System-level whole vehicle vibration prediction method based on deep learning

Abstract

The invention discloses a system-level whole vehicle vibration prediction method based on deep learning, and provides a method for constructing a proper correlation model through an LSTM (Long Short Term Memory) model by using a deep learning method. By adding a dropout layer and discarding unimportant vibration moments, extraction of time sequence vibration features is realized, and the constructed correlation model has higher generalization and can adapt to prediction of various complex problems. Compared with the prior art, the correlation model constructed by the method has lower time and space cost compared with a traditional multi-body dynamic model; compared with an FNN method in the same field, the method can better reflect the system-level vibration characteristics, and finally obtains higher prediction precision.

Description

Technical Field

[0001] The present invention relates to the technical field of whole vehicle vibration prediction of a tracked vehicle system, and in particular to a system-level whole vehicle vibration prediction method based on deep learning. Background Art

[0002] Vibration prediction of tracked vehicles refers to analyzing the existing vibration response data and analyzing and predicting the vibration characteristics of the whole vehicle under specific driving environments such as different road conditions and different vehicle speeds.

[0003] The vibration analysis of tracked vehicles usually adopts multi-body dynamics models and random road models, with multi-objective optimization design requirements. If the simulation model method based on physical mechanism is used to analyze its dynamic parameters, due to its complex sub-component simulation model, multi-source data and heterogeneous simulation interface, the integration of sub-components of the simulation model of the tracked vehicle system is time-consuming and laborious, resulting in high time and space costs for predicting its vibration characteristics. In order to optimize the vibration prediction requirements of the tracked vehicle system, it is necessary to establish a whole vehicle vibration prediction method that can adapt to its internal complex mechanism and effectively extract the vehicle vibration characteristics. Summary of the invention

[0004] The present disclosure provides a method for predicting the vibration of a tracked vehicle system based on deep learning technology, which uses a multi-layer neural network to learn data features and rules from high-dimensional, unstructured and complex data to solve the problem of predicting the vibration response of a tracked vehicle during driving. The core problem of deep learning is how to train the neural network, that is, how to adjust the weights and biases of neurons so that the neural network can learn useful information from the data and achieve the desired goal.

[0005] The present invention discloses a system-level vehicle vibration prediction method that reduces overfitting by extracting time series vibration features: it adopts LSTM as a nonlinear model to avoid the gradient vanishing problem in the learning process of long time series data sets, so that the gradient decrease rate over time will be linear rather than exponential; by adding a dropout layer that is randomly inactivated with a certain probability to extract time series vibration features, the constructed correlation model is prevented from overfitting.

[0006] The specific steps include:

[0007] 1. Establish a multi-level correlation model architecture based on the transmission path

[0008] There is a multi-level parameter transfer relationship between the typical environmental working conditions of tracked vehicles and the system-level vibration state: overall operating state → overall working condition parameters → local load parameters of each excitation source → vibration response of key positions of the system. Among them, the overall operating state includes factors such as drive mode, action control mode, and mission load; the overall working condition parameters include the force on the left driving wheel, the torque of the left driving wheel, the speed of the right driving wheel, the torque of the right driving wheel, etc.; the local load working condition parameters corresponding to each excitation source include the dynamic load of the first to sixth wheel axle head, the torque of the first to sixth wheel axle head, etc.; the system-level vibration state parameters include the vertical displacement of the vehicle body, the vertical velocity of the vehicle body, the vertical acceleration of the vehicle body, the pitch angular displacement of the vehicle body, the pitch angular velocity of the vehicle body, and the pitch angular acceleration of the vehicle body.

[0009] Directly using intelligent algorithms such as machine learning to establish a correlation model between typical working conditions and system-level vibration states across levels would require too large a number of experimental samples and may lead to the problem of being unable to explain the rationality of the learning results. Therefore, a hierarchical modeling approach is adopted in this disclosure to reduce learning costs and improve the rationality of the model.

[0010] 2. Construct a correlation model between operating status and overall operating parameters

[0011] In order to determine the typical vibration response and sensitive factors under different mission loads, it is necessary to determine the mapping set of typical overall operating states and overall operating parameters of tracked vehicles. Each operating state is a combination of road surface, vehicle speed, engine speed, etc. Among them, the overall operating state includes factors such as drive mode, action control mode, and mission load; the overall operating parameters include left driving wheel force, left driving wheel torque, right driving wheel speed, right driving wheel torque, etc.

[0012] The overall operating condition parameter mapping association model is a different operating condition parameter mapping set for each driving mode. The operating condition parameter mapping association model is preferably constructed in a database manner.

[0013] 3. Construct the correlation model between the working condition parameters and the loads of each local excitation source

[0014] The model is used to predict the loads of local excitation sources in a certain time domain in the future during driving under different working conditions. The overall working condition parameters such as the driving wheel force and the driving wheel output torque are recorded, and the local loads at the excitation source positions such as the road wheel and the power transmission support are tested. The neural network model is used to establish the correlation between the overall working condition parameters and the loads of each local excitation source.

[0015] The input of this correlation model is 4 parameters, namely, the force on the left driving wheel, the torque on the left driving wheel, the force on the right driving wheel, and the torque on the right driving wheel. Its output is 24 parameters, namely, the axle head load of the first to sixth wheels on the left, the axle head torque of the first to sixth wheels on the left, the axle head load of the first to sixth wheels on the right, and the axle head torque of the first to sixth wheels on the right.

[0016] The deep learning model of system-level vehicle vibration needs to meet the needs of predicting time series characteristics, so this disclosure introduces a neural network with a "memory" function: Long-Short Term Memory (LSTM). As a nonlinear model, LSTM can be used as a complex nonlinear unit to construct a larger deep neural network. This disclosure uses LSTM as a nonlinear model, so that the rate of gradient decrease over time will be linear rather than exponential, avoiding the gradient vanishing problem in the learning process of long time series data sets;

[0017] The structural complexity of the neural network will affect its computational efficiency and prediction performance. When the number of neurons is too large, the network has a stronger learning ability, but it requires more computing resources and is prone to overfitting the prediction model, reducing its generalization ability. In this regard, the present disclosure adopts the dropout random deactivation method. Even when the input data propagates forward, some neurons stop working with a certain probability, so that the model will not rely too much on local features, prevent the association model constructed by the neural network from overfitting, and effectively improve the versatility of the model.

[0018] The model building process specifically includes:

[0019] Step (1): Based on the LSTM nonlinear model, select appropriate neural network parameters and configure a model that meets the requirements.

[0020] In the neural network model configuration process, you need to adjust the three parameters of network layers, number of neurons, and number of training cycles. When configuring a neural network, first find the most optimal network parameters with training samples, then gradually reduce the training data, check the performance of the network when the training samples are insufficient, and finally select different networks based on different sample numbers.

[0021] Furthermore, the specific configuration process includes:

[0022] First, the number of network layers and neurons is set, and then the number of neurons is gradually increased to find the minimum number of neurons that enables the neural network to achieve the highest prediction accuracy;

[0023] Next, fix the number of neurons, gradually increase the number of training steps, and find the number of steps that achieves the best accuracy under a certain training time;

[0024] Then gradually reduce the number of training samples to clarify the degree of dependence of the model prediction accuracy on the training samples, and finally compare to obtain the optimal network structure.

[0025] Step (2), add a dropout layer to the above neural network model and set an appropriate deactivation probability to prevent overfitting of the deep learning model.

[0026] Furthermore, the model structure includes input layer, LSTM layer, dropout layer, fully connected layer and output layer. First is the input layer, which inputs a data matrix of the number of data points multiplied by the number of input parameters, and predicts the next data point. Then there are several LSTM layers, each of which is connected to a dropout layer, and the number of neurons is set separately. After all LSTM layers are built, they are connected to the fully connected layer, and each fully connected layer is connected to a dropout layer. Finally, there is the output layer, and the number of neurons should be consistent with the number of output parameters.

[0027] Step (3) applies the neural network model to conduct deep learning on the data set and constructs an association model suitable for known input and output parameters. Then, the association model is compared with the test data to verify the reliability of the association model.

[0028] 4. Construct a correlation model between local excitation source and system vibration state parameters

[0029] The input of the third layer is the output parameters of the second layer, and the output of the third layer is six system-level vibration parameters, including vehicle body vertical displacement, vehicle body vertical velocity, vehicle body vertical acceleration, vehicle body pitch angular displacement, vehicle body pitch angular velocity, and vehicle body pitch angular acceleration.

[0030] Preferably, the association model of this layer adopts the same architecture as the association model of the second layer, and the difference lies only in the input and output parameters.

[0031] Compared with the prior art, the present invention has the following beneficial effects: (1) based on deep learning technology, a whole vehicle vibration prediction method suitable for tracked vehicle systems is established. Based on the deep learning method, a system-level whole vehicle vibration prediction model constructed using an existing data set can better predict the local and global characteristics of system-level vibrations than the prior art, better reflect the characteristics of system-level vibrations, and have a higher prediction accuracy; (2) LSTM is used as a nonlinear model to construct a correlation model suitable for nonlinear strong and multi-input and output problems. By extracting features at different moments, the gradient vanishing problem in the learning process of long time series data sets is avoided, so that the gradient decrease rate over time will be linear rather than exponential, which is more suitable for the prediction of whole vehicle vibration characteristics at the system level of tracked vehicles; (3) by connecting a dropout layer that is randomly inactivated with a certain probability after each LSTM layer and fully connected layer, the interaction between hidden nodes is reduced, and moments without obvious feature information are discarded, so that the extraction of time series vibration features is realized, the generalization of the model is improved, and it can have a better prediction effect for different working conditions; (4) It has powerful nonlinear approximation characteristics and adaptive capabilities, and can be easily applied to multivariable control systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The above and other objects, features and advantages of the present disclosure will become more apparent through a more detailed description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present disclosure.

[0033] Figure 1 Temporal representation of long short-term memory networks;

[0034] Figure 2 Configure the process for neural network parameters;

[0035] Figure 3 is an exemplary LSTM association model architecture according to the present disclosure;

[0036] Figure 4 Training loss change process for the second layer LSTM model;

[0037] Figure 5 The training accuracy change process of the second-layer LSTM model;

[0038] Figure 6 Comparison of the prediction results of FNN and LSTM for a certain wheel axle head torque;

[0039] Figure 7 Training loss change process for the second layer LSTM model;

[0040] Figure 8 The training accuracy change process of the second-layer LSTM model;

[0041] Fig. 9 Comparison of vehicle center of mass displacement prediction results between FNN and LSTM;

[0042] Fig.10 Comparison of vehicle center of mass speed prediction results between FNN and LSTM;

[0043] Fig.11 Comparison of vehicle center of mass acceleration prediction results between FNN and LSTM;

[0044] Fig.12 Comparison of vehicle pitch angle prediction results between FNN and LSTM;

[0045] Fig.13 Comparison of vehicle pitch angular velocity prediction results between FNN and LSTM;

[0046] Fig.14 Comparison of vehicle pitch angular acceleration prediction results between FNN and LSTM. DETAILED DESCRIPTION

[0047] The preferred embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0048] This paper establishes a system-level vehicle vibration prediction method to reduce overfitting by extracting time-series vibration features. This method is based on deep learning technology, discards unimportant vibration information through the dropout method, extracts time-series vibration features, and combines LSTM (Long-Short Term Memory) technology to build a model, which can effectively predict the vibration response characteristics of tracked vehicles when they are driving.

[0049] This disclosure adopts a hierarchical modeling approach to reduce learning costs and improve the rationality of the model.

[0050] In one exemplary embodiment:

[0051] The first-level operating parameter mapping association model is mainly constructed in the form of SQL database, which represents the mapping relationship between the overall operating status and the operating parameters.

[0052] The construction of the second and third layer association models specifically includes the following steps:

[0053] Step (1): In order to analyze the dependence of each parameter on the time series, find the optimal dynamic model, and prevent the gradient from disappearing, LSTM is used as the model. Its time series structure is as follows: Figure 1 As shown:

[0054] Whenever the hidden layer output h at the previous moment t-1 When entering the current moment of calculation, it will be combined with the current input x t Combined and added to the information flow C in two ways t-1 Thus, we get the hidden layer input C at the current moment. t , and then through the third effect, we finally get the hidden layer output h at the current moment t These three functions are called forget gate, input gate and output gate, and are represented by A, B and C in the figure respectively. Among them:

[0055] The forget gate is a sigmoid function. By reading the input matrix, a matrix with the same dimension and element size between 0 and 1 is obtained. The matrix is ​​multiplied by the information flow. The information corresponding to the position of 1 in the matrix is ​​completely retained, while the information corresponding to the position of 0 is completely forgotten. Its mathematical expression can be written as:

[0056] f t =σ(W f ·[h t-1 , x t ]+b f ) (1)

[0057] Among them, f t That is, the binary matrix that passes through the forget gate, W f is the weight matrix of the forget gate, b f is the bias matrix of the forget gate;

[0058] The function of the input gate is to add the information that needs to be updated. t-1 With x t After confluence, they are multiplied by the weight matrix and passed through the activation function f1 to obtain new important information, which is then added to the information flow to form the hidden layer input C at the current moment. t , the mathematical expression of this process is:

[0059] i t =o(W i ·[h t-1 , x t ]+b l ) (2)

[0060]

[0061] Among them, i t is the screening matrix through the sigmoid function, is the processed information of the current moment and the previous moment. Therefore, the input of the hidden layer can be expressed as:

[0062]

[0063] The function of the output gate is similar to that of the hidden layer of other neural networks, except that the data is further filtered by the sigmoid function while passing through the activation function, and is passed to the output layer as the hidden layer output to obtain the final result.

[0064] Step (2): Configure the parameters of the LSTM model. The configuration criteria and process are as follows: Figure 2 shown.

[0065] In this embodiment, it is preferred to use the control variable method to gradually determine the three parameters of the number of network layers, the number of neurons, and the number of training cycles. When configuring a certain neural network, first find the optimal network parameters with training samples, then gradually reduce the training data, check the performance of the network when the training samples are insufficient, and finally select different networks according to different sample numbers. Specifically including:

[0066] First, the network is set as a two-layer network, and the initial number of neurons is set to 20. Then, the number of neurons is gradually increased to find the minimum number of neurons that enables the neural network to achieve the highest prediction accuracy. Next, the number of neurons is fixed, and the number of training steps is gradually increased to find the number of steps that achieve the best accuracy when the training time does not exceed 90 seconds. Then the number of training samples is gradually reduced to study the degree of dependence of the model prediction accuracy on the training samples, and finally the optimal network structure is obtained by comparison.

[0067] Step (3): Based on the above configuration criteria, the above construction method based on the LSTM association model is verified. The LSTM model is configured with 6 system-level vibration parameters as output values. Its input is a matrix with a length of Time Sequance and a width of the number of input parameters, that is, the parameter value at the next moment is predicted through the historical data trend of the previous Time sequence data points.

[0068] Step (4): Use LSTM to construct a model that associates operating parameters with excitation loads.

[0069] The LSTM model constructed in this embodiment is as follows Figure 3 As shown in the figure, it contains three LSTM layers and three fully connected layers. Each layer is also connected to a dropout layer to prevent overfitting of the deep learning model.

[0070] First is the input layer, the input is a data matrix of 40 times the number of input parameters, that is, the input sequence length is 40, and the 41st data point is predicted.

[0071] Next comes the first LSTM layer, with 100 LSTM neurons. It is followed by a random dropout layer. As a preferred method, the dropout probability of all random dropout layers in this LSTM model is uniformly set to 0.4. Then comes the second LSTM layer, with 100 neurons, followed by a random dropout layer. Then comes the third LSTM layer, with 50 neurons.

[0072] After all LSTM layers are built, a fully connected layer is connected. The number of neurons in the first fully connected layer is 64, followed by a random dropout layer. The number of neurons in the second fully connected layer is 32, and it is also connected to a random dropout layer.

[0073] Finally, there is the output layer. The number of neurons in the output layer is consistent with the number of output parameters of the model architecture of the association model between the operating parameters and each local excitation load (the second-layer association model) and the association model between the local excitation source and the system vibration state parameters (the third layer).

[0074] Step (5): Train the LSTM model and build an association model.

[0075] Refer to the configuration parameters obtained in step (2) and select an appropriate number of training times. The changes of loss and accuracy during training with the number of training times are shown in Figure 4 , Figure 5 shown.

[0076] Step (6): Use the constructed correlation model to predict the loads on the local key positions of the tracked vehicle system and verify the prediction effect of the model.

[0077] The input of this model structure is 4 parameters, namely, the force on the left driving wheel, the torque on the left driving wheel, the force on the right driving wheel, and the torque on the right driving wheel. The output is 11 parameters, namely, the local excitation source. The prediction results of a certain part and the comparison with the true value and the prediction results of the traditional fully-connected neural network (FNN) are as follows: Figure 6 As shown in the comparison results, it can be seen that the second-layer LSTM association model can better predict the local and global characteristics of system-level vibration and better reflect the system-level vibration characteristics.

[0078] Step (7): Use LSTM to construct a correlation model between excitation load and vibration parameters.

[0079] The structure of the third-layer LSTM model is consistent with that of the second layer. The difference is that the input of the third-layer LSTM model is 11 local excitation source parameters, and the output is 6 system-level vibration parameters. Therefore, the number of input neurons is 11, and the number of output neurons is 6. The loss and accuracy of the third-layer LSTM model change with the training process as shown in Figure 2. Figure 7 , Figure 8 shown.

[0080] Step (8): Use the constructed correlation model to predict the loads on the local key positions of the tracked vehicle system and verify the prediction effect of the model.

[0081] The vehicle center of mass displacement predicted by the model and the comparison with the true value and the FNN model prediction results are shown in the figure. Figures 9 to 14 The results show that the LSTM model can improve the prediction accuracy by 4.4% compared with the FNN model, and is more suitable for the construction of tracked vehicle association models.

[0082] Using the above system-level vehicle vibration prediction method based on deep learning, a method for constructing an appropriate association model through the LSTM model is given using a deep learning method. By adding a dropout layer, unimportant vibration moments are discarded to extract the time series vibration features. The constructed association model has higher generalization and can adapt to the prediction of various complex problems.

[0083] Compared with the existing technology, the constructed correlation model has lower time and space costs than the traditional multi-body dynamics model; compared with the FNN method in the same field, it can better reflect the system-level vibration characteristics and ultimately obtain higher prediction accuracy.

[0084] The above technical scheme is only an exemplary embodiment of the present invention. For those skilled in the art, it is easy to make various types of improvements or modifications based on the application methods and principles disclosed in the present invention, and it is not limited to the method described in the above specific embodiment of the present invention. Therefore, the method described above is only preferred and does not have a restrictive meaning.

Claims

1. A system-level vehicle vibration prediction method based on deep learning, The following steps are involved: S0, establish a multi-level parameter transfer relationship between the typical environmental conditions of tracked vehicles and the system-level vibration state: overall operating state → overall operating condition parameters → local load parameters of each excitation source → vibration response of key positions of the system; S1, using database to build a mapping association model between the overall operating status and the overall operating condition parameters; S2, based on LSTM, constructs a correlation model structure between the working condition parameters and the local load parameters of each excitation source, which is used to predict the parameter value at the next moment through the historical data trend of several previous data points; wherein, the correlation model includes a dropout layer that is randomly inactivated with a certain probability, which is used to extract the time series vibration features to prevent the constructed correlation model from overfitting; S3, using the constructed correlation model between the working condition parameters and each local excitation load, predict the load on the local key position of the tracked vehicle system to verify the prediction effect of the model; S4, based on LSTM, builds a correlation model between the local load parameters of the excitation source and the vibration state parameters of the system; S5, the constructed correlation model between the local load parameters of the excitation source and the system vibration state parameters is used to predict the system-level vibration parameters of the tracked vehicle.

2. The method according to claim 1, It is characterized in that In the step S0: The overall operating state includes: one or more of a driving mode, an action control mode, and a mission load change mode; The overall working condition parameters include: one or more of the force on the left driving wheel, the torque on the left driving wheel, the speed of the right driving wheel, and the torque on the right driving wheel; The local load condition parameters corresponding to each excitation source include: one or more of the following: the dynamic loads of the first to sixth wheel axle heads and the torques of the first to sixth wheel axle heads; The system-level vibration state parameters include: one or more of vehicle body vertical displacement, vehicle body vertical velocity, vehicle body vertical acceleration, vehicle body pitch angular displacement, vehicle body pitch angular velocity, and vehicle body pitch angular acceleration.

3. The method according to claim 1 or 2, It is characterized in that The step S2 comprises the following steps: S21, based on the LSTM nonlinear model, selecting appropriate neural network parameters and configuring a model that meets the requirements, wherein the neural network parameters include: the number of network layers, the number of neurons, and the number of training cycles; S22, adding a dropout layer to the above neural network model and setting an appropriate inactivation probability to prevent overfitting of the deep learning model; S23, applying the neural network model to perform deep learning on the data set; and verifying the reliability of the model by comparing it with the test data, and finally obtaining an association model suitable for known input and output parameters.

4. The method according to claim 3, It is characterized in that In step S21, the step of configuring a model that meets the requirements specifically includes: Initially set the number of network layers and neurons, then gradually increase the number of neurons to find the minimum number of neurons that enables the neural network to achieve the highest prediction accuracy; Then fix the number of neurons, gradually increase the number of training steps, and find the number of steps that achieves the best accuracy under a certain training time; Then gradually reduce the number of training samples to clarify the degree of dependence of the model prediction accuracy on the training samples, and finally compare to obtain the optimal network structure.

5. The method according to claim 1, It is characterized in that In step S2, the constructed working condition parameter and excitation load association model includes: an input layer, an LSTM layer, a dropout random deactivation layer, a fully connected layer and an output layer, wherein: First, the input layer, which inputs a data matrix that is the number of data points multiplied by the number of input parameters, is used to predict the next data point; Then there are several LSTM layers, each of which is connected to a dropout layer, and the number of neurons is set respectively; After all LSTM layers are built, they are connected to the fully connected layer. Similarly, each fully connected layer is connected to a dropout layer. The last layer is the output layer, whose number of neurons is consistent with the number of output parameters of the associated model.

6. The method according to claim 5, It is characterized in that In step S2, the constructed working condition parameter and excitation load association model specifically includes: three LSTM layers and three fully connected layers, wherein: First is the input layer; Next comes the first LSTM layer, followed by a random dropout layer; Then comes the second LSTM layer, followed by a random dropout layer; Then the third LSTM layer; Then connect a fully connected layer, followed by a random dropout layer; Then comes the second fully connected layer, followed by a random dropout layer; Then the third fully connected layer; The last layer is the output layer. The number of neurons in the output layer is consistent with the number of output parameters of the correlation model architecture between the operating parameters and each local excitation load.

7. The method according to claim 1 or 5, It is characterized in that The correlation model structure in step S4 is the same as the correlation model structure in step S2, except that the input is the local load parameter of the excitation source, and the output is the system-level vibration parameter.

8. The method according to claim 1, It is characterized in that The inactivation probability of the dropout random inactivation layer in the association model is set to 0.4.

Images