Train stability prediction method and system based on MOGRIFIER-GRU deep learning

Abstract

The present invention relates to a train stability prediction method and system based on MOGRIFIER-GRU deep learning, the method comprising the following steps: step S1: obtaining train acceleration and wheel-rail force data under different road conditions, and normalizing the data to obtain training data for training a prediction model; step S2: building a MOGRIFIER-GRU deep learning prediction model, and training the MOGRIFIER-GRU deep learning prediction model through the training data to obtain a trained MOGRIFIER-GRU deep learning prediction model; step S3: collecting data transmitted back in real time by an actual running train, and predicting train stability through a trained MOGRIFIER-GRU deep learning prediction model. The method and system are conducive to reducing the cost of train stability monitoring and improving the accuracy of wheel-rail force prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of train stability prediction, and in particular relates to a train stability prediction method and system based on MOGRIFIER-GRU deep learning. Background Art

[0002] In recent years, the rapid development of high-speed railways in the world has not only brought huge economic benefits to the national economy and people's livelihood, but also raised new challenges to the research of related theories. With the increase in running speed, once a derailment accident occurs, it will cause greater casualties and property losses than lower-speed vehicles, seriously affecting the economic and social benefits of railway transportation. Therefore, real-time detection of stability indicators during train travel and early warning are of great practical significance.

[0003] my country's "Road Vehicle Dynamics Performance Evaluation and Test Identification Specification" (GB / T 5599-2019) stipulates that train stability indicators include derailment coefficient, wheel weight reduction rate, and wheel axle lateral force. To calculate these stability indicators, it is necessary to obtain the wheel-rail force between the track and the wheel, including vertical wheel-rail force and lateral wheel-rail force. The traditional wheel-rail force measurement method is achieved by using a dynamometer wheelset, but the processing of the dynamometer wheelset is quite complicated, the economic cost is huge, the failure rate is high, and the maintenance is inconvenient, which limits the promotion and use of the dynamometer wheelset. Therefore, the indirect measurement method of wheel-rail force has become a research difficulty in vehicle dynamics.

[0004] With the development of computer science and the improvement of the level of dynamics theory, existing dynamics software can establish multi-body dynamics models according to the actual parameters of the vehicle and perform dynamics calculation and analysis according to different working conditions, such as SIMPACK, ADAMS, etc. The train operation parameters obtained through simulation are used to predict wheel-rail forces and stability indicators. American scholars use multi-layer perceptron (MPL) neural networks to predict wheel-rail forces and derailment coefficients through track geometry parameters and running speed; professors at Southwest Jiaotong University use radial basis function neural networks to predict wheel-rail forces; Beijing Jiaotong University proposes the use of nonlinear autoregressive neural networks to predict derailment coefficients. These technologies all achieve the measurement and prediction of wheel-rail forces, but the accuracy is not ideal, the model performance is poor, and there are many shortcomings. Summary of the invention

[0005] The purpose of the present invention is to provide a train stability prediction method and system based on MOGRIFIER-GRU deep learning, which is conducive to reducing the cost of train stability monitoring and improving the wheel-rail force prediction accuracy.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a train stability prediction method based on MOGRIFIER-GRU deep learning, comprising the following steps:

[0007] Step S1: obtaining train acceleration and wheel-rail force data under different road conditions, and normalizing the data to obtain training data for training the prediction model;

[0008] Step S2: building a MOGRIFIER-GRU deep learning prediction model, and training the MOGRIFIER-GRU deep learning prediction model through training data to obtain a trained MOGRIFIER-GRU deep learning prediction model;

[0009] Step S3: Collect the data sent back in real time by the actual running train, and predict the train stability through the trained MOGRIFIER-GRU deep learning prediction model.

[0010] Furthermore, the step S1 specifically includes the following steps:

[0011] Step S11: According to the actual situation, SIMPACK software is used to establish a train dynamics model, and the vertical acceleration and lateral acceleration of the left and right axle boxes and bogies of the train and the vertical wheel-rail force and lateral wheel-rail force of the track are simulated under different road conditions as the training data of the prediction model, and the input parameters x of the prediction model are defined. i ,y i As shown below:

[0012] x i =[Lab_LA i ,Lab_VA i ,Rab_LA i ,Rab_VA i ,Ltf_LA i ,Ltf_VA i ,Rtf_LA i ,Rtf_VA i ,F_LL i-1 ,F_LV i-1 ,F_RL i-1 ,F_RV i-1 ,v]

[0013] y i =[P_L i ,Q_L i ,P_R i ,Q_R i ]

[0014] Where i represents the i-th data point; Lab_LA i ,Lab_VA i Indicates the lateral and vertical acceleration values ​​at the left axle box of the train; Rab_LA i ,Rab_VAi Indicates the lateral and vertical acceleration values ​​at the right axle box of the train; Ltf_LA i ,Ltf_VA i Indicates the lateral and vertical acceleration values ​​at the left side of the train bogie; Rtf_LA i ,Rtf_VA i Indicates the lateral and vertical acceleration values ​​at the right side of the train bogie; P_L i ,Q_L i Indicates the vertical wheel-rail force and lateral wheel-rail force on the left side of the track; P_R i ,Q_R i It represents the vertical wheel-rail force and the lateral wheel-rail force on the right side of the track; v represents the train running speed;

[0015] Step S12: normalize the training data:

[0016]

[0017]

[0018] in, is the normalized data, is the data before normalization, where are the minimum and maximum values ​​of variables x and y respectively.

[0019] Furthermore, the step S2 specifically includes the following steps:

[0020] Step S21: Establish a GRU single neuron prediction model, by updating the gate z t , reset gate t The composition is as follows:

[0021] z t =sigmoid(ω z *[h t-1 ,x t ]+b z )

[0022] r t =sigmoid(ω r *[h t-1 ,x t ]+b r )

[0023] Among them, ω z ,ω r 、b z 、b r are the weight coefficient matrix and bias term of the update gate and reset gate respectively, h t-1 is the hidden layer output at the previous moment, xt Represents the input matrix at the current moment, [] represents matrix concatenation, and * represents matrix multiplication;

[0024] The update gate is used to control the extent to which the state information of the previous moment is brought into the current state. The larger the value of the update gate, the more state information of the previous moment is brought in. The reset gate is used to control how much information of the previous state is written to the current candidate set. The smaller the reset gate, the less information of the previous state is written.

[0025] Step S22: By updating gate z t , reset gate t Calculate the hidden layer output h at the current moment t :

[0026]

[0027]

[0028] in, is the candidate state value at the current moment, is the weight matrix and bias term of the candidate state;

[0029] Step S23: Introduce the MOGRIFIER mechanism, without changing the structure of the GRU itself, and let the input and state interact first to enhance the association ability of adjacent time information. The formula is as follows:

[0030] When i is an odd number and i∈[1,2,3,4,...,r]

[0031] When i is even and i∈[1,2,3,4,...,r]

[0032] Among them, Q i ,R i where represents the i-th Q interaction matrix and R interaction matrix; r represents the number of interaction rounds;

[0033] Step S24: construct a network structure, set the number of input layer nodes, the first hidden layer is composed of GRU neurons, the number of nodes is set, and the MOGRIFIER mechanism is adopted between the input layer and the first hidden layer; the second hidden layer is composed of GRU neurons, the number of nodes is set; set the number of output layer nodes; obtain the MOGRIFIER-GRU deep learning prediction model;

[0034] Step S25: training the prediction model, setting the hidden layer to a dropout mechanism with a set value, and randomly failing the hidden layer neurons with a set probability during the training process; using the L2 norm loss function MSE to evaluate the gap between the prediction model target and the actual output, where the smaller the function value, the smaller the difference between the actual output and the target output. The formula of the loss function is as follows:

[0035]

[0036] Where N is the number of data samples, y is the target value of the data samples, is the model prediction value after denormalization;

[0037] When the loss function value during training is lower than the set threshold or reaches the upper limit of the training cycle, the prediction model training is considered completed.

[0038] Furthermore, in step S3, the data transmitted back by the acceleration sensors on the axle box and bogie of the actual running train and the train data are first band-pass filtered to remove interference caused by the sensor DC power supply and the high-frequency noise between the wheel and rail; the filtered data are normalized and input into the MOGRIFIER-GRU deep learning prediction model, and the output value is denormalized to obtain the final predicted wheel-rail force, thereby obtaining the train stability index.

[0039] The present invention also provides a train stability prediction system based on MOGRIFIER-GRU deep learning, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the above-mentioned method steps can be implemented.

[0040] Compared with the prior art, the present invention has the following beneficial effects: it provides a train stability prediction method and system based on MOGRIFIER-GRU deep learning, which can reliably and accurately predict wheel-rail forces and stability indicators without using high-cost and high-complexity dynamometer wheelsets. In addition, the present invention uses the MOGRIFIER-GRU deep learning algorithm, which is superior to the traditional neural network algorithm in both prediction model performance and data time correlation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart of a method implementation of an embodiment of the present invention;

[0042] Figure 2 Schematic diagram of a single GRU neuron in an embodiment of the present invention;

[0043] Figure 3 A schematic diagram of the MOGRIFIER mechanism in an embodiment of the present invention;

[0044] Figure 4 This is a structural diagram of the MOGRIFIER-GRU deep learning prediction model in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0046] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0047] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0048] like Figure 1 As shown, this embodiment provides a train stability prediction method based on MOGRIFIER-GRU deep learning, including the following steps:

[0049] Step S1: Obtain train acceleration and wheel-rail force data under different road conditions, and normalize the data to obtain training data for training the prediction model.

[0050] Step S11: According to the actual situation, SIMPACK software is used to establish a train dynamics model, and the vertical acceleration and lateral acceleration of the left and right axle boxes and bogies of the train and the vertical wheel-rail force and lateral wheel-rail force of the track are simulated under different road conditions as the training data of the prediction model, and the input parameters x of the prediction model are defined. i ,y i As shown below:

[0051] x i =[Lab_LA i ,Lab_VA i ,Rab_LA i ,Rab_VA i ,Ltf_LA i ,Ltf_VA i ,Rtf_LA i ,Rtf_VA i ,F_LL i-1 ,F_LV i-1 ,F_RL i-1 ,F_RVi-1 ,v]

[0052] y i =[P_L i ,Q_L i ,P_R i ,Q_R i ]

[0053] Where i represents the i-th data point; Lab_LA i ,Lab_VA i Indicates the lateral and vertical acceleration values ​​at the left axle box of the train; Rab_LA i ,Rab_VA i Indicates the lateral and vertical acceleration values ​​at the right axle box of the train; Ltf_LA i ,Ltf_VA i Indicates the lateral and vertical acceleration values ​​at the left side of the train bogie; Rtf_LA i ,Rtf_VA i Indicates the lateral and vertical acceleration values ​​at the right side of the train bogie; P_L i ,Q_L i Indicates the vertical wheel-rail force and lateral wheel-rail force on the left side of the track; P_R i ,Q_R i It represents the vertical wheel-rail force and lateral wheel-rail force on the right side of the track; v represents the train speed.

[0054] Step S12: normalize the training data:

[0055]

[0056]

[0057] in, is the normalized data, is the data before normalization, where are the minimum and maximum values ​​of variables x and y respectively.

[0058] Step S2: Build a MOGRIFIER-GRU deep learning prediction model, and train the MOGRIFIER-GRU deep learning prediction model through training data to obtain a trained MOGRIFIER-GRU deep learning prediction model.

[0059] Step S21: Establish a GRU single neuron prediction model, such as Figure 2 As shown, it is updated by the gate z t , reset gate t The composition is as follows:

[0060] zt =sigmoid(ω z *[h t-1 ,x t ]+b z )

[0061] r t =sigmoid(ω r *[h t-1 ,x t ]+b r )

[0062] Among them, ω z ,ω r 、b z 、b r are the weight coefficient matrix and bias term of the update gate and reset gate respectively, h t-1 is the hidden layer output at the previous moment, x t Represents the input matrix at the current moment, [] represents matrix concatenation, and * represents matrix multiplication.

[0063] The update gate is used to control the extent to which the state information of the previous moment is brought into the current state. The larger the value of the update gate, the more state information of the previous moment is brought in. The reset gate is used to control how much information of the previous state is written to the current candidate set. The smaller the reset gate, the less information of the previous state is written.

[0064] Step S22: By updating gate z t , reset gate t Calculate the hidden layer output h at the current moment t :

[0065]

[0066]

[0067] in, is the candidate state value at the current moment, is the weight matrix and bias term of the candidate state.

[0068] Step S23: Since the current input and the state of the previous moment are independent of each other, they only interact in the gate and lack interaction before that, which may lead to the loss of information before and after. Figure 3 The MOGRIFIER mechanism shown does not change the structure of the GRU itself, but allows the input and state to interact first to enhance the association ability of adjacent time information. The formula is as follows:

[0069] When i is an odd number and i∈[1,2,3,4,...,r]

[0070] When i is even and i∈[1,2,3,4,...,r]

[0071] Among them, Q i ,R i where represents the i-th Q interaction matrix and R interaction matrix; r represents the number of interaction rounds, and generally 4 or 5 has better effect.

[0072] Step S24: Construct a network structure, such as Figure 4 As shown in the figure, the number of nodes in the input layer is 13, the first hidden layer is composed of GRU neurons, the number of nodes is 50, the MOGRIFIER mechanism is adopted between the input layer and the first hidden layer; the second hidden layer is composed of GRU neurons, the number of nodes is 20; the number of nodes in the output layer is 4. The MOGRIFIER-GRU deep learning prediction model is obtained.

[0073] Step S25: Train the prediction model, set the hidden layer to a dropout mechanism of 0.3, and randomly fail the hidden layer neurons with a probability of 30% during the training process, which can effectively alleviate the occurrence of overfitting and achieve the regularization effect to a certain extent. The L2 norm loss function (MSE) is used to evaluate the gap between the prediction model target and the actual output, where the smaller the function value, the smaller the difference between the actual output and the target output. The formula of the loss function is as follows:

[0074]

[0075] Where N is the number of data samples, y is the target value of the data samples, is the model prediction value after denormalization.

[0076] When the loss function value during training is lower than the set threshold or reaches the upper limit of the training cycle, the prediction model training is considered completed.

[0077] Step S3: Collect the data sent back in real time by the actual running train, and predict the train stability through the trained MOGRIFIER-GRU deep learning prediction model.

[0078] Specifically, the data sent back by the acceleration sensors on the axle box and bogie of the actual running train and the train data are first band-pass filtered at 0.5 to 40 Hz to remove the interference caused by the sensor DC power supply and the high-frequency noise between the wheel and rail; the filtered data are normalized and input into the MOGRIFIER-GRU deep learning prediction model, and the output value is denormalized to obtain the final predicted wheel-rail force, and then the train stability index is obtained.

[0079] The stability index includes derailment coefficient, wheel load reduction rate, and wheel axle lateral force. The calculation formula is as follows:

[0080]

[0081]

[0082] Axle lateral force = |Q_L-Q_R|

[0083] According to the indicators in the "Specifications for Evaluation and Test of Locomotive and Rolling Stock Dynamics Performance" (GB / T 5599-2019), they are as follows:

[0084]

[0085]

[0086] P0 is the static axle weight (kN)

[0087] In summary, the present invention is conducive to finding the global optimal planning scheme while considering the distribution network reconstruction and the scale allocation of distributed power sources, greatly reducing the active power loss of the distribution network, improving the voltage stability of the distribution network, and thus improving its stable operation capability.

[0088] This embodiment also provides a train stability prediction system based on MOGRIFIER-GRU deep learning, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the above-mentioned method steps can be implemented.

[0089] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A train stability prediction method based on MOGRIFIER-GRU deep learning, characterized in that: The following steps are involved: Step S1: obtaining train acceleration and wheel-rail force data under different road conditions, and normalizing the data to obtain training data for training the prediction model; Step S2: Building a MOGRIFIER-GRU deep learning prediction model, and training the MOGRIFIER-GRU deep learning prediction model through training data to obtain a trained MOGRIFIER-GRU deep learning prediction model; Step S3: Collect the data sent back in real time by the actual running train, and predict the train stability through the trained MOGRIFIER-GRU deep learning prediction model; The step S1 specifically includes the following steps: Step S11: According to the actual situation, SIMPACK software is used to establish a train dynamics model, and the vertical acceleration and lateral acceleration of the left and right axle boxes and bogies of the train and the vertical wheel-rail force and lateral wheel-rail force of the track are simulated under different road conditions as the training data of the prediction model, and the input parameters x of the prediction model are defined. i ,y i As shown below: <h2 style=";text-align:left;direction:ltr">x<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> =[Lab_LA<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> ,Lab_VA<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> ,Rab_LA<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> ,Rab_VA<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> Ltf_LA<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> Ltf_VA<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> Rtf_LA<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> Rtf_VA<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> , F_LL i-1 ,F_LV i-1 ,F_RL i-1 ,F_RV i-1 ,v] y i =[P_L i ,Q_L i ,P_R i ,Q_R i ] Where i represents the i-th data point; Lab_LA i ,Lab_VA i Indicates the lateral and vertical acceleration values ​​at the left axle box of the train; Rab_LA i ,Rab_VA i Indicates the lateral and vertical acceleration values ​​at the right axle box of the train; Ltf_LA i ,Ltf_VA i Indicates the lateral and vertical acceleration values ​​at the left side of the train bogie; Rtf_LA i ,Rtf_VA i Indicates the lateral and vertical acceleration values ​​at the right side of the train bogie; P_L i ,Q_L i Indicates the vertical wheel-rail force and lateral wheel-rail force on the left side of the track; P_R i ,Q_R i It represents the vertical wheel-rail force and the lateral wheel-rail force on the right side of the track; v represents the train running speed; Step S12: normalize the training data: in, is the normalized data, is the data before normalization, where are the minimum and maximum values ​​of variables x and y respectively; The step S2 specifically includes the following steps: Step S21: Establish a GRU single neuron prediction model, by updating the gate z t , reset gate t The composition is as follows: z t =sigmoid(ω z *[h t-1 ,x t ]+b z ) r t =sigmoid(ω r *[h t-1 ,x t ]+b r ) Among them, ω z ,ω r , b z , b r are the weight coefficient matrix and bias term of the update gate and reset gate respectively, h t-1 is the hidden layer output at the previous moment, x t Represents the input matrix at the current moment, [] represents matrix concatenation, and * represents matrix multiplication; The update gate is used to control the extent to which the state information of the previous moment is brought into the current state. The larger the value of the update gate, the more state information of the previous moment is brought in. The reset gate is used to control how much information of the previous state is written to the current candidate set. The smaller the reset gate, the less information of the previous state is written. Step S22: By updating gate z t , reset gate t Calculate the hidden layer output h at the current moment t : in, is the candidate state value at the current moment, is the weight matrix and bias term of the candidate state; Step S23: Introduce the MOGRIFIER mechanism, without changing the structure of the GRU itself, and let the input and state interact first to enhance the association ability of adjacent time information. The formula is as follows: When i is an odd number and i∈[1,2,3,4,...,r] When i is even and i∈[1,2,3,4,...,r] Among them, Q i ,R i where represents the i-th Q interaction matrix and R interaction matrix; r represents the number of interaction rounds; Step S24: construct a network structure, set the number of input layer nodes, the first hidden layer is composed of GRU neurons, the number of nodes is set, and the MOGRIFIER mechanism is adopted between the input layer and the first hidden layer; the second hidden layer is composed of GRU neurons, the number of nodes is set; set the number of output layer nodes; obtain the MOGRIFIER-GRU deep learning prediction model; Step S25: training the prediction model, setting the hidden layer to a dropout mechanism with a set value, and randomly failing the hidden layer neurons with a set probability during the training process; using the L2 norm loss function MSE to evaluate the gap between the prediction model target and the actual output, where the smaller the function value, the smaller the difference between the actual output and the target output. The formula of the loss function is as follows: Where N is the number of data samples, y is the target value of the data samples, is the model prediction value after denormalization; When the loss function value during training is lower than the set threshold or reaches the upper limit of the training cycle, the prediction model training is considered completed.

2. The train stability prediction method based on MOGRIFIER-GRU deep learning according to claim 1 is characterized in that: In step S3, the data and train data transmitted back by the acceleration sensors on the axle box and bogie of the actual running train are firstly band-pass filtered to remove the interference caused by the sensor DC power supply and the high-frequency noise between the wheel and the rail; The filtered data is normalized and input into the MOGRIFIER-GRU deep learning prediction model, and the output value is denormalized to obtain the final predicted wheel-rail force, and then the train stability index is obtained.

3. A train stability prediction system based on MOGRIFIER-GRU deep learning, characterized in that: The method comprises a memory, a processor, and computer program instructions stored in the memory and executable by the processor. When the processor executes the computer program instructions, the method steps as described in any one of claims 1 to 2 can be implemented.

Images