is LSTM neural unit, which receives the state vector x(t) of the current step and the hidden memory information h(t-1) of the last step. And through nonlinear transformation, the memory information h(t) of the current step is obtained. The last neural unit of the encoder outputs the memory information h(t) of the entire time sequence to the decoder. The entire encoder part is to integrate and nonlinearly transform the effective features in the state vector of the entire time sequence.

[0093] 2.2.2 Decoder part: The basic unit of the decoder is also LSTM neural unit, which receives two parts: the memory information h(t) output by the encoder and the output y(t-1) of the last time sequence. It can organize the information of the input state variable of the current time sequence and the output structure of the last time sequence, learn the nonlinear mapping relationship between the target output and the two kinds of information, and thus obtain the target output.

[0094] 2.2.3 Attention mechanism part: The attention mechanism part is essentially a small neural network, which focuses on the appropriate encoder output information in each time step, and filters useful information. The attention mechanism layer receives all the outputs from the encoder and all the hidden states from the decoder at the last time sequence, and uses weight coefficients e(i,j) to evaluate the alignment degree. In this way, the importance of each encoder information can be obtained, and the information is multiplied by the corresponding weight coefficient to output. Finally, all scores are passed through the softmax layer to obtain the maximum weight a(i,j) of each encoder output.

[0095] 2.2.4 Select activation function

[0096] LeakyReLU is selected as the activation function of each layer, referring to Figure 4 .

[0097] LeakyReLU α (x)=max(αx,x)

[0098] The hyperparameter a defines the degree of "leak" of the function: it is the slope when x < 0, usually set to 0.01. This variant of the ReLU function makes up for the disadvantage of the ReLU function that cannot output less than zero, and retains the advantage of fast convergence speed.

[0099] 3) Giorot initialization hyperparameters

[0100] Normal distribution, mean is 0, variance is

[0101] Or uniform distribution between -r and +r, where

[0102] where:

[0103] fan avg = (fan in + fan out ) / 2

[0104] fan in is the number of input neurons

[0105] fan out is the number of output neurons

[0106] Glorot initialization solves the problem of unstable gradient in deep neural networks, and can ensure normal forward prediction and reverse optimization of signals.

[0107] 4) Use Bayesian optimization to select hyperparameters

[0108] The Bayesian optimization algorithm learns the shape of the target function and finds the parameters that make the target function improve towards the global optimal value. First, according to the prior distribution, assume a search function; then, every time a new sampling point is used to test the target function, use this information to update the prior distribution of the target function; finally, the algorithm tests the point where the global maximum value given by the posterior distribution is most likely to appear.

[0109] Its pseudo code is as follows:

[0110]

[0111] where:

[0112] f: So-called black box, that is, input a set of hyperparameters, get an output value.

[0113] X: Hyperparameter search space.

[0114] D: Represents a data set consisting of several pairs of data.

[0115] S: is the collection function, used to select x.

[0116] M: is the Gaussian model obtained by fitting the data set D.

[0117] Bayesian optimization does not search every possible combination, but randomly selects the first few. Then according to the performance of these hyperparameters, select the next possible best value. Therefore, the selection of each hyperparameter depends on the previous attempt. According to the historical record, select the next set of hyperparameters and evaluate the performance until the best combination of hyperparameters is found.

[0118] Advantages: Bayesian optimization not only optimizes the accuracy of the model, but also reduces the time cost compared with grid search, random search and other methods.

[0119] 5) Use mini-batch stochastic gradient descent and momentum optimization to optimize the state observer model

[0120] 5.1 Mini-batch stochastic gradient descent

[0121] For neural networks, the optimization process is the process of constantly updating the model weights. Gradient descent is one of the most common methods to update model weights.

[0122] First, determine the loss function:

[0123]

[0124] Where:

[0125] The goal is to minimize the value of the loss function J (θ) According to the gradient descent method, so take the partial derivative of J (θ) with respect to θ and update the weights:

[0126]

[0127] θ k is the kth weight before updating

[0128] θ' k is the kth weight after updating

[0129] η is the learning rate

[0130] J (θ) is the loss function

[0131] m represents how many samples are taken for training each time. If single-value stochastic gradient descent is used for training, a group of samples is randomly taken each time, and m = 1; if batch processing is used, m is equal to the number of samples taken as training samples each time.

[0132] Advantages: Since training with a single sample can bring a lot of noise, single-value stochastic gradient descent may not always converge in the direction of global optimization, so it may converge quickly at the beginning of training, but it will become very slow after a period of training. On this basis, using mini-batch gradient descent, a small batch of samples can be randomly taken for training each time instead of a group. This can avoid getting stuck in local optima and converge to the global optimum faster.

[0133] 5.2 Momentum optimization

[0134] The conventional gradient descent method updates the weights by subtracting the gradient of the loss function of the weights multiplied by the learning rate . Only a small, regular step is taken on the slope, so the algorithm will take more time to optimize to the bottom.

[0135] Momentum optimization cares very much about what the previous gradient was: at each iteration, it subtracts the local gradient from the momentum vector g and updates the weights by adding that momentum vector. The formula is as follows:

[0136]

[0137] θ k →θ ′ k +g ′

[0138] In the formula:

[0139] g is the momentum vector before update

[0140] g ′ is the momentum vector after update

[0141] β is the momentum parameter

[0142] The gradient in momentum optimization is used for acceleration, not velocity, to simulate some kind of friction mechanism and prevent momentum from becoming too large, the algorithm introduces a new hyperparameter β, called momentum, which is typically set to 0.9.

[0143] Advantages: ordinary gradient descent method will be quite fast along the larger slope, but when the slope is small, it takes too long. The corresponding momentum optimization gradient descent method can be faster and faster until it is optimal. Because of the existence of momentum parameter β, it can also avoid the problem of oscillation and overshoot at the optimal point.

[0144] 6) Use dropout regularization and early stopping to avoid overfitting

[0145] 6.1 Use dropout regularization:

[0146] When the number of learning data samples is too small or the network is too complex, the neural network focuses on the special features of the training samples, and loses the acquisition of the general features of the learning samples. Such a model fits the data of the training set very well, but its generalization ability is not strong, and its fitting ability for data that the model "does not recognize" is poor. This phenomenon is called model overfitting, and the method to prevent this situation is called regularization.

[0147] For deep neural networks, dropout is one of the most popular regularization techniques. The principle is simple: at each training step, each neuron has a probability p of being temporarily "deleted", which means that it is completely ignored for this training step, but can be active in the next step. The hyperparameter p is called the dropout rate and is usually set to 10-50%. Neurons trained with dropout cannot adapt to their neighboring neurons, they have to do their best.

[0148] Advantages: Each neuron trained with dropout cannot rely too much on a few input neurons, but has to pay attention to all of them, and is ultimately less sensitive to small changes in the input. A more robust model can be obtained, which is also more generalizable.

[0149] 6.2 Prevent overfitting with early stopping

[0150] The callback function can save the model periodically during training, and stop training when there is no further optimization on the validation set, and can choose to roll back to the best model.

[0151] At the end of each round of training, the callback function compares the error of the current validation set with the previously saved error. Only when the model performance on the validation set reaches the best so far, the model is selected to be saved. This way, there is no need to worry about overfitting due to too long training time. To prevent wasting time by setting too large a number of training cycles, the early stopping method can be used. If there is no progress on the validation set for a number of rounds (10 is chosen in this paper), the training will be interrupted and the best model saved previously will be chosen to roll back.

[0152] Advantages: both the problem of model overfitting due to too many training cycles and the saving of training time by early stopping are solved.

[0153] 7) Model training and saving

[0154] The preprocessed data in the training set and validation set part are imported into the model for training, the body pitch angle θ, the body acceleration Front suspension relative displacement Rear suspension relative displacement Front suspension unsprung mass acceleration Rear suspension unsprung mass acceleration As input; vehicle front suspension unsprung mass displacement Rear suspension unsprung mass displacement State variables and front wheel road displacement Rear wheel road displacement Front wheel tire dynamic load Rear wheel tire dynamic load Front wheel road force input F l f Rear wheel road force input As output, the number of training is 4000. And save the best model. The state observer model schematic diagram reference Figure 5 .

[0155] 8) State observer model prediction and accuracy verification

[0156] The body pitch angle θ, body acceleration Front suspension relative displacement Rear suspension relative displacement Front suspension unsprung mass acceleration Rear suspension unsprung mass acceleration Input into the model to make predictions. Then compare the state observer output with the real test set corresponding output, test results show that the model accuracy up to 96.35%.

Claims

1. A method for constructing a state observer for a vehicle suspension system based on an attention mechanism using an LSTM neural network, comprising the following steps: Step 1: Build a dynamic model of the half-vehicle suspension system on simulation software and simulate the suspension system state variables and road surface information. Step 2 includes: 2.1 data preprocessing and 2.2 designing the state observer model. 2.1 The data preprocessing mentioned above includes: 2.1.1 State variable data are divided into inputs and outputs. State variables that are easy for sensors to measure are inputs, while state variables that are difficult to measure and road surface information are outputs. 2.1.2 Convert the state variables into time series with fixed time steps and normalize the data; 2.1.3 Using random sampling, all data are divided into training set, validation set and test set, with an allocation ratio of 10:2:2; The training set data is used to train the state observer model, the validation set data is used to cross-validate the model's accuracy during training, and to iteratively optimize the model; the test set data is used to finally test the model's generalization ability and verify the state observation model's estimation accuracy for unknown state variables. 2.2 Design of the state observer model: An encoder-decoder neural network model is constructed, with an attention mechanism layer built between the encoder and decoder layers. The encoder layer processes the information in the input sequence; each neuron receives the input of the current time step and the output of the previous time step, and retains the sequence information. The attention mechanism layer combines the output of the encoder layer with the output of the decoder of the previous time step and evaluates their correspondence, focusing on learning data closely related to the target output. The decoder layer takes the output of the attention mechanism layer as input and outputs the state variable information corresponding to each time step. Step 3, initialize hyperparameters using Glorot: determine the distribution range of randomly initialized parameters based on the number of inputs and outputs of each layer; Step 4: Use Bayesian optimization to select hyperparameters: learning rate and number of neurons in each layer; first, determine the objective function, then determine the distribution domain space of each hyperparameter, and then design and optimize the state observer model. Finally, show all the optimal model hyperparameters. Step 5: Optimize the state observer model using mini-batch stochastic descent and momentum optimization; Step 6: Optimize the model using dropout regularization and early stopping to prevent the state observer model from overfitting. Step 7, Model Training and Saving: Record the vehicle pitch angle θ and vehicle acceleration... Relative displacement of front suspension Rear suspension relative displacement Front suspension unsprung mass acceleration Rear suspension unsprung mass acceleration The time series of state variables is used as input; the unsprung mass displacement of the vehicle's front suspension is used as input. Rear suspension unsprung mass displacement State variables and front wheel road surface displacement Rear wheel road surface displacement Front tire dynamic load Rear tire dynamic load Front wheel road force input Rear wheel road force input The time series data is used as output, imported data is used to train the state observer model, and the observer model is saved. Step 8, Model Accuracy Validation: The predicted unknown state variables are compared with the outputs in the validation set to obtain a model accuracy of 96.35%, which is more accurate than traditional modeling methods.

2. The method for constructing a state observer for a vehicle suspension system based on an attention mechanism using an LSTM neural network, as described in claim 1, is characterized in that... Step 1: The dynamic model of the half-vehicle suspension system is an air spring suspension model. Its spring stiffness is obtained by polynomial fitting, and the damper damping force is learned by a neural network model. Both are nonlinear models. The state observer is used to solve the problem of difficulty in nonlinear solution. However, the state observer is not limited to the suspension model mentioned above. The dynamic equations of the half-vehicle suspension system can be expressed as: In the formula: m s For vehicle weight, I s For the vehicle body's rotational inertia, x represents the unsprung mass of the front and rear wheels. s Let θ be the vehicle body displacement, and θ be the vehicle body pitch angle. For the front and rear wheel road surface displacement, For the front and rear unsprung displacement, For the front and rear spring displacement, For front and rear tire stiffness, For the nonlinear air spring forces of the front and rear suspensions, The values ​​represent the nonlinear damping forces of the front and rear suspensions, where 'a' is the distance from the vehicle's center of gravity to the front axle, and 'b' is the distance from the vehicle's center of gravity to the rear axle. in: The fitted nonlinear air spring force is: Using the time-domain expression of filtered white noise as the road surface input model, the road surface input equations for the front and rear wheels are as follows: In the formula: f0 is the lower cutoff frequency, G0 is the road surface roughness coefficient, and u c w1 represents the vehicle speed, and w2 represents the random white noise input from the front and rear road surfaces.

3. The method for constructing a state observer for a vehicle suspension system based on an attention mechanism using an LSTM neural network, as described in claim 1, is characterized in that... Step 2.2 also includes: LSTM calculation of the equation: In this equation: W xi W xf W xo W xg It is the relationship between each of the four layers and the input vector x. (t) Connection weight matrix W hi W hf W ho W hg Each of the four layers is related to the previous short-term state h. (t-1) Connection weight matrix b i b f b o b g It is the bias term of each of the four layers. Step 2.1.3 also includes: splitting the sequential data into a training set, a validation set, and a test set: using the first 80% of the data as the training set, the middle 10% as the validation set, and the last 10% as the test set; Step 2.1.2 also includes: normalizing the data to unify the order of magnitude of the data and prevent erroneous learning results due to differences in the magnitude of the values; and dividing the sequential dataset into multiple time series: determining the time step of each state vector and then using a staggered data collection method to divide all the data into sequences with the same time step.

4. The method for constructing a state observer for a vehicle suspension system based on an attention mechanism using an LSTM neural network, as described in claim 3, is characterized in that... Step 2.1.2 also includes: Scaling function: In the formula: F' represents the normalized data, F text For the data before normalization, F max F is the maximum value among all input data. min The maximum value among all input data.

5. The method for constructing a state observer for a vehicle suspension system based on an attention mechanism using an LSTM neural network, as described in claim 3, is characterized in that: Step 2.1.5 also includes: State variables that are easily measured using sensors are used as training inputs for the model, including: vehicle pitch angle θ and vehicle acceleration. Relative displacement of front suspension Rear suspension relative displacement Front suspension unsprung mass acceleration Rear suspension unsprung mass acceleration The model's training output includes difficult-to-measure state variables and road surface information, including: unsprung mass displacement of the front suspension. Rear suspension unsprung mass displacement State variables and front wheel road surface displacement Rear wheel road surface displacement Front tire dynamic load Rear tire dynamic load Front wheel road force input Rear wheel road force input 6. The method for constructing a state observer for a vehicle suspension system based on an attention mechanism using an LSTM neural network, as described in claim 1, is characterized in that: Step 2.2 includes: 2.2.1 Encoder section: The basic unit of the encoder is an LSTM neural unit, which receives the state vector x(t) of the current step size and the hidden memory information h(t-1) of the previous step size, and obtains the memory information h(t) of the current step size through nonlinear transformation. The last neural unit of the encoder outputs the memory information h(t) of the entire time series to the decoder. 2.2.2 Decoder Section: The basic unit of the decoder is also an LSTM neural unit. It receives two parts: the memory information h(t) output by the encoder and the output y(t-1) of the previous time series. It can organize the information of the input state variables of the current time series and the output structure of the previous time series, learn the nonlinear mapping relationship between the target output and these two types of information, and thus obtain the target output. 2.2.3 Attention Mechanism: The attention mechanism layer receives all outputs from the encoder and all hidden states from the decoder in the previous time series. It evaluates the alignment using the weight coefficient e(i,j), thus obtaining the importance of each encoder information. The information is multiplied by the corresponding weight coefficient and output. Finally, all scores pass through the softmax layer to obtain the final weight α(i,j) of each encoder output. Step 2.2.4: The activation function LeakyReLU is: LeakyReLU α (x)=max(αx,x) The hyperparameter α defines the degree to which the function "leaks": it is the slope when x < 0, and is usually set to 0.

01.

7. The method for constructing a state observer for a vehicle suspension system based on an attention mechanism using an LSTM neural network, as described in claim 1, is characterized in that: In step 3, the distribution range of the random initialization parameters is as follows: The distribution is normal, with a mean of 0 and a variance of . Or a uniform distribution between -r and +r, where in: from avg =(from in +of out ) / 2 In the formula: fan in For the number of input neurons, fan out This represents the number of output neurons.

8. The method for constructing a state observer for a vehicle suspension system based on an attention mechanism using an LSTM neural network, as described in claim 1, is characterized in that: In step 4: The Bayesian optimization algorithm learns the shape of the objective function to find the parameters that improve the objective function toward the global optimum. First, it assumes a collection function based on the prior distribution. Then, each time a new sampling point is used to test the objective function, this information is used to update the prior distribution of the objective function. Finally, the algorithm tests the point where the global optimum is most likely to occur, as given by the posterior distribution. Bayesian optimization does not search every possible combination, but rather randomly selects the first few; then, based on the performance of these hyperparameters, it selects the next possible optimal value; therefore, the selection of each hyperparameter depends on previous attempts; the next set of hyperparameters is selected based on historical records and the performance is evaluated until the optimal combination of hyperparameters is found.

9. The method for constructing a state observer for a vehicle suspension system based on an attention mechanism using an LSTM neural network, as described in claim 8, is characterized in that: Its pseudocode is as follows: Input: f, X, S, M D←InitSamples(f,X) for i←|D|to T do: p(y|x,D)←FitModel(M,D) x i ←argmax S(x,p(y|x,D)) and i ←f(x i ) D←D∪(x i ,and i ) end for In the formula: f: The so-called black box, which is a set of hyperparameters input and an output value output. X: Hyperparameter search space, D: Represents a dataset consisting of several pairs of data. S: This is the acquisition function used to select x. M: is the Gaussian model obtained by fitting the dataset D.

10. The method for constructing a state observer of a vehicle suspension system based on an attention mechanism using an LSTM neural network according to claim 1, characterized in that: Step 5 includes: 5.1 First, determine the loss function: in: Perform weight updates: θ k The k-th weight before the update θ' k For the k-th weight after the update η is the learning rate J (θ) loss function 5.2 In each iteration, it subtracts the local gradient from the momentum vector g and updates the weights by adding the momentum vector, as shown in the following formula: i k →θ ′ k +g ′ In the formula: g is the momentum vector before the update. g ′ For the updated momentum vector, β is the momentum parameter, typically set to 0.

9.

11. The method for constructing a state observer for a vehicle suspension system based on an attention mechanism using an LSTM neural network, as described in claim 1, is characterized in that: Step 6.1: In each training step, each neuron has a probability p of being temporarily "dropped out", which means that it is completely ignored in this training step, but may be active in the next step. The hyperparameter p is called the dropout rate, which is usually set to 10% to 50%. Step 6.2: Only save the model when its performance on the validation set reaches its best so far. This way, you don't have to worry about overfitting due to excessive training time. If there is no progress on the validation set after multiple rounds, including 10 rounds, training will be interrupted and you will choose to roll back to the previously saved best model.

Images